Merge staging into main (#3)

* feat(engine): Architect agent + cross-scene StoryState coherence

Add a dedicated Architect LLM call at session start that expands the terse
world/style prompt into a persistent story bible (logline, genre, second-
person protagonist, cast, engineered opening hook). The bible seeds a
StoryState the Writer reads and patches every scene, carried + merged
across cuts (applyStoryStatePatch) so the story keeps a spine from beat
one instead of jumping between scenes.

- prompts: inject web-novel / short-drama / galgame craft into Writer +
  Architect; Writer emits storyStatePatch to update the running bible
- director: parallelize voice + non-entry portraits with the Painter
  (only entry-beat portraits block paint) to offset Architect latency
- architect: chat/parse guarded so a malformed response never aborts start
- types: StoryState / StoryStatePatch; required on Start/SceneResponse

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* docs: add AGPL-3.0 license, README i18n, and TTS accuracy fix (#2)

* docs: add AGPL-3.0 license, README i18n, and TTS accuracy fix

- LICENSE: add GNU AGPL v3 with InfiPlot copyright notice
- README.md: rewrite for open-source project, fix TTS description
  (TTS uses MiMo's own protocol, not OpenAI-compatible)
- README.zh-CN.md: add Simplified Chinese translation
- README.ja.md: add Japanese translation
- package.json: change license from UNLICENSED to AGPL-3.0-only

* fix: address Copilot review — .env.example TTS comment, zh-CN formatting

- .env.example: clarify TTS uses MiMo's own protocol, not OpenAI-compatible
- README.md: 'land paper after paper' → 'publish paper after paper'
- README.zh-CN.md: add spaces around '5 月', fix code formatting
  for model names (deepseek-v4-flash)

---------

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
Zonghao Yuan
2026-06-02 13:41:37 +08:00
committed by GitHub
parent 16707cc255
commit 588b668d14
16 changed files with 1639 additions and 162 deletions
+2
View File
@@ -20,3 +20,5 @@ npm-debug.log*
pnpm-debug.log*
repomix-output.xml
users.md
+663
View File
@@ -0,0 +1,663 @@
InfiPlot Copyright (C) 2025-2026 InfiPlot Contributors
GNU AFFERO GENERAL PUBLIC LICENSE
Version 3, 19 November 2007
Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU Affero General Public License is a free, copyleft license for
software and other kinds of works, specifically designed to ensure
cooperation with the community in the case of network server software.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
our General Public Licenses are intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
Developers that use our General Public Licenses protect your rights
with two steps: (1) assert copyright on the software, and (2) offer
you this License which gives you legal permission to copy, distribute
and/or modify the software.
A secondary benefit of defending all users' freedom is that
improvements made in alternate versions of the program, if they
receive widespread use, become available for other developers to
incorporate. Many developers of free software are heartened and
encouraged by the resulting cooperation. However, in the case of
software used on network servers, this result may fail to come about.
The GNU General Public License permits making a modified version and
letting the public access it on a server without ever releasing its
source code to the public.
The GNU Affero General Public License is designed specifically to
ensure that, in such cases, the modified source code becomes available
to the community. It requires the operator of a network server to
provide the source code of the modified version running there to the
users of that server. Therefore, public use of a modified version, on
a publicly accessible server, gives the public access to the source
code of the modified version.
An older license, called the Affero General Public License and
published by Affero, was designed to accomplish similar goals. This is
a different license, not a version of the Affero GPL, but Affero has
released a new version of the Affero GPL which permits relicensing under
this license.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU Affero General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Remote Network Interaction; Use with the GNU General Public License.
Notwithstanding any other provision of this License, if you modify the
Program, your modified version must prominently offer all users
interacting with it remotely through a computer network (if your version
supports such interaction) an opportunity to receive the Corresponding
Source of your version by providing access to the Corresponding Source
from a network server at no charge, through some standard or customary
means of facilitating copying of software. This Corresponding Source
shall include the Corresponding Source for any work covered by version 3
of the GNU General Public License that is incorporated pursuant to the
following paragraph.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the work with which it is combined will remain governed by version
3 of the GNU General Public License.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU Affero General Public License from time to time. Such new versions
will be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU Affero General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU Affero General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU Affero General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU Affero General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Affero General Public License for more details.
You should have received a copy of the GNU Affero General Public License
along with this program. If not, see <https://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If your software can interact with users remotely through a computer
network, you should also make sure that it provides a way for users to
get its source. For example, if your program is a web application, its
interface could display a "Source" link that leads users to an archive
of the code. There are many ways you could offer source, and different
solutions will be better for different programs; see section 13 for the
specific requirements.
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU AGPL, see
<https://www.gnu.org/licenses/>.
+148
View File
@@ -0,0 +1,148 @@
[English](README.md) · [简体中文](README.zh-CN.md) · 日本語
# InfiPlot
> AI がリアルタイムに生成する、初のインタラクティブ・ストーリーゲーム —— あなたが思い描く場面を入力すれば、それが目の前に、没入感たっぷりのビジュアルとして立ち上がり、あなた自身がその中へ飛び込めます。一幕一幕の筋書きも、一枚一枚の画像も、一人ひとりのキャラクターも、すべてマルチモーダル AI エンジンがその場で設計・生成します —— 幼い頃に夢見た「アニメの中に飛び込む」あの願いを叶えるために。
[![License: AGPL v3](https://img.shields.io/badge/License-AGPL%20v3-blue.svg)](LICENSE)
[![GitHub stars](https://img.shields.io/github/stars/zonghaoyuan/infiplot?style=social)](https://github.com/zonghaoyuan/infiplot/stargazers)
**▶ ベータ期間中は無料でプレイ、セットアップ不要 —— [infiplot.com](https://infiplot.com)**
このプロジェクトを面白いと思っていただけたら、リポジトリへの ⭐ が私たちにとって何よりの励みになり、より多くの人に届く助けにもなります。ありがとうございます!
---
## InfiPlot とは
InfiPlot は、AI がコンテンツをリアルタイムに生成する世界初のインタラクティブ・ストーリーゲームです。あらかじめ用意された筋書きもキャラクターも、声色さえもありません。すべてはあなたの求めに応じて、その場でカスタマイズして生成されます。私たちは、美少女ゲーム(galgame)、乙女ゲーム、フルモーション・ビデオ(FMV)ゲームに比肩する体験を目指しつつ、一人ひとりに合わせた*参加型のファンタジー*を提供します —— より没入感のある視聴体験で、あなたの想像力と好奇心を存分に満たします。
ひとことで言えば、私たちが作っているのは、AI がリアルタイムにコンテンツを生成する『Love Is All Around(完蛋!我被美女包围了!)』です。
6 歳の子どもでも、20 代の若者でも、35 歳でも、60 歳でも —— ここには、あなただけのファンタジーがあります:
ハリー・ポッターの世界で魔法を学ぶ。学校で誰もが憧れ、想いを寄せる存在になる。トップ誌・トップ会議に論文を出し続け、研究費にも事欠かない。『宮廷の諍い女(甄嬛传)』の世界で宮廷の駆け引きを味わう。あるいは若い頃に戻り、悔いの残るあの選択をやり直す……
---
## 開発の動機
私たちは、清華大学をはじめとする大学に集う若者のグループです。
一方で、私たち自身が galgame、乙女ゲーム、FMV、AI ロールプレイといったゲームのヘビーユーザーでした。楽しみながらも、もし筋書きが固定された選択肢に縛られず、チャットアプリ越しの会話ではなく AI キャラクターと深く関われたら、どれほど愉快で刺激的だろうと想像していました。
もう一方で、私たちはたまたま大規模モデルの技術を少しばかり理解しており、AI でアイデアを素早く形にでき、技術の道筋や既存技術で実現できる製品の限界について、ささやかな考えを持っていました。
きっかけは 2026 年 4 月 22 日、[@zan2434](https://x.com/zan2434) たちが [flipbook](https://flipbook.page/) を公開したことでした。この全く新しいインタラクションの形に、私たちは驚き、心を躍らせました。そして 5 月のある日、意気投合し、こうした製品を作ろうと決めました —— かつて諦めた幻想を叶える手助けをしつつ、マルチモーダルモデルがもたらす新しいインタラクションの形を探るために。
プロジェクトはまだごく初期で、多くの機能が未完成です。[issue](https://github.com/zonghaoyuan/infiplot/issues) でのフィードバックを歓迎します。あるいは開発チームに加わって、一緒に新たな可能性を探り、あなた自身の好奇心を満たしてください。
お問い合わせ:hi@infiplot.com
---
## 仕組み
テキスト・画像・音声モデルを基盤に、私たちは InfiPlot の目標を実現するためのマルチエージェント・フレームワークを構築しました。エージェントを **アーキテクト(Architect)・脚本家(Writer)・キャラクターデザイナー(Character Designer)・撮影監督(Cinematographer)・絵師(Painter** の 5 つの役割に分け、互いに連携させることで、物語の一貫性・キャラクターの一貫性・シーンの連続性を保ちつつ、できる限り魅力的な物語を目指します。
一回のプレイ全体を、私たちは**ストーリー(story)**と呼んでいます。
物語は一連の**シーン(scene)**として展開します。各シーンは、AI が描いた 1 枚の背景画と、短い**ビート(beat)**のツリー —— ナレーション、セリフ、ときおりの選択肢 —— で構成されます。シーン内のビートをタップしていく間、画像はそのまま動きません。選択肢が本当に新しい場所 —— 別の空間、新しい視点、時間の跳躍 —— へ導いたときだけ、AI は次のシーンを描きます。
```mermaid
flowchart TD
U["あなたの入力:世界設定 + 画風"] --> A["アーキテクト / Architect<br/>入力を解析 → 物語の全体構造(第一歩)"]
A --> W["脚本家 / Writer<br/>このシーンのビートを生成:ナレーション · セリフ · 選択肢"]
subgraph SCENE["1 シーンの生成"]
direction TB
W --> C["キャラクターデザイナー / Character Designer<br/>立ち絵 + 声(新キャラごとに並行)"]
W --> S["撮影監督 / Cinematographer<br/>カメラ構成 + 背景プロンプト"]
C --> P["絵師 / Painter<br/>立ち絵を参照に 16:9 背景を描画"]
S --> P
end
P --> SC["1 シーン:背景画 + ビートツリー"]
SC -. 次のシーンを先回り生成 .-> W
```
あなたがひとつのシーンを読んでいる間に、エンジンは**選択肢が導きうるシーンを先回りして生成**します —— 避けられない次の一歩については、そのさらに先のシーンまで。あなたが方向を選ぶ頃には、その画像はたいてい描き上がっているので、切り替えは一瞬に感じられます。いまはまだ多少の遅延を感じるかもしれませんが、ご安心ください —— 私たちは鋭意改善に取り組んでいます。
ボタンではなく背景そのものをクリックすると、**ビジョン(vision)**モデルを経由します。タップした位置を読み取り、いまのシーンを探索しているのか(新しい画像なしでビートを挿入)、先へ進もうとしているのか(新しいシーン)を判断します。これは flipbook から学んだ貴重な知見に基づくもので、この機能はいずれ InfiPlot を特徴づける鍵となり、プレイ体験をもう一段引き上げてくれると信じています。
アートの中には、従来型のゲーム UI は一切焼き込まれていません。AI は、あなたが選んだ任意のスタイル —— 「方眼紙の棒人間」でも「サイバーパンク・ノワール」でも —— で世界を描きます。セリフ枠と選択肢ボタンは、その上に重ねた軽量な HTML レイヤーで、シーンになじむよう調整されています。つまり UI は、毎回同じではなく、そのプレイの物語に寄り添って変化するのです。
---
## ワンクリックデプロイ
[![Deploy with Vercel](https://vercel.com/button)](https://vercel.com/new/clone?repository-url=https://github.com/zonghaoyuan/infiplot&root-directory=apps/web&env=TEXT_BASE_URL,TEXT_API_KEY,TEXT_MODEL,IMAGE_BASE_URL,IMAGE_API_KEY,IMAGE_MODEL,VISION_BASE_URL,VISION_API_KEY,VISION_MODEL,TTS_BASE_URL,TTS_API_KEY,TTS_SPEECH_MODEL,MOCK_IMAGE&envDescription=Three%20required%20providers%20%2B%20optional%20TTS.%20Any%20OpenAI-compatible%20endpoint%20works%20for%20text%2Fvision.%20TTS%20uses%20MiMo%27s%20own%20protocol.&envLink=https://github.com/zonghaoyuan/infiplot%23configuration-guide)
デプロイ後、Vercel プロジェクトで環境変数を設定してください —— 下記の[設定ガイド](#設定ガイド)を参照。Vercel プロジェクトの **Root Directory**`apps/web` に設定する必要があります(デプロイボタンが自動で渡します。手動設定の場合は Project Settings で指定してください)。
---
## 設定ガイド
InfiPlot は 4 種類のモデルプロバイダと通信します。**テキスト(Text)・ビジョン(Vision)は、任意の OpenAI 互換エンドポイント**OpenAI、OpenAI 互換プロキシ経由の Anthropic、Gemini、OpenRouter、DeepSeek、ローカルの Ollama など)を使用でき、自由に組み合わせられます。**画像(Image)**は現在 **Runware**(OpenAI 互換ではなく、独自の task-array プロトコル)を使用します —— レイテンシとコストを総合的に考慮した選択です。**音声(TTS)**は **Xiaomi MiMo** の独自音声デザイン/クローンプロトコルを使用します(これも OpenAI 互換ではありません)—— キャラクターごとの音声デザイン、クローン、行ごとの抑揚指示に対応します。
**1. プロバイダを選ぶ**
| プロバイダ | 環境変数 | 必須? | 推奨 |
|---|---|---|---|
| Text · ストーリー監督 | `TEXT_BASE_URL` `TEXT_API_KEY` `TEXT_MODEL` | ✅ | DeepSeek の `deepseek-v4-flash` |
| Image · シーン描画 | `IMAGE_BASE_URL` `IMAGE_API_KEY` `IMAGE_MODEL` | ✅ | [Runware](https://runware.ai) の `runware:400@6`FLUX.2 [klein] 9B KV |
| Vision · クリック解釈 | `VISION_BASE_URL` `VISION_API_KEY` `VISION_MODEL` | ✅ | Google の `gemini-3.5-flash` |
| TTS · キャラクター音声 | `TTS_BASE_URL` `TTS_API_KEY` `TTS_SPEECH_MODEL` | 任意 —— 空欄なら無音で動作 | Xiaomi MiMo の `mimo-v2.5-tts`(独自プロトコル、OpenAI 互換ではない) |
**2. 環境変数を設定する**
Vercel プロジェクト(**Settings → Environment Variables**)、またはローカル実行時は `apps/web/.env.local` に設定します。9 つの変数が必須で、TTS は任意です(空欄なら無音で動作)。低コストなテスト用のフラグもあります。
| 変数 | 効果 |
|---|---|
| `MOCK_IMAGE=true` | 画像生成をスキップし、レンダラが静的なプレースホルダを返します。ストーリー・音声・選択肢は通常どおり動作します。Runware のクレジットを消費せずに TTS を調整するのに最適です。 |
正確なフォーマットは `apps/web/.env.example` を参照してください。
**3. コストに注意**
推奨の 3 点セットでは、各**シーン**のコストの大半はテキスト LLM の呼び出しです。FLUX.2 [klein] 9B KV の画像は 1 シーンあたり概ね **\$0.00078**1792×1024、4 ステップ、サブ秒)で、残りはテキスト呼び出しです(deepseek-v4-flash を使えば非常に安価)。シーン内のビートをタップしていくのは無料です。切り替えを一瞬に保つため、エンジンは**選ぶ可能性はあるが実際には選ばないシーンも先行生成**します —— そのため実際の支出は、あなたが実際に見るシーン数よりやや高くなります。標準ではレート制限も認証もありません —— デプロイを公開すれば、請求額にそのまま反映されます。広く共有する前に、制限を追加し(必要に応じてプリフェッチの深さを下げ)てください。
---
## ロードマップ
- [ ] 知覚できないほどの低遅延
- [ ] モバイルブラウザ対応
- [ ] 大半のプロバイダに対応
- [ ] ユーザー登録・ログイン機能
- [ ] 静止画から動画へのアップグレード
- [ ] プレイ中の自由入力対応
- [ ] 音声インタラクション
- [ ] ストーリー共有機能
- [ ] モバイルアプリ
---
## コントリビュート
Issue と Pull Request を歓迎します。InfiPlot をローカルで実行するには(Node 20+ と pnpm 9+ が必要):
```bash
pnpm install
cp apps/web/.env.example apps/web/.env.local # キーを記入 —— 設定ガイドを参照
pnpm dev # http://localhost:3000 を開く
```
コントリビュートすることで、あなたの貢献が AGPL-3.0 の下でライセンスされることに同意したものとみなされます。
---
## スター推移
[![Star History Chart](https://api.star-history.com/svg?repos=zonghaoyuan/infiplot&type=Date)](https://star-history.com/#zonghaoyuan/infiplot&Date)
---
## ライセンス
[AGPL-3.0](LICENSE) © InfiPlot。コアは完全にオープンソースです。AGPL の「ネットワーク利用は配布とみなす」条項により、改変版をネットワークサービスとして運用する者は、そのソースコードも公開しなければなりません —— これによりコアをオープンに保ちつつ、将来のホスティング版や商用版の余地も残しています。
+100 -51
View File
@@ -1,60 +1,101 @@
English · [简体中文](README.zh-CN.md) · [日本語](README.ja.md)
# InfiPlot
> A real-time AI-generated interactive story game — painted scene by scene. You talk and explore within a scene; when the story turns a corner, it paints the next. You click. It paints. The story unfolds.
> The first real-time, AI-generated interactive story game — describe the scene you've always fantasized about and watch it come alive in front of you: immersive, visual, and yours to step into. Every plot beat, every image, every character is designed and generated on the fly by a multimodal AI engine — all to grant that childhood wish of crossing over into the cartoon.
[![License: AGPL v3](https://img.shields.io/badge/License-AGPL%20v3-blue.svg)](LICENSE)
[![GitHub stars](https://img.shields.io/github/stars/zonghaoyuan/infiplot?style=social)](https://github.com/zonghaoyuan/infiplot/stargazers)
**▶ Free to play during the beta, no setup required — [infiplot.com](https://infiplot.com)**
If you find this project interesting, a free ⭐ on the repo means the world to us — and it genuinely helps more people discover it. Thank you!
---
## What is InfiPlot
InfiPlot is the world's first interactive story game with content generated by AI in real time. There are no pre-written plots, no pre-made characters, not even pre-recorded voices — everything is generated on demand, tailored to you. We're aiming for an experience on par with bishōjo games (galgame), otome games, and full-motion-video (FMV) games, while adding a personal, *participatory fantasy* — a richer, more immersive audio-visual experience that gives your imagination and curiosity free rein.
In one line: what we're building is an AI-generated, real-time take on *Love Is All Around* (《完蛋!我被美女包围了!》).
Whether you're a six-year-old, a twenty-something, thirty-five, or sixty, there's a fantasy here that belongs to you and you alone:
Learn magic in the world of Harry Potter; become the one everyone at school adores and confesses to; publish paper after paper in top journals and conferences with grant money to spare; step into *Empresses in the Palace* and live out the court intrigue; or return to your younger self and make a different choice about something you regret…
---
## Why we built it
We're a group of young people from Tsinghua University and other schools.
On one hand, we're longtime, devoted players of galgames, otome games, FMV, and AI role-play games. Even while enjoying them, we kept imagining how much more delightful and thrilling it would be if the story choices weren't fixed in advance — or if you could truly interact with an AI character in depth, instead of just texting it through a chat app.
On the other hand, we happen to know a little about large-model technology: enough to turn ideas into working software quickly with AI, and to have formed some modest views on the technical paths available and the limits of what today's tech can build.
The spark came on April 22, 2026, when [@zan2434](https://x.com/zan2434) and others released [flipbook](https://flipbook.page/). We were stunned and delighted by this entirely new form of interaction. So one day in May, we agreed on the spot to build something like this — both to help people live out the fantasies they'd once set aside, and to explore the new modes of interaction that multimodal models make possible.
The project is still very early and many features are far from polished. We'd love your feedback — open an [issue](https://github.com/zonghaoyuan/infiplot/issues), or join our dev team and explore the new possibilities with us, and satisfy your own curiosity.
Get in touch: hi@infiplot.com
---
## How it works
The story unfolds as a sequence of **scenes**. Each scene is one AI-painted background plus a short tree of **beats** — moments of narration, dialogue, and the occasional choice. You tap through a scene's beats and the image stays put; only when a choice leads somewhere genuinely new — another place, a new point of view, a jump in time — does the AI paint the next scene.
Built on text, image, and audio models, we've assembled a multi-agent framework to deliver on InfiPlot's goal. We split the agents into five roles — **Architect, Writer, Character Designer, Cinematographer, and Painter** — that work together to keep the plot coherent, the characters consistent, and the scenes continuous, all while making the story as compelling as we can.
```
entering a scene
1. Text LLM directs the whole scene at once — a background prompt
plus a tree of beats (narration / dialogue / choices)
2. Image model paints the background once, 16:9, no UI baked in
[ tap through beats — no model calls, instant ]
├─ in-scene choice ──────▶ jump to another beat (instant)
└─ scene-change choice ──▶ the next scene
(usually pre-generated — see below)
We call each complete playthrough a **story**.
A story unfolds as a sequence of **scenes**. Each scene is one AI-painted background plus a short tree of **beats** — moments of narration, dialogue, and the occasional choice. You tap through a scene's beats and the image stays put; only when a choice leads somewhere genuinely new — another place, a new point of view, a jump in time — does the AI paint the next scene.
```mermaid
flowchart TD
U["Your input: world setting + art style"] --> A["Architect<br/>parses input → full story structure (first step)"]
A --> W["Writer<br/>directs this scene's beats: narration · dialogue · choices"]
subgraph SCENE["Generating one scene"]
direction TB
W --> C["Character Designer<br/>portrait + voice (parallel, per new character)"]
W --> S["Cinematographer<br/>shot composition + background prompt"]
C --> P["Painter<br/>renders the 16:9 background using portraits as reference"]
S --> P
end
P --> SC["One scene: background image + beat tree"]
SC -. speculatively pre-generate the next scene .-> W
```
While you're reading one scene, the engine **speculatively generates the scenes your choices could lead to** — and, for unavoidable next steps, the scene after that. By the time you pick a direction, its image is usually already painted, so the cut feels instant.
While you're reading one scene, the engine **speculatively generates the scenes your choices could lead to** — and, for unavoidable next steps, the scene after that. By the time you pick a direction, its image is usually already painted, so the cut feels instant. If you still notice some lag today, don't worry — we're working hard to bring it down.
Clicking the background itself (not a button) routes through a **vision** model: it reads where you tapped and decides whether you're exploring the current scene (it inserts a beat — no new image) or moving on (a new scene).
Clicking the background itself (not a button) routes through a **vision** model: it reads where you tapped and decides whether you're exploring the current scene (it inserts a beat — no new image) or moving on (a new scene). This builds on a valuable lesson we learned from flipbook, and we believe it will become one of InfiPlot's defining features — taking the experience to the next level.
There is no traditional game UI baked into the art. The AI paints the world in whatever style you pick — "stick figure on grid paper" or "cyberpunk noir" — and the dialogue panel and choice buttons are a light HTML layer drawn on top, tuned to sit over the scene.
There is no traditional game UI baked into the art. The AI paints the world in whatever style you pick — "stick figure on grid paper" or "cyberpunk noir" — and the dialogue panel and choice buttons are a light HTML layer drawn on top, tuned to sit over the scene. In other words, the UI fits the story of each playthrough, rather than staying the same every time.
---
## One-click deploy
[![Deploy with Vercel](https://vercel.com/button)](https://vercel.com/new/clone?repository-url=https://github.com/zonghaoyuan/infiplot&root-directory=apps/web&env=TEXT_BASE_URL,TEXT_API_KEY,TEXT_MODEL,IMAGE_BASE_URL,IMAGE_API_KEY,IMAGE_MODEL,VISION_BASE_URL,VISION_API_KEY,VISION_MODEL,TTS_BASE_URL,TTS_API_KEY,TTS_SPEECH_MODEL,MOCK_IMAGE&envDescription=Three%20required%20providers%20%2B%20optional%20TTS.%20Any%20OpenAI-compatible%20endpoint%20works%20for%20text%2Fvision%2Ftts.&envLink=https://github.com/zonghaoyuan/infiplot%23environment-variables)
[![Deploy with Vercel](https://vercel.com/button)](https://vercel.com/new/clone?repository-url=https://github.com/zonghaoyuan/infiplot&root-directory=apps/web&env=TEXT_BASE_URL,TEXT_API_KEY,TEXT_MODEL,IMAGE_BASE_URL,IMAGE_API_KEY,IMAGE_MODEL,VISION_BASE_URL,VISION_API_KEY,VISION_MODEL,TTS_BASE_URL,TTS_API_KEY,TTS_SPEECH_MODEL,MOCK_IMAGE&envDescription=Three%20required%20providers%20%2B%20optional%20TTS.%20Any%20OpenAI-compatible%20endpoint%20works%20for%20text%2Fvision.%20TTS%20uses%20MiMo%27s%20own%20protocol.&envLink=https://github.com/zonghaoyuan/infiplot%23configuration-guide)
After deploy, set the environment variables (see below) in your Vercel project. Nine are required; TTS is optional (leave blank to run silently); `MOCK_IMAGE=true` skips image generation for cheap TTS-only testing. The Vercel project's **Root Directory** must be set to `apps/web` (the deploy button passes this; if you configure manually, set it in Project Settings).
After deploy, set your environment variables in the Vercel project — see the [Configuration guide](#configuration-guide) below. The Vercel project's **Root Directory** must be `apps/web` (the deploy button passes this automatically; if you configure manually, set it in Project Settings).
---
## Environment variables
## Configuration guide
Three required providers + optional TTS. Text, Vision, and TTS accept any OpenAI-compatible endpoint (OpenAI, Anthropic via OpenAI-compat proxy, Gemini, OpenRouter, DeepSeek, local Ollama, …). Image goes to **Runware** (its own task-array protocol, not OpenAI-compatible).
InfiPlot talks to four kinds of model providers. **Text and Vision use any OpenAI-compatible endpoint** (OpenAI, Anthropic via an OpenAI-compat proxy, Gemini, OpenRouter, DeepSeek, local Ollama, …), so you can mix and match freely. **Image** currently goes to **Runware** (its own task-array protocol, not OpenAI-compatible), chosen for its combination of latency and cost. **TTS** uses **Xiaomi MiMo**'s own voice design / clone protocol (also not OpenAI-compatible) — per-character voice design, clone, and per-line delivery direction.
**1. Choose your providers**
| Provider | Variables | Required? | Recommended |
|---|---|---|---|
| Text · story director | `TEXT_BASE_URL` `TEXT_API_KEY` `TEXT_MODEL` | ✅ | `claude-opus-4-7` via Anthropic |
| Image · UI renderer | `IMAGE_BASE_URL` `IMAGE_API_KEY` `IMAGE_MODEL` | ✅ | `runware:400@6` (FLUX.2 [klein] 9B KV) via [Runware](https://runware.ai) |
| Vision · click reader | `VISION_BASE_URL` `VISION_API_KEY` `VISION_MODEL` | ✅ | `gemini-3-flash` via Google |
| TTS · per-character voice | `TTS_BASE_URL` `TTS_API_KEY` `TTS_SPEECH_MODEL` | optional — leave blank to run silently | `mimo-v2.5-tts` via Xiaomi MiMo |
| Text · story director | `TEXT_BASE_URL` `TEXT_API_KEY` `TEXT_MODEL` | ✅ | `deepseek-v4-flash` via DeepSeek |
| Image · scene renderer | `IMAGE_BASE_URL` `IMAGE_API_KEY` `IMAGE_MODEL` | ✅ | `runware:400@6` (FLUX.2 [klein] 9B KV) via [Runware](https://runware.ai) |
| Vision · click reader | `VISION_BASE_URL` `VISION_API_KEY` `VISION_MODEL` | ✅ | `gemini-3.5-flash` via Google |
| TTS · per-character voice | `TTS_BASE_URL` `TTS_API_KEY` `TTS_SPEECH_MODEL` | optional — leave blank to run silently | `mimo-v2.5-tts` via Xiaomi MiMo (own protocol, not OpenAI-compat) |
There's also a flag for cheap testing:
**2. Set the environment variables**
Set these in your Vercel project (**Settings → Environment Variables**), or in `apps/web/.env.local` for local runs. Nine variables are required; TTS is optional (leave blank to run silently). There's also a flag for cheap testing:
| Variable | Effect |
|---|---|
@@ -62,38 +103,46 @@ There's also a flag for cheap testing:
See `apps/web/.env.example` for the exact shape.
**3. Mind the cost**
With the recommended trio, each **scene** is dominated by the text-LLM call. The FLUX.2 [klein] 9B KV image is roughly **\$0.00078** per scene (1792×1024, 4 steps, sub-second); the text call is the rest (very cheap with deepseek-v4-flash). Tapping through a scene's beats is free. To keep transitions instant, the engine also **pre-generates scenes you might pick but don't** — so real spend runs somewhat higher than the scenes you actually see. There is no rate limiting or auth out of the box — if you make your deployment public, your bill will reflect that. Add limits (and consider lowering the prefetch depth) before sharing widely.
---
## Local development
## Roadmap
Requires Node 20+ and pnpm 9+.
- [ ] Latency you can't perceive
- [ ] Mobile browser support
- [ ] Compatibility with most providers
- [ ] User accounts and login
- [ ] Upgrade from static images to motion video
- [ ] Free-form player input mid-story
- [ ] Voice interaction
- [ ] Story sharing
- [ ] Mobile app
---
## Contributing
Issues and pull requests are welcome. To run InfiPlot locally (Node 20+, pnpm 9+):
```bash
pnpm install
cp apps/web/.env.example apps/web/.env.local
# fill in env vars (9 required + optional TTS/MOCK_IMAGE)
pnpm dev
# open http://localhost:3000
cp apps/web/.env.example apps/web/.env.local # fill in your keys — see the Configuration guide
pnpm dev # open http://localhost:3000
```
By contributing, you agree that your contributions are licensed under the AGPL-3.0.
---
## Project layout
## Star history
```
infiplot/
├── apps/web/ Next.js 16 app — pages + API routes (Vercel root)
└── packages/
├── types/ shared TypeScript types
├── ai-client/ unified OpenAI-compatible clients + Runware adapter
├── tts-client/ Xiaomi MiMo TTS adapter
└── engine/ multi-agent AI orchestration (open core)
```
`packages/engine` is the open core — pure TS, no Next.js or browser dependency. Import it directly to build your own interactive-narrative front-end (Tauri, Electron, CLI, anywhere).
[![Star History Chart](https://api.star-history.com/svg?repos=zonghaoyuan/infiplot&type=Date)](https://star-history.com/#zonghaoyuan/infiplot&Date)
---
## Cost & limits
## License
With the recommended trio, each **scene** is dominated by the text-LLM call. The FLUX.2 [klein] 9B KV image is roughly **\$0.001** per scene (1792×1024, 4 steps, sub-second); the text call is the rest. Tapping through a scene's beats is free. To keep transitions instant, the engine also **pre-generates scenes you might pick but don't** — so real spend runs somewhat higher than the scenes you actually see. There is no rate limiting or auth out of the box — if you make your deployment public, your bill will reflect that. Add limits (and consider lowering the prefetch depth) before sharing widely.
[AGPL-3.0](LICENSE) © InfiPlot. The core is fully open source. AGPL's "network use is distribution" clause means anyone who runs a modified version as a network service must also publish their source — this keeps the core open while leaving room for a future hosted or commercial edition.
+148
View File
@@ -0,0 +1,148 @@
[English](README.md) · 简体中文 · [日本語](README.ja.md)
# InfiPlot
> 第一款 AI 实时生成的交互式剧情游戏 —— 输入你幻想中的场景,然后让它沉浸式、可视化地呈现在你面前,并且亲身参与其中。每一幕剧情、每一张图片、每一个角色,都由多模态AI引擎实时设计和生成,只为满足你儿时那个穿越进动画里的梦想。
[![License: AGPL v3](https://img.shields.io/badge/License-AGPL%20v3-blue.svg)](LICENSE)
[![GitHub stars](https://img.shields.io/github/stars/zonghaoyuan/infiplot?style=social)](https://github.com/zonghaoyuan/infiplot/stargazers)
**▶ 内测期间,免费在线试玩,无需本地部署 —— [infiplot.com](https://infiplot.com)**
如果你觉得这个项目有意思,麻烦给仓库点一个免费的star ⭐ 这是对我们努力的最大肯定,也能实实在在地帮助更多人发现它。谢谢你!
---
## 项目介绍
InfiPlot是世界上第一款实现了AI实时生成内容的互动剧情游戏,这里没有预设好的剧情、角色,甚至没有预设好的音色,所有内容都根据你的需求定制化的生成。我们力求实现比肩美少女游戏(galgame)、乙女游戏(Otome games)和全动态真人互动游戏(FMV)的体验,同时又能提供个性化的”参与式幻想”,让视听体验更沉浸,充分满足你的想象力和好奇心。
用一句话说,我们要做的是一款用AI实时生成内容的《完蛋!我被美女包围了!》
无论你是六岁的小朋友,20岁的年轻人,35岁的青年还是60岁的长者,都能在这里满足独属于你的幻想:
穿越到哈利波特世界学习魔法、成为学校里所有异性青睐和表达爱意的对象、顶刊顶会发不停科研经费拿到手软、穿越到甄嬛传体验宫廷斗争、或者重返年轻为遗憾的事情重新做选择......
---
## 开发初衷
我们是一群来自清华大学等高校的年轻人。
一方面,我们本来就是galgame、乙女游戏、FMV、AI角色扮演游戏这类游戏的深度用户,在享受游戏体验的同时,也会想象如果能选择不被预设的剧情选项,或者和对话的AI角色深度互动而不只是通过聊天软件聊天,该是多么愉快刺激的体验。
另一方面,我们恰好又对大模型技术有些了解,能用AI快速实现想法,对技术路线和基于已有技术的产品能力边界有一些浅薄的思考。
契机发生在 2026 年 4 月 22 日,[@zan2434](https://x.com/zan2434) 等人发布了 [flipbook](https://flipbook.page/),我们对这种全新的交互形态感到震惊和欣喜。于是在 5 月的某一天,我们一拍即合,决定做一款这样的产品,既帮助大家满足那些曾经遗憾过的幻想,又能够探索多模态模型所带来的新的交互形态。
目前我们的项目还很早期,有许多功能尚不完善,欢迎提交 [issues](https://github.com/zonghaoyuan/infiplot/issues) 反馈问题,或者加入我们的开发团队一起探索新的可能性,满足你的好奇心。
联系方式:hi@infiplot.com
---
## 工作原理
基于文本、图像和音频模型,我们搭建了一个多智能体框架来实现InfiPlot的目标。我们把agent分为架构师、编剧、角色设计师、场景布置师和画家五个职能,让他们之间相互配合,在保证剧情连贯性、角色一致性、场景一致性的基础上,尽可能使得剧情足够富有吸引力。
我们把每一次游玩的整体体验称为故事(story)。
故事以一连串**场景(scene)**的形式展开。每个场景由一张 AI 绘制的背景图,加上一棵简短的**节拍(beat)**树组成 —— 也就是旁白、对话和偶尔出现的选项。你逐拍点过一个场景时,画面始终不变;只有当某个选项把你带到真正全新的地方 —— 换了空间、换了视角、跳跃了时间 —— AI 才会绘制下一幕场景。
```mermaid
flowchart TD
U["你的输入:世界设定 + 画风"] --> A["架构师 / Architect<br/>解析输入 → 完整剧情结构(第一步)"]
A --> W["编剧 / Writer<br/>生成本幕节拍:旁白 · 对话 · 选项"]
subgraph SCENE["每一幕场景的生成"]
direction TB
W --> C["角色设计师 / Character Designer<br/>立绘 + 音色(每个新角色并行)"]
W --> S["场景布置师 / Cinematographer<br/>镜头编排 + 背景提示词"]
C --> P["画家 / Painter<br/>以立绘为参考渲染 16:9 背景"]
S --> P
end
P --> SC["一幕场景:背景图 + 节拍树"]
SC -. 预测式预生成下一幕 .-> W
```
当你正在阅读一幕场景时,引擎会**预测式地生成你的选项可能通向的那些场景** —— 对于无法回避的下一步,还会再往前生成一幕。等你真正选定方向时,那一幕的图通常已经画好了,于是切换瞬间完成、毫无停顿。如果你现在仍然感到有些延迟,别担心,我们正在努力优化它。
直接点击背景本身(而非按钮)会走一个**视觉(vision)**模型:它读取你点击的位置,判断你是在探索当前场景(于是插入一个节拍 —— 不生成新图),还是要继续前进(生成一幕新场景)。这是基于我们从flipbook那里学到的宝贵认知,我们相信这个功能会在未来成为InfiPlot的关键功能,让你的游玩体验更上一层楼。
未来,画面里将没有烤进任何传统的游戏 UI。AI 会用你选择的任意风格来描绘整个世界 —— 「方格纸上的火柴人」也好,「赛博朋克黑色电影」也罢 —— 而对话框和选项按钮,只是叠在画面之上、并为贴合场景而精心调校过的一层轻量 HTML。也就是说,每次游玩时,UI都会契合当前的故事,而不是一成不变。
---
## 一键部署
[![Deploy with Vercel](https://vercel.com/button)](https://vercel.com/new/clone?repository-url=https://github.com/zonghaoyuan/infiplot&root-directory=apps/web&env=TEXT_BASE_URL,TEXT_API_KEY,TEXT_MODEL,IMAGE_BASE_URL,IMAGE_API_KEY,IMAGE_MODEL,VISION_BASE_URL,VISION_API_KEY,VISION_MODEL,TTS_BASE_URL,TTS_API_KEY,TTS_SPEECH_MODEL,MOCK_IMAGE&envDescription=Three%20required%20providers%20%2B%20optional%20TTS.%20Any%20OpenAI-compatible%20endpoint%20works%20for%20text%2Fvision.%20TTS%20uses%20MiMo%27s%20own%20protocol.&envLink=https://github.com/zonghaoyuan/infiplot%23configuration-guide)
部署完成后,在 Vercel 项目里填好环境变量 —— 详见下方的[配置教程](#配置教程)。Vercel 项目的 **Root Directory** 必须设为 `apps/web`(一键部署按钮会自动带上;若手动配置,请在 Project Settings 里设置)。
---
## 配置教程
InfiPlot 会与四类模型供应商通信。**文本(Text)和视觉(Vision)都使用 OpenAI 兼容的接口**OpenAI、通过 OpenAI 兼容代理的 Anthropic、Gemini、OpenRouter、DeepSeek、本地 Ollama……),可以自由搭配。**图像(Image)**目前接入 **Runware**(其自有的 task-array 协议,并非 OpenAI 兼容),因为延迟和成本的叠加考量。**语音(TTS)**使用**小米 MiMo** 自有的音色设计/克隆协议(同样不是 OpenAI 兼容)——支持角色级音色设计、克隆与逐行演绎指导。
**1. 选择你的供应商**
| 供应商 | 环境变量 | 是否必填 | 推荐 |
|---|---|---|---|
| Text · 剧情导演 | `TEXT_BASE_URL` `TEXT_API_KEY` `TEXT_MODEL` | ✅ | DeepSeek 的 `deepseek-v4-flash` |
| Image · 场景渲染 | `IMAGE_BASE_URL` `IMAGE_API_KEY` `IMAGE_MODEL` | ✅ | [Runware](https://runware.ai) 的 `runware:400@6`FLUX.2 [klein] 9B KV |
| Vision · 点击解读 | `VISION_BASE_URL` `VISION_API_KEY` `VISION_MODEL` | ✅ | Google 的 `gemini-3.5-flash` |
| TTS · 角色配音 | `TTS_BASE_URL` `TTS_API_KEY` `TTS_SPEECH_MODEL` | 可选 —— 留空则静音运行 | 小米 MiMo 的 `mimo-v2.5-tts`(自有协议,非 OpenAI 兼容) |
**2. 填写环境变量**
在 Vercel 项目里设置(**Settings → Environment Variables**),或在本地运行时写进 `apps/web/.env.local`。九个变量为必填;TTS 可选(留空则静音运行)。此外还有一个用于低成本测试的开关:
| 变量 | 作用 |
|---|---|
| `MOCK_IMAGE=true` | 跳过图像生成,渲染器返回一张静态占位图。剧情、语音、选项照常运行。非常适合在不消耗 Runware 额度的情况下调试 TTS。 |
确切的字段格式见 `apps/web/.env.example`
**3. 注意成本**
使用推荐的三件套时,每一幕**场景**的开销主要来自文本 LLM 调用。FLUX.2 [klein] 9B KV 的图像大约 **\$0.00078** 一张(1792×1024,4 步,亚秒级);其余主要是文本调用(使用 `deepseek-v4-flash` 模型时成本极低)。逐拍点过一个场景是免费的。为了让切换瞬间完成,引擎还会**预先生成那些你可能选、但最终没选的场景** —— 所以真实花费会比你实际看到的场景数略高一些。开箱状态下没有任何限流或鉴权 —— 如果你把部署公开出去,账单会如实反映这一点。在大范围分享之前,请先加上限流(并酌情降低预取深度)。
---
## 路线图
- [ ] 用户感知不到的延迟
- [ ] 移动端浏览器适配
- [ ] 兼容大多数 provider
- [ ] 用户注册登录系统
- [ ] 由静态图升级为动态视频
- [ ] 游玩过程中支持用户自定义输入
- [ ] 语音交互
- [ ] 分享故事的功能
- [ ] 移动端 app
---
## 参与贡献
欢迎提交 Issue 和 Pull Request。在本地运行 InfiPlot(需要 Node 20+ 和 pnpm 9+):
```bash
pnpm install
cp apps/web/.env.example apps/web/.env.local # 填入你的密钥 —— 参见「配置教程」
pnpm dev # 打开 http://localhost:3000
```
提交贡献即表示你同意你的贡献以 AGPL-3.0 协议授权。
---
## Star 趋势
[![Star History Chart](https://api.star-history.com/svg?repos=zonghaoyuan/infiplot&type=Date)](https://star-history.com/#zonghaoyuan/infiplot&Date)
---
## 许可证
[AGPL-3.0](LICENSE) © InfiPlot。内核完全开源。AGPL 的「网络使用即分发」条款意味着:任何人若将修改后的版本作为网络服务运行,也必须公开其源代码 —— 这既让内核保持开放,又为未来的托管版或商业版本留下了空间。
+3 -1
View File
@@ -3,9 +3,11 @@
# Recommended setup: Xiaomi MiMo Token Plan for TEXT / VISION / TTS
# (one API key covers all three) + Runware for IMAGE (FLUX.2 [klein]).
#
# TEXT / VISION / TTS use OpenAI-compatible endpoints (any OpenAI-
# TEXT / VISION use any OpenAI-compatible endpoint (any OpenAI-
# compatible host works: OpenRouter, OpenAI, Anthropic via proxy,
# Gemini, DeepSeek, Ollama, ...).
# TTS uses Xiaomi MiMo's own voice design / clone protocol
# (not OpenAI-compatible; appends -voicedesign / -voiceclone).
#
# IMAGE uses Runware's own task-array protocol (not OpenAI-compatible);
# the adapter posts an `imageInference` task to IMAGE_BASE_URL.
+3
View File
@@ -187,6 +187,7 @@ function prefetchScenePath(
const carriedBase: Session = {
...baseSession,
characters: data.characters,
storyState: data.storyState,
};
prefetchScenePath(pool, carriedBase, [...steps, nextStep], depth + 1);
}
@@ -539,6 +540,7 @@ function PlayInner() {
},
],
characters: data.characters,
storyState: data.storyState,
};
visitedBeatsRef.current = [data.scene.entryBeatId];
setSession(initial);
@@ -635,6 +637,7 @@ function PlayInner() {
},
],
characters: result.characters,
storyState: result.storyState,
};
visitedBeatsRef.current = [result.scene.entryBeatId];
setSession(newSession);
+1 -1
View File
@@ -3,7 +3,7 @@
"version": "0.1.0",
"private": true,
"description": "AI 实时交互剧情游戏",
"license": "UNLICENSED",
"license": "AGPL-3.0-only",
"packageManager": "pnpm@9.12.0",
"engines": {
"node": ">=20"
+90
View File
@@ -0,0 +1,90 @@
import { chat } from "@infiplot/ai-client";
import type { ProviderConfig, Session, StoryState } from "@infiplot/types";
import { parseJsonLoose } from "../jsonParser";
import { ARCHITECT_SYSTEM, buildArchitectUserMessage } from "../prompts";
// ──────────────────────────────────────────────────────────────────────
// Architect agent — ONE LLM call at session start.
//
// Expands the user's (often terse) world + style prompt into a real story
// bible: a second-person protagonist with a want and a flaw, a single
// central dramatic question (logline), a genre frame that anchors the
// 爽点 rhythm, an engineered cold-open for scene 1 (nextHook), and a small
// intentional cast. Seeds the StoryState that the Writer reads and updates
// every scene — so the story has a spine from beat one instead of being
// improvised cold.
//
// Everything is best-effort coerced with fallbacks: a malformed LLM
// response can never abort session start — worst case the Writer just gets
// a thinner bible and improvises more.
// ──────────────────────────────────────────────────────────────────────
type RawStoryState = {
logline?: unknown;
genreTags?: unknown;
protagonist?: unknown;
castNotes?: unknown;
synopsis?: unknown;
openThreads?: unknown;
relationships?: unknown;
nextHook?: unknown;
};
function str(raw: unknown): string {
return typeof raw === "string" ? raw.trim() : "";
}
function strArray(raw: unknown): string[] | undefined {
if (!Array.isArray(raw)) return undefined;
const out = raw
.map((x) => (typeof x === "string" ? x.trim() : ""))
.filter((x) => x.length > 0);
return out.length > 0 ? out : undefined;
}
export async function runArchitect(
config: ProviderConfig,
session: Session,
): Promise<StoryState> {
try {
const raw = await chat(
config,
[
{ role: "system", content: ARCHITECT_SYSTEM },
{ role: "user", content: buildArchitectUserMessage(session) },
],
{ temperature: 0.85, responseFormat: "json_object" },
);
const parsed = parseJsonLoose<RawStoryState>(raw);
return {
// Stable spine — fall back to the raw world/style prompt so the bible is
// never wholly empty even if the model returns garbage.
logline: str(parsed.logline) || session.worldSetting,
genreTags: str(parsed.genreTags),
protagonist:
str(parsed.protagonist) ||
"你是这个故事的主角(第二人称视角,永不出现在画面里)。",
castNotes: str(parsed.castNotes) || undefined,
// Volatile seeds — the opening Writer will rewrite these via its patch.
synopsis: str(parsed.synopsis) || "故事即将开始。",
openThreads: strArray(parsed.openThreads),
relationships: strArray(parsed.relationships),
nextHook: str(parsed.nextHook) || undefined,
};
} catch (err) {
// chat() or parseJsonLoose() can throw (network / unrepairable JSON).
// The Architect is best-effort: never let it abort session start — return
// a minimal bible seeded from the raw prompt and let the Writer improvise.
const msg = err instanceof Error ? err.message : String(err);
console.error(`[architect] failed, using minimal bible: ${msg}`);
return {
logline: session.worldSetting,
genreTags: "",
protagonist:
"你是这个故事的主角(第二人称视角,永不出现在画面里)。",
synopsis: "故事即将开始。",
};
}
}
+30 -49
View File
@@ -15,25 +15,20 @@ import {
} from "../prompts";
// ──────────────────────────────────────────────────────────────────────
// CharacterDesigner agent — designs ONE new character end-to-end.
// CharacterDesigner agent — designs ONE new character.
//
// Pipeline (per character, all the slow parts are parallelized):
// Exposed as three GRANULAR stages so the director can schedule the slow
// parts around the Painter (a voice is never needed to paint a scene, and
// only entry-beat characters' portraits are referenced by the Painter):
//
// 1. LLM call — designs BOTH visual + voice cards in one shot
// (intentional: same agent thinks about who this character IS,
// which keeps appearance and vocal personality coherent)
// 1. designCharacterCard — ONE LLM call → visual + voice TEXT cards
// (intentional bundling: the same agent thinks about who this character
// IS, keeping appearance and vocal personality coherent)
// 2. renderCharacterPortrait — base portrait image (Runware URL + UUID)
// 3. provisionCharacterVoice — Xiaomi MiMo voicedesign → reference audio
//
// 2. In parallel:
// a. Image gen — base portrait (Runware returns URL + UUID in one shot;
// no separate upload round-trip is needed for cheap re-reference)
// b. Voice provisioning — Xiaomi MiMo voicedesign from voiceDescription
// → reference audio for later voiceclone synth
//
// 3. Returns merged Character ready to be added to session.characters
//
// Each step degrades gracefully — if image gen fails we return the
// character without a portrait; if voice gen fails we return without
// voice. The game keeps running even when sub-components fail.
// Each step degrades gracefully — if image gen fails the character just has
// no portrait; if voice gen fails it has no voice. The game keeps running.
// ──────────────────────────────────────────────────────────────────────
type CharacterDesignOutput = {
@@ -77,7 +72,7 @@ async function runDesignLLM(
//
// In mock mode we return the data URI as basePortraitUrl with no UUID
// (Painter is short-circuited anyway, so the lack of a UUID is moot).
async function renderPortrait(
export async function renderCharacterPortrait(
config: EngineConfig,
charName: string,
visualDescription: string,
@@ -101,7 +96,7 @@ async function renderPortrait(
}
}
async function provisionVoiceSafe(
export async function provisionCharacterVoice(
config: EngineConfig,
voiceDescription: string,
charName: string,
@@ -116,45 +111,31 @@ async function provisionVoiceSafe(
}
}
// Single-character design pipeline. Called by the orchestrator once per
// NEW character name; multiple characters in the same scene run their
// pipelines in parallel at the orchestrator level.
export async function designCharacter(
// The cheap first stage: design the visual + voice TEXT cards in one LLM
// call. The director then schedules renderCharacterPortrait /
// provisionCharacterVoice around the Painter. Multiple new characters in the
// same scene run this stage in parallel at the director level.
export type CharacterCard = {
name: string;
visualDescription?: string;
voiceDescription: string;
};
export async function designCharacterCard(
config: EngineConfig,
session: Session,
charName: string,
): Promise<Character> {
const tTotal = Date.now();
// Step 1 — LLM design (visual + voice). Must complete first.
): Promise<CharacterCard> {
const tDesign = Date.now();
const design = await runDesignLLM(config, session, charName);
tlog(`[charDesigner ${charName}] design LLM`, tDesign);
const visualDescription = design.visualDescription?.trim();
const voiceDescription =
design.voiceDescription?.trim() ||
`请根据角色名「${charName}」推断其性别、年龄与气质,生成最贴合的音色。所属世界观:${session.worldSetting}`;
// Step 2 — parallel: portrait + voice provisioning.
const tProvision = Date.now();
const portraitPromise = visualDescription
? renderPortrait(config, charName, visualDescription, session.styleGuide)
: Promise.resolve({} as Awaited<ReturnType<typeof renderPortrait>>);
const voicePromise = provisionVoiceSafe(config, voiceDescription, charName);
const [portrait, voice] = await Promise.all([portraitPromise, voicePromise]);
tlog(`[charDesigner ${charName}] portrait+voice parallel`, tProvision);
tlog(`[charDesigner ${charName}] TOTAL`, tTotal);
return {
name: charName,
voiceDescription,
visualDescription,
basePortraitUrl: portrait.basePortraitUrl,
basePortraitUuid: portrait.basePortraitUuid,
voice,
visualDescription: design.visualDescription?.trim() || undefined,
voiceDescription:
design.voiceDescription?.trim() ||
`请根据角色名「${charName}」推断其性别、年龄与气质,生成最贴合的音色。所属世界观:${session.worldSetting}`,
};
}
@@ -169,6 +150,6 @@ export async function provisionVoiceForName(
charName: string,
): Promise<Character> {
const voiceDescription = `请根据角色名「${charName}」推断其性别、年龄与气质,生成最贴合的音色。所属世界观:${session.worldSetting}`;
const voice = await provisionVoiceSafe(config, voiceDescription, charName);
const voice = await provisionCharacterVoice(config, voiceDescription, charName);
return { name: charName, voiceDescription, voice };
}
+39
View File
@@ -7,6 +7,7 @@ import type {
BeatNext,
ProviderConfig,
Session,
StoryStatePatch,
} from "@infiplot/types";
import { parseJsonLoose } from "../jsonParser";
import { WRITER_SYSTEM, buildWriterUserMessage } from "../prompts";
@@ -28,6 +29,9 @@ export type WriterOutput = {
sceneKey?: string;
entryBeatId: string;
beats: Beat[];
/** Rewritten volatile story memory — merged onto the carried StoryState by
* the director. Absent when the model omitted it (rare; bible just stales). */
storyStatePatch?: StoryStatePatch;
};
// Raw shapes — what the LLM produces before validation / coercion.
@@ -59,11 +63,18 @@ type RawBeat = {
activeCharacters?: RawActiveCharacter[];
next?: RawNext;
};
type RawStoryStatePatch = {
synopsis?: unknown;
openThreads?: unknown;
relationships?: unknown;
nextHook?: unknown;
};
type RawScene = {
sceneSummary?: string;
sceneKey?: string;
entryBeatId?: string;
beats?: RawBeat[];
storyStatePatch?: RawStoryStatePatch;
};
// ──────────────────────────────────────────────────────────────────────
@@ -321,6 +332,33 @@ function normalizeSceneKey(raw: string | undefined): string | undefined {
return slug.length > 0 ? slug : undefined;
}
function coerceStringArray(raw: unknown): string[] | undefined {
if (!Array.isArray(raw)) return undefined;
const out = raw
.map((x) => (typeof x === "string" ? x.trim() : ""))
.filter((x) => x.length > 0);
return out.length > 0 ? out : undefined;
}
// Pull the volatile story-memory rewrite out of the Writer's JSON. Only
// non-empty fields are kept; an all-empty/absent patch returns undefined so
// the director leaves the carried StoryState untouched.
function coerceStoryStatePatch(
raw: RawStoryStatePatch | undefined,
): StoryStatePatch | undefined {
if (!raw || typeof raw !== "object") return undefined;
const patch: StoryStatePatch = {};
const synopsis = typeof raw.synopsis === "string" ? raw.synopsis.trim() : "";
if (synopsis) patch.synopsis = synopsis;
const openThreads = coerceStringArray(raw.openThreads);
if (openThreads) patch.openThreads = openThreads;
const relationships = coerceStringArray(raw.relationships);
if (relationships) patch.relationships = relationships;
const nextHook = typeof raw.nextHook === "string" ? raw.nextHook.trim() : "";
if (nextHook) patch.nextHook = nextHook;
return Object.keys(patch).length > 0 ? patch : undefined;
}
export async function runWriter(
config: ProviderConfig,
session: Session,
@@ -359,6 +397,7 @@ export async function runWriter(
sceneKey: normalizeSceneKey(parsed.sceneKey),
entryBeatId,
beats,
storyStatePatch: coerceStoryStatePatch(parsed.storyStatePatch),
};
}
+178 -48
View File
@@ -6,8 +6,16 @@ import type {
ProviderConfig,
Scene,
Session,
StoryState,
StoryStatePatch,
} from "@infiplot/types";
import { designCharacter, provisionVoiceForName } from "./agents/characterDesigner";
import type { CharacterCard } from "./agents/characterDesigner";
import {
designCharacterCard,
provisionCharacterVoice,
provisionVoiceForName,
renderCharacterPortrait,
} from "./agents/characterDesigner";
import { runCinematographer } from "./agents/cinematographer";
import { runPainter } from "./agents/painter";
import {
@@ -27,26 +35,29 @@ import { INSERT_BEAT_SYSTEM, buildInsertBeatUserMessage } from "./prompts";
//
// Writer LLM (~3s, serial)
// │
// ├─ CharacterDesigner LLM × N (parallel per new char)
// │ │
// │ ├─ portrait gen (Runware returns URL + UUID in one call)
// │ └─ voice provisioning (parallel within agent)
// ├─ CharacterCard LLM × N (parallel per new char — TEXT only)
// ├─ Cinematographer LLM (parallel with the cards)
// │
// Cinematographer LLM (parallel with all of the above)
// wait for cards + cinema
// │
// wait for all parallel branches
// entry-beat portraits ──┐ (block the Painter — its refs)
// ▼ │
// Painter — generateImage │ (overlapped, NOT on the paint path):
// with referenceImages ├─ non-entry-beat portraits
// │ └─ ALL voice provisioning + orphan voices
// ▼
// await the overlapped work, fold into the registry
// │
// ▼
// Painter — generateImage with referenceImages (UUID/URL refs only;
// no base64 to upload, since outputType=URL gives both back)
// │
// ▼
// return { scene, sceneImageUrl, characters }
// return { scene, sceneImageUrl, characters, storyState }
//
// The Cinematographer intentionally does NOT depend on CharacterDesigner
// output — it only positions named characters in the frame, not their
// appearance. This unlocks the parallelism that makes the full pipeline
// ~9-12s instead of ~15-18s serial.
// Two deliberate decouplings unlock the parallelism:
// 1. The Cinematographer only POSITIONS named characters, so it needs no
// visualDescription and runs alongside the card LLMs.
// 2. The Painter only needs visualDescription TEXT (all on-stage) + the
// entry-beat characters' PORTRAITS (its referenceImages). Voices are
// never needed to paint, and non-entry portraits are never referenced —
// so both overlap the (longest) paint call instead of blocking it.
// ══════════════════════════════════════════════════════════════════════
function newSceneId(): string {
@@ -112,10 +123,33 @@ function pickPriorSceneReference(
return {};
}
// Merge the Writer's volatile story-memory patch onto the carried StoryState.
// The stable spine (logline/genreTags/protagonist/castNotes) is preserved;
// only the volatile fields the Writer is allowed to rewrite are overwritten,
// and only when the patch actually provided them. A missing carried state
// (legacy session from before the Architect existed) degrades to an empty
// spine rather than throwing.
function applyStoryStatePatch(
base: StoryState | undefined,
patch: StoryStatePatch | undefined,
): StoryState {
const start: StoryState =
base ?? { logline: "", genreTags: "", protagonist: "", synopsis: "" };
if (!patch) return start;
return {
...start,
synopsis: patch.synopsis ?? start.synopsis,
openThreads: patch.openThreads ?? start.openThreads,
relationships: patch.relationships ?? start.relationships,
nextHook: patch.nextHook ?? start.nextHook,
};
}
export type SceneResult = {
scene: Scene;
sceneImageUrl: string;
characters: Character[];
storyState: StoryState;
};
// ──────────────────────────────────────────────────────────────────────
@@ -156,17 +190,19 @@ export async function directScene(
writerOut.sceneKey,
);
// Stage 2 — parallel: CharacterDesigner(s) and Cinematographer.
// Cinematographer doesn't need character visualDescriptions (those are
// appended at Painter stage), so it runs concurrently with chardesign.
// ── Stage 2 — character cards (LLM) ∥ Cinematographer ──────────────────
// Both are cheap LLM calls and neither needs the other's output, so they
// run concurrently. The cards give us each new character's visualDescription
// TEXT; portraits + voices are deferred to Stage 3 so they can overlap the
// paint instead of blocking it.
const tParallel = Date.now();
const designPromises = newCharNames.map((name) =>
designCharacter(config, session, name).catch((err): Character => {
const cardPromises = newCharNames.map((name) =>
designCharacterCard(config, session, name).catch((err): CharacterCard => {
const msg = err instanceof Error ? err.message : String(err);
console.error(`[directScene] designCharacter(${name}) failed: ${msg}`);
// Last-resort fallback: register with name only so the speaker isn't
// unknown. Caller may try voice provisioning later or skip.
console.error(`[directScene] designCharacterCard(${name}) failed: ${msg}`);
// Last-resort fallback: a name + generic voice card so the speaker isn't
// unknown. No visualDescription → no portrait is attempted for them.
return {
name,
voiceDescription: `请根据角色名「${name}」推断其性别、年龄与气质。所属世界观:${session.worldSetting}`,
@@ -183,40 +219,102 @@ export async function directScene(
currentSceneKey: writerOut.sceneKey,
});
const [designedChars, cinemaOut] = await Promise.all([
Promise.all(designPromises),
const [cards, cinemaOut] = await Promise.all([
Promise.all(cardPromises),
cinemaPromise,
]);
tlog("[directScene] CharacterDesigner+Cinematographer parallel", tParallel);
tlog("[directScene] CharacterCards+Cinematographer parallel", tParallel);
// Merge new chars into a working registry that we'll pass to the Painter.
const characters = mergeCharacters(session.characters, designedChars);
// Working registry: existing characters + new cards. visualDescription text
// is present now; portraits + voices fill in over the next two phases.
let characters = mergeCharacters(
session.characters,
cards.map((c) => ({
name: c.name,
voiceDescription: c.voiceDescription,
visualDescription: c.visualDescription,
})),
);
// Edge case: a speaker referenced by the Writer might not have been in
// `activeCharacters` of any beat (LLM oversight), so they got skipped by
// newCharNames. Catch them here and at least provision a voice so the
// beat-audio path doesn't render silent. No portrait they weren't
// visible in the scene, so visual consistency doesn't matter for them.
// ── Stage 3 — portraits + voices, scheduled around the Painter ─────────
const tProvision = Date.now();
// Entry-beat character names: the ONLY portraits the Painter references
// (collectReferenceImages slots in the entry beat's speaker + activeChars).
const entryNames = new Set<string>();
if (entryBeat?.speaker && !isPovName(entryBeat.speaker)) {
entryNames.add(entryBeat.speaker);
}
for (const c of entryBeatActive) {
if (!isPovName(c.name)) entryNames.add(c.name);
}
type NamedPortrait = {
name: string;
basePortraitUrl?: string;
basePortraitUuid?: string;
};
// Kick off portrait gen for every NEW char that has a visualDescription.
// Entry-beat portraits block the Painter; the rest overlap it.
const entryPortraitPromises: Promise<NamedPortrait>[] = [];
const restPortraitPromises: Promise<NamedPortrait>[] = [];
for (const card of cards) {
const vd = card.visualDescription;
if (!vd) continue;
const p = renderCharacterPortrait(
config,
card.name,
vd,
session.styleGuide,
).then((res): NamedPortrait => ({ name: card.name, ...res }));
(entryNames.has(card.name) ? entryPortraitPromises : restPortraitPromises).push(p);
}
// Kick off voice provisioning for every NEW char (never on the paint path).
const voicePromises = cards.map((card) =>
provisionCharacterVoice(config, card.voiceDescription, card.name).then(
(voice): Character => ({
name: card.name,
voiceDescription: card.voiceDescription,
voice,
}),
),
);
// Edge case: a speaker the Writer referenced without listing in any beat's
// activeCharacters. collectActiveCharacterNames already includes speakers,
// so this is a rare defensive net. Provision a voice only (never on-screen).
const speakerNames = new Set(
writerOut.beats.map((b) => b.speaker).filter((n): n is string => Boolean(n)),
);
const orphanSpeakers = [...speakerNames].filter(
// Pattern B: "你" (player) is a valid speaker but never gets a Character
// record — TTS is intentionally skipped on the client. Filter POV out so
// provisionVoiceForName isn't accidentally invoked for the player.
(n) => !isPovName(n) && !characters.some((c) => c.name === n),
// record — TTS is intentionally skipped on the client.
(n) =>
!isPovName(n) &&
!characters.some((c) => c.name === n) &&
!cards.some((c) => c.name === n),
);
const orphanPromises = orphanSpeakers.map((n) =>
provisionVoiceForName(config, session, n),
);
if (orphanSpeakers.length > 0) {
const orphans = await Promise.all(
orphanSpeakers.map((n) => provisionVoiceForName(config, session, n)),
);
const merged = mergeCharacters(characters, orphans);
characters.splice(0, characters.length, ...merged);
}
// Stage 3 — Painter (depends on cinemaOut + characters).
// On-stage characters for THIS scene are the ones in any beat — pass them
// all so the archetype block covers anyone the player might encounter.
// Block the Painter ONLY on entry-beat portraits (its referenceImages).
const entryPortraits = await Promise.all(entryPortraitPromises);
characters = mergeCharacters(
characters,
entryPortraits.map((p) => ({
name: p.name,
voiceDescription: "", // preserved from the card by mergeCharacters
basePortraitUrl: p.basePortraitUrl,
basePortraitUuid: p.basePortraitUuid,
})),
);
tlog("[directScene] entry-beat portraits", tProvision);
// ── Stage 4 — Painter (depends on cinemaOut + on-stage visual cards +
// entry portraits). On-stage = everyone named in any beat, so the archetype
// block covers anyone the player might encounter in this scene.
const onStageCharacters = characters.filter((c) =>
allActiveNames.includes(c.name),
);
@@ -234,6 +332,30 @@ export async function directScene(
);
tlog("[directScene] Painter", tPainter);
// Fold in the work that overlapped the paint: remaining portraits, all
// voices, and any orphan-speaker voices. Awaited before returning so the
// session the client persists is fully provisioned for later scenes.
const tOverlap = Date.now();
const [restPortraits, voicedChars, orphanChars] = await Promise.all([
Promise.all(restPortraitPromises),
Promise.all(voicePromises),
Promise.all(orphanPromises),
]);
characters = mergeCharacters(
characters,
restPortraits.map((p) => ({
name: p.name,
voiceDescription: "",
basePortraitUrl: p.basePortraitUrl,
basePortraitUuid: p.basePortraitUuid,
})),
);
characters = mergeCharacters(characters, voicedChars);
if (orphanChars.length > 0) {
characters = mergeCharacters(characters, orphanChars);
}
tlog("[directScene] overlapped portraits+voices", tOverlap);
const scene: Scene = {
id: newSceneId(),
// scenePrompt is the cinematographer's English compositional output;
@@ -249,9 +371,17 @@ export async function directScene(
imageUrl: painted.imageUrl,
};
// Merge the Writer's volatile memory rewrite onto the carried bible so the
// throughline survives the next scene cut (orchestrator returns it; the
// client persists it back into the session).
const storyState = applyStoryStatePatch(
session.storyState,
writerOut.storyStatePatch,
);
tlog("[directScene] TOTAL", tTotal);
return { scene, sceneImageUrl: painted.imageUrl, characters };
return { scene, sceneImageUrl: painted.imageUrl, characters, storyState };
}
// ──────────────────────────────────────────────────────────────────────
+1
View File
@@ -9,6 +9,7 @@ export { annotateClick } from "./annotate";
export { synthesizeBeat } from "./voice";
export { mergeCharacters } from "./director";
export type { SceneResult } from "./director";
export { runArchitect } from "./agents/architect";
export type { WriterOutput } from "./agents/writer";
export type { CinematographerOutput } from "./agents/cinematographer";
export type { InsertBeatPartial } from "@infiplot/types";
+16 -2
View File
@@ -12,6 +12,7 @@ import type {
VisionRequest,
VisionResponse,
} from "@infiplot/types";
import { runArchitect } from "./agents/architect";
import { annotateClick } from "./annotate";
import { directInsertBeat, directScene } from "./director";
import { synthesizeBeat } from "./voice";
@@ -49,7 +50,18 @@ export async function startSession(
characters: [],
};
const { scene, sceneImageUrl, characters } = await directScene(config, session);
// Stage 0 — Architect: expand the terse world/style prompt into a story
// bible BEFORE the first scene. Serial by necessity (the opening Writer
// reads session.storyState), but it gives the whole story a spine from beat
// one — the latency is offset by the director's portrait/voice overlap win.
const tArchitect = Date.now();
session.storyState = await runArchitect(config.text, session);
tlog("[start] Architect", tArchitect);
const { scene, sceneImageUrl, characters, storyState } = await directScene(
config,
session,
);
tlog("[start] TOTAL", tTotal);
@@ -58,6 +70,7 @@ export async function startSession(
scene,
imageUrl: sceneImageUrl,
characters,
storyState,
};
}
@@ -71,7 +84,7 @@ export async function requestScene(
): Promise<SceneResponse> {
const tTotal = Date.now();
const { scene, sceneImageUrl, characters } = await directScene(
const { scene, sceneImageUrl, characters, storyState } = await directScene(
config,
req.session,
);
@@ -82,6 +95,7 @@ export async function requestScene(
scene,
imageUrl: sceneImageUrl,
characters,
storyState,
};
}
+156 -10
View File
@@ -3,20 +3,106 @@ import type {
Character,
Scene,
Session,
StoryState,
} from "@infiplot/types";
// ══════════════════════════════════════════════════════════════════════
// Multi-agent scene generation pipeline:
// Writer (编剧) — narrative + beats[] + per-beat activeCharacters
// Architect (编剧) — ONE-TIME at session start: the story bible
// (protagonist / logline / genre / opening hook /
// planned cast) → seeds StoryState
// Writer (编剧) — narrative + beats[] + per-beat activeCharacters,
// reads StoryState and emits a StoryStatePatch
// CharacterDesigner — per-new-character visual + voice cards
// Cinematographer (分镜导演) — sceneKey + English compositional prompt
// Painter (画师) — FLUX rendering with character archetypes
//
// Each agent owns one system prompt + one user-message builder below.
// All four agents see the same world / style guide, but each only reads
// the slice of session state it needs to make its decision.
// All agents see the same world / style guide, but each only reads the
// slice of session state it needs to make its decision.
// ══════════════════════════════════════════════════════════════════════
// ──────────────────────────────────────────────────────────────────────
// Shared — render the StoryState bible into a compact prompt block read
// by the Writer (and Architect, on revisions). Keeping one renderer means
// the bible looks identical to every agent that consumes it.
// ──────────────────────────────────────────────────────────────────────
export function renderStoryState(s: StoryState | undefined): string {
if (!s) return "";
const lines: string[] = ["【故事档案 / 主线记忆】"];
if (s.logline) lines.push(`主线(中心钩子):${s.logline}`);
if (s.genreTags) lines.push(`题材基调:${s.genreTags}`);
if (s.protagonist) lines.push(`主角「你」:${s.protagonist}`);
if (s.castNotes) lines.push(`核心配角:\n${s.castNotes}`);
if (s.synopsis) lines.push(`已发生(梗概):${s.synopsis}`);
if (s.relationships?.length) {
lines.push(`当前关系/情绪:\n${s.relationships.map((r) => `- ${r}`).join("\n")}`);
}
if (s.openThreads?.length) {
lines.push(`未收的悬念/伏笔:\n${s.openThreads.map((t) => `- ${t}`).join("\n")}`);
}
if (s.nextHook) lines.push(`接下来要往哪走(下一个钩子方向):${s.nextHook}`);
return lines.join("\n");
}
// ──────────────────────────────────────────────────────────────────────
// 0. Architect (总编剧) — ONE LLM call at session start.
//
// Turns the (often terse) user world + style prompt into a real story
// bible: a second-person protagonist with a want and a flaw, a single
// central dramatic question, a genre frame that anchors the 爽点 rhythm,
// an engineered opening hook (前3秒冷开场), and a small intentional cast.
// Everything downstream — Writer, CharacterDesigner — reads this so the
// story has a spine from beat one instead of being improvised cold.
// ──────────────────────────────────────────────────────────────────────
export const ARCHITECT_SYSTEM = `你是一部交互视觉小说的「总编剧 / 故事架构师」。玩家只给了你一句到几句的世界观和画风,你要在开拍前把它扩写成一份**故事档案(story bible)**,为后续每一幕定下脊梁。你不写具体台词、不写分镜、不设计立绘——你只搭骨架。
你深谙网文(番茄)、短剧(红果)与视觉小说(galgame)的爆款心法:
- **开篇即钩子**:黄金三章 / 前3秒法则。开场不铺垫世界观,直接抛出冲突、悬念或一个反常的瞬间。
- **代入感**:主角是第二人称「你」,是玩家的化身——要让玩家一进场就清楚"我是谁、我此刻卡在什么处境里、我想要什么"。
- **题材锚定爽点**:先选定一个清晰的题材框架(如 甜宠 / 校园暗恋 / 悬疑追凶 / 复仇逆袭 / 救赎治愈),它决定了情绪回报的节奏与类型。
- **戏剧问题**:整部故事由一个悬而未决的中心问题驱动(她到底是谁?你能否在记忆消失前查明真相?这场暗恋会走向哪里?)。
- **人设要鲜明且有反差**:每个核心角色一个强标签 + 一个反差面(外冷内热 / 傲娇 / 看似柔弱实则腹黑)。
你要产出(全部用中文,except 不需要英文):
- logline:一句话主线 / 中心戏剧问题,必须带钩子,让人想看下去
- genreTags:题材+基调标签,斜杠分隔,如 "甜宠 / 校园 / 慢热治愈带点伤感"
- protagonist:第二人称主角卡。包含:你是谁、你此刻正卡在什么具体处境里(要有即时张力)、你想要什么、一个软肋或秘密。50–120 字。
- castNotes:2–3 个核心配角,每行一个「名字:一句话人设(强标签+反差)+ 与你的关系/张力」。给真实好记的中文名字(不要"神秘女子"这种占位)。
- synopsis:开场此刻的情境梗概(故事尚未展开,就写"故事从……开始"),13 句。
- openThreads:开场就埋下的 1–3 个悬念/问题(数组)。
- nextHook:**第一幕**应当如何冷开场——具体描述开场那个抓人的瞬间/冲突(这会直接指导编剧写开场)。要画面感强、有张力。
设计硬规则:
- 主角「你」永不出现在画面里(第二人称 POV),所以 castNotes 里**不要**把"你/主角"当成一个角色。
- 配角名字要符合世界观(年代、地域、文化)。
- 一切服从玩家给的世界观与画风,不要擅自跑题;玩家信息少时,做最贴合、最有戏的合理扩写。
必须输出严格 JSON
{
"logline": "...",
"genreTags": "...",
"protagonist": "...",
"castNotes": "夏海:表面开朗的天台诗人,实则在用诗逃避家里的变故;与你是同班转学的邻座,对你有种说不清的在意。\\n班主任老周:…",
"synopsis": "...",
"openThreads": ["...", "..."],
"nextHook": "第一幕冷开场:……"
}
不要输出 JSON 以外的任何文本。`;
export function buildArchitectUserMessage(session: Session): string {
const parts: string[] = [];
parts.push(`世界观:${session.worldSetting}`);
parts.push(`画风:${session.styleGuide}`);
parts.push(
"\n请据此产出这部交互剧的故事档案(story bible),严格以 JSON 格式返回。",
);
return parts.join("\n");
}
// ──────────────────────────────────────────────────────────────────────
// 1. Writer (编剧) — drives the narrative.
//
@@ -27,7 +113,26 @@ import type {
// session.characters.
// ──────────────────────────────────────────────────────────────────────
export const WRITER_SYSTEM = `你是一交互视觉小说的「编剧」。每次基于世界观、画风、玩家历史、已登记角色,写出**一个完整场景的剧本**:场景背景概要 + 一组对话节拍 beats。你只负责**剧情和台词**——不设计角色形象、不写出图提示词、不做镜头调度,这些由其他 agent 完成。
export const WRITER_SYSTEM = `你是一交互视觉小说的「编剧」。每次基于【故事档案 / 主线记忆】、世界观、画风、玩家历史、已登记角色,写出**一个完整场景的剧本**:场景背景概要 + 一组对话节拍 beats,并在最后更新主线记忆。你只负责**剧情和台词**——不设计角色形象、不写出图提示词、不做镜头调度,这些由其他 agent 完成。
═══════════════════════════════════════════════════════════════════
爆款心法(番茄网文 / 红果短剧 / galgame 的叙事手感)—— 必须贯彻
═══════════════════════════════════════════════════════════════════
- **每个场景都要有钩子**:开头 1–2 个 beat 内就抛出新信息、悬念、冲突或情绪冲击,绝不平铺直叙地交代背景;结尾 beat 留一个让玩家"想知道接下来"的扣子。
- **兑现爽点 / 情绪回报**:按题材给观众想要的情绪(甜宠的心动、暗恋的暧昧拉扯、逆袭的扬眉吐气、悬疑的真相一角)。让玩家这一场"有所得"。
- **反转与反差**:适时打破预期——以为是 A 结果是 B、角色露出与第一印象相反的一面;但反转要可信、要扣主线。
- **快节奏、入戏快**:进场即冲突,少铺陈,删掉一切"为完整而存在"却不推进情绪的对话。
- **show, don't tell**:用动作、神态、潜台词、环境细节传递情绪,别直接旁白"她很难过"——让玩家自己读出来。
- **人设鲜明有反差**:每个角色一个强标签 + 一个反差面,台词紧贴其腔调(傲娇嘴硬心软、外冷内热、看似柔弱实则强势)。
- **选择要有分量**:choice 只出现在真正的岔路口,每个选项都要让玩家感到"通向不同的东西"(情绪指向不同 / 关系走向不同),别给等价的废选项。
═══════════════════════════════════════════════════════════════════
连贯性铁律(跨场景切换不能跳戏 —— 最重要)
═══════════════════════════════════════════════════════════════════
- 你会收到【故事档案 / 主线记忆】和上一场的结尾。**新场景必须从上一刻自然承接**——承接上一场的情绪、地点逻辑、人物状态与未收的悬念。
- 若给了「转场种子 nextSceneSeed」,把它当作"下一场的命题"去兑现,而不是另起炉灶;开场要让玩家感到"这正是我上一个动作 / 选择导致的结果"。
- 沿用主线记忆里的人物关系与情绪温度——别让刚告白的人下一场形同陌路,也别凭空遗忘已埋的伏笔。
- 推进、但别重置:每一场都让主线问题往前走一点(关系变化 / 真相揭露一角 / 新悬念浮现)。
一个场景包含:
- sceneSummary:当前场景的中文概要(地点、时间、氛围、关键事件——给后续的分镜导演看)
@@ -98,6 +203,13 @@ sceneKey 设计原则(重要 — 用于跨场景视觉一致性):
例:speaker="你" line="学姐,这把伞你拿着。"
- 同一个 beat 可以同时有 narration(心理活动 / 动作)和 speaker="你" + line(说出口的话)
更新主线记忆(storyStatePatch)—— 写完这一场后必做:
- synopsis:把这一场并入后的整体梗概,**压缩**到 3–5 句(别越写越长,旧细节该丢就丢)
- relationships:每个核心角色此刻与「你」的关系 / 情绪温度,每条一句(如 "夏海:暗恋升温,刚向你说了一半的告白被打断")
- openThreads:仍未收的悬念 / 伏笔——已收束的可移除、新埋的加入(但至少保留一条正在推进的主线,别把列表清空)
- nextHook:基于这一场的结尾,下一场应往哪走(给"下一次的你"一个明确命题,接住本场留下的扣子)
这些字段是写给"未来的你"的连贯性记忆,请认真写。
必须输出严格 JSON,结构如下:
{
"sceneSummary": "中文场景概要:地点+时间+氛围+关键事件",
@@ -149,13 +261,26 @@ sceneKey 设计原则(重要 — 用于跨场景视觉一致性):
]
}
}
]
],
"storyStatePatch": {
"synopsis": "把这一场并入后的滚动梗概,压缩到 3–5 句",
"relationships": ["夏海:暗恋升温,刚向你说了一半的告白被打断"],
"openThreads": ["夏海没说完的那句话到底是什么", "她书包里掉出的那张旧照片"],
"nextHook": "下一场:放学后的天台,她把你单独叫上去,要把话说完"
}
}
不要输出 JSON 以外的任何文本。`;
export function buildWriterUserMessage(session: Session): string {
const parts: string[] = [];
const bible = renderStoryState(session.storyState);
if (bible) {
parts.push(bible);
parts.push("");
}
parts.push(`世界观:${session.worldSetting}`);
parts.push(`画风:${session.styleGuide}`);
@@ -173,7 +298,9 @@ export function buildWriterUserMessage(session: Session): string {
}
if (session.history.length === 0) {
parts.push("\n这是故事的开场。请生成第一个场景,严格以 JSON 格式返回。");
parts.push(
"\n这是故事的开场。请按【故事档案】里的 nextHook 把第一幕的冷开场写出来——开场即抓人,别花笔墨铺垫世界观。写完后更新 storyStatePatch。严格以 JSON 格式返回。",
);
return parts.join("\n");
}
@@ -210,22 +337,40 @@ export function buildWriterUserMessage(session: Session): string {
});
const last = session.history.at(-1);
// The exact last moment the player stopped on — the new scene must continue
// seamlessly from this emotional beat, not reset to a neutral state.
if (last) {
const lastBeatId = last.visitedBeatIds.at(-1) ?? last.scene.entryBeatId;
const lastBeat = last.scene.beats.find((b) => b.id === lastBeatId);
if (lastBeat) {
const frag: string[] = [];
if (lastBeat.narration) frag.push(`旁白:${lastBeat.narration}`);
if (lastBeat.line) frag.push(`${lastBeat.speaker ?? "?"}${lastBeat.line}`);
if (frag.length) {
parts.push(
`\n上一刻(玩家停留的最后一个画面,新场景要从这里的情绪无缝承接):\n ${frag.join(" / ")}`,
);
}
}
}
const lastExit = last?.exit;
if (lastExit) {
if (lastExit.kind === "choice") {
parts.push(
`\n请基于「玩家在上一场选择了:${lastExit.label},生成下一个场景(参考种子${lastExit.nextSceneSeed})。`,
`\n承接「玩家在上一场选择了:${lastExit.label}无缝续写下一个场景(转场命题${lastExit.nextSceneSeed})。开场要让玩家感到这正是上一步的结果,并延续此刻的情绪。`,
);
} else {
parts.push(
`\n请基于「玩家自由动作:${lastExit.action},生成下一个场景。`,
`\n承接「玩家自由动作:${lastExit.action}无缝续写下一个场景,延续此刻的情绪与处境`,
);
}
} else {
parts.push("\n请生成下一个场景。");
parts.push("\n无缝续写下一个场景,延续上一刻的情绪。");
}
parts.push("严格以 JSON 格式返回。");
parts.push("写完后别忘了更新 storyStatePatch。严格以 JSON 格式返回。");
return parts.join("\n");
}
@@ -506,6 +651,7 @@ export const INSERT_BEAT_SYSTEM = `你是视觉小说编剧。玩家在当前场
- narration 与 line 加起来 ≤80 字
- 不要打破当前场景的物理状态(玩家仍在原地、对面仍是同一个角色)
- 不要生成选项或下一步指引 —— 玩家点击会自然回到原 beat
- 这个 beat 也要"有所得"——给玩家一个新细节、一丝潜台词或情绪波动(show, don't tell),别写成无意义的空台词
speaker 字段允许的取值**只有两种**(与主路径 Writer 一致 — Pattern B galgame 标准):
1. **已登记角色**里的 NPC 真名(**绝不允许引入新角色**)
+61
View File
@@ -139,6 +139,53 @@ export type BeatAudio = {
mime: string;
};
// ──────────────────────────────────────────────────────────────────────
// StoryState — the persistent "story bible" + evolving narrative memory.
//
// Created once at session start by the Architect agent (rich opening
// planning), then carried across every scene and incrementally updated by
// the Writer. This is the single throughline that keeps tone, cast, and
// stakes coherent across scene cuts — without it each Writer call would
// re-derive the whole arc from a flat beat log and drift.
//
// Split into STABLE fields (set by the Architect, rarely change) and
// VOLATILE fields (rewritten each scene via StoryStatePatch).
// ──────────────────────────────────────────────────────────────────────
export type StoryState = {
// ── Stable (Architect-authored; persists unless deliberately revised) ──
/** One-line central dramatic question / 主线钩子. */
logline: string;
/** Genre + tone tags anchoring the 爽点 framework, e.g. "甜宠 / 校园 / 慢热治愈". */
genreTags: string;
/** Second-person protagonist card: who 你 are, the immediate situation, the
* core want, and a flaw/secret. The audience proxy — never rendered. */
protagonist: string;
/** Key supporting cast and their relationship/tension with 你 (one per line). */
castNotes?: string;
// ── Volatile (rewritten each scene by the Writer's StoryStatePatch) ──
/** Rolling, compressed synopsis of what has happened so far (~3-5 句). */
synopsis: string;
/** Unresolved hooks / mysteries / questions still owed to the player. */
openThreads?: string[];
/** Current relationship/emotion state per character, e.g.
* "夏海:好感升温,刚向你告白了一半". */
relationships?: string[];
/** Where the story is heading next — the conflict/reversal/suspense the
* next scene should drive toward. Seeds the next scene's hook. */
nextHook?: string;
};
/** The volatile subset the Writer rewrites after each scene. Stable fields
* (logline/genreTags/protagonist/castNotes) are preserved by the merge. */
export type StoryStatePatch = {
synopsis?: string;
openThreads?: string[];
relationships?: string[];
nextHook?: string;
};
// ──────────────────────────────────────────────────────────────────────
// Session
// ──────────────────────────────────────────────────────────────────────
@@ -151,6 +198,13 @@ export type Session = {
history: SceneHistoryEntry[];
/** Character registry — accumulates across scenes; voices + portraits persist for reuse. */
characters: Character[];
/**
* Persistent story bible + evolving narrative memory. Set at session start
* by the Architect, carried by the client across every /api/scene call, and
* updated by the Writer each scene. Optional for back-compat with any
* session payload created before this field existed.
*/
storyState?: StoryState;
};
// ──────────────────────────────────────────────────────────────────────
@@ -207,6 +261,9 @@ export type StartResponse = {
imageUrl: string;
/** Character registry with voice references + visual cards provisioned. */
characters: Character[];
/** Story bible created by the Architect + updated by the opening scene's
* Writer. The client persists this into the session for later /api/scene calls. */
storyState: StoryState;
};
// /api/scene — generates the next Scene, given session whose latest
@@ -221,6 +278,10 @@ export type SceneResponse = {
/** Public CDN URL (or data URI in MOCK_IMAGE mode) for the rendered scene background. */
imageUrl: string;
characters: Character[];
/** Story bible after this scene's Writer applied its update. The client
* must persist this back into the session so the throughline survives the
* next scene cut. */
storyState: StoryState;
};
// /api/beat-audio — lazily synthesize one beat's voice. Client fires this