Multiic [v0.7.68]

0.7.68: custom contexts and character cards. hyperwoo. one thousand and one nights. Mika. story

→ reworked custom contexts
→ png/json character cards and loras for custom characters
→ ava is no longer ava — her new name is Mika
→ story:
– the prison section has been completed and connected to the rest of the story;
– a new d&d hospital section has been added, with a new “hospital (one thousand and one nights)” point in story navigation;
– playing hyperwoo with Mika: the choice about humanity from level 6 of the mini-game is now saved to Mika’s context. she will offer to play it in prison and when you choose her at night, after the park scene with Jo, Fiona, Dina, or Finn
→ a new one thousand and one nights d&d scenario in sandbox mode
→ tamagotchi: added grok voice support and a voice selector in the tray menu
→ new music
→ koboldcpp: contextshift/swa settings are now available directly in the game
→ android: the game now runs fully edge-to-edge, without those annoying borders
→ hyperwoo voxel mini-game:
– the ridge in the intro has been replaced with an arched stone fortress;
– the ending of level 2 is now a nighttime trail;
– added comets streaking across the sky on the trail and during the storm intro;
– lava and the sea of stars now fade smoothly into the fog without a visible rectangular edge;
– respawning now happens by sector instead of restarting the entire level;
– improved performance and forced use of the dedicated gpu;
– android version
→ the main game now prioritizes the dedicated gpu, which is especially relevant for laptops and primarily needed for mini-games and visual effects
→ changing the lora used by the koboldcpp image model no longer requires restarting the model
→ koboldcpp updated to 1.117.1; electron and three.js have also been updated
→ android: fixed a story-mode bug during the morning scenes in jo’s, dina’s, and fiona’s apartments that prevented further progress
→ collapsed settings tabs now display active models and other enabled features inside them
→ improved start screen performance
→ character thoughts now disappear after the text field is updated
→ rewrote the narrator context
→ updated the list of recommended local text models and api models for claude, gemini, chatgpt, and grok
→ minor interface improvements and code refactoring

———

0.7.45: skills, trust, video generation, a new mini-game, local models on android

→ story: added summaries after long free-play segments; skill system and character trust meter; rewritten the event with dina where she offers to play a video game; in ruin 3042, the reward for winning is 40⦵
→ uncensored-local-ai for android
→ gamepad support
→ koboldcpp: support for the anima model; short videos using wan models; more than one lora
→ you can now start a new line with shift + enter
→ fiona’s instagram link in the phone and in the story — in the prologue, when you unpack your things in the room
→ quick access to mini-games has been added to the main screen: “save princess”, “poker”, and “ruin 3042”
→ new music
→ sexy images for dina; images for ava
→ gallery items
→ tamagotchi: koboldcpp support, text model selection through the tray menu
→ d20 roll modifiers in dnd based on skills
→ story: rewrote a small part with dina in the hospital; fixed the room panel — it now hides when leaving characters’ apartments
→ koboldcpp: updated to version 1.115.2
→ comfyui: you no longer need to select the workflow file in the game again after making changes in comfyui
→ ruin 3042: fixed weapon stands; adjusted stun timings when jumping onto someone’s head
→ promptchan: proper error output if something goes wrong
→ rewrote and shortened the default negative prompt for images
→ fixed browser speech recognition
→ fixed nvl mode
→ updated the help section
→ response options and the d20 roll window no longer cover long text — you can scroll it now
→ for claude-fable-5, claude-opus-4-7/4-8, and newer models, deprecated sampling and temperature parameters are now correctly ignored so smaller bots don’t fail with an error
→ fixed the background zoom bug when switching from a 3d character to a 2d one
→ improved d20 roll detection — the bot no longer invents the value and waits for the roll instead

when choosing the skills, i was inspired by disco elysium — of course — modern analytic philosophers of mind, as well as nietzsche, horney’s work on neurosis, and byung-chul han’s the burnout society

the skill and trust system works both in the story and in sandbox mode. to reset skills, you need to start a new game after the short airplane video. by default, every character starts with 0% trust. in sandbox mode, skill and trust analysis/counting activates after sleeping and moving to the next day. in the story, it activates after long free-play conversation segments.
to save system context, skill information is currently given to the bot in a very minimal form, as a simple list. in the next step, once i optimize the characters’ system contexts, skills will affect gameplay more significantly

video generation: works with wan models, short clips up to 5 seconds. in the local image generation help section, at the very end, you’ll find the necessary instructions and links.

uncensored-local-ai on android: the models are very small, and the context window is small too. the full system context does not fit, so you will most likely need to trim the system context and shorten the message history

0.7.22: story, 3d city, multi-dialogues, grok voice, Fiona’s vrm, bug fixes

→ “grok voice think fast 1.0” speech
→ story, prologue:

• added an option to walk around the city before entering the house and before going to sleep;
• completely rewritten the first day after getting caught in the loop;
• dinner event with Fiona, Laura, and Kasper;
• poker event;

→ “multi-dialogues”:

• a 3rd character on screen;
• improved prompts for interactions between multiple characters;
• everyone in the dialogue tries to speak only for themselves;
• new “stay silent” button — active when there are more than 2 characters on screen;
• reworked character entrances/exits;

→ new images for Laura, Fiona, Josephine, Dina, Finn, and Kasper
→ new gallery items
→ updated images on the start screen
→ updated Fiona’s 3d models (vrm); added underwear outfit; changed eye behavior when tracking the mouse; increased the model’s tilt angle
→ the log now has a button for deleting all lines in the history below a selected point
→ music: three new tracks
→ updated story editor
→ updated help section
→ poker: added a hand-combination cheat sheet to the instructions
→ redesigned main screen; color correction and WebGL lighting for the background and characters
→ added a new story navigation point: “prologue, start of the loop”
→ new images for Fiona
→ redesigned settings tabs for speech and images
→ default image model ilustmix updated to v11
→ koboldcpp updated to 1.112.2
→ electron updated to 39.8.9
→ fixed a small branch if u choose to sleep with Ava on Fiona’s couch on day four
→ poker opening changed to a lighter circular transition effect
→ headers in story choices are now displayed above the options
→ fixed a bug with generated images getting blurred when certain triggers fired
→ redrew some weak locations
→ fixed the custom cursor arrow
→ button for skipping long dialogue chains — fixed problematic spots
→ android — images reduced and optimized

grok voice
uses the grok API key set in the text models section. Unfortunately, the trigger system based on emoji is not usable with fully voice-based models 

that means location changes / outfit changes and all other events will not work. but image generation works fine.
it seems like a new trigger system should be possible — I’ll think about it

Fiona’s vrm
I’ve gotten a bit better with blender and VRoid studio, but because of VRoid’s fundamental editor limitations, making the model I originally had in mind is pretty painful. VRoid has a very recognizable base: the same facial plasticity, weird limits on head/body shapes, and hair that lives by its own laws.
still, I managed to pull something out of it: adjusted the face proportions, improved the textures, made the hair feel more alive, tweaked the materials, and polished the model a bit in blender. Fiona no longer looks like a default dummy, and at least starts to resemble the character from the game.
i’m planning to keep improving the models, and later the animations for them too.
pls let me know what u think — compared to the other 3D characters in the game. maybe I’m wasting my time trying to squeeze something decent out of VRoid, and should just move to something else

main screen
It had been bothering me for a long time that the characters on the main screen didn’t really match the background in terms of color and lighting. from the outside, it might seem like a small thing, but in my view the start/menu screen is the first thing the player sees – it sets the tone – the user doesn’t know anything about the game yet, but they already read its visual coherence.
so yeah, now it looks ~ the way it was meant to. yoohoo •ᴗ•
the only thing is, I had to sacrifice the custom cursor on the main screen – because the canvas redraw causes an annoying cursor flicker. i’ll probably find a solution eventually, but for now this is the compromise

0.6.98: poker, story, laura, crypto, kobold gpu layers, bugfix

→ poker minigame
→ redesign of the bottom settings panel
→ new sandbox images for laura and fiona
→ koboldcpp: added gpu layers setting; updated to 1.110
→ added the ability to buy the person key with crypto
→ laura’s arc in the prologue rewritten (if u skip yoga and the variant after talking to her);
→ reworked falling asleep on day one (night-1);
→ reworked entering the loop in the prologue;
→ reworked meeting fiona
→ sped up the animation of navigation buttons appearing in story
→ fixed fiona’s 3d model animation in story
→ updated error and no-internet messages
→ fixed suggestions when using a local koboldcpp model
→ fixed gemini “freezing” issue
→ fixed contexts for english in story
→ android: fixed image generation settings via promptchan

poker:
launched via the phone button in sandbox.
strip poker starts after one participant is eliminated; if at the end it’s just you and one character left, after win/loss you’re taken to sandbox with that character

0.6.82: custom contexts, cursor, story content, image prompts, Jo, music

→ custom contexts: you can now save them either locally or publish them for everyone
→ a new cursor replaces the “fucking awesome cursor” and is now the default
→ koboldcpp: added a choice of 4 modes — auto/cuda/vulkan/cpu
→ in some story sections, dialogue choices now include character icons showing who you’ll meet on that route
→ a kas event in the prologue (in story navigation: “prologue, night”)
→ reworked the suggestions bot (quick replies): instructions were changed; replies are now more varied, higher quality, and more appropriate to what’s happening; they now also take the protagonist’s system context (yours) into account
→ new music; the currently playing track name is shown, plus loop and previous track buttons
→ the cafe and kitchen locations have been updated
→ new gallery items
→ images added to the start screen
→ jo: added new images and updated the bad ones
→ image prompt generation setting: last n messages, up to 3; when set to 1, it uses only the user’s message
→ ui/ux improvements; refactoring; transition animations fixed
→ koboldcpp: fixed startup on amd; updated to 1.109.2
→ reworked the shortcut from the prologue into “part 0”
→ improved the instruction for deciding when to roll a d20 in dnd mode
→ changed the trigger emoji for moving to the next day to reduce false positives
→ fixed the english in the prologue dinner scene
→ added to recommended models: claude-sonnet-4-6, grok-4-1-fast-non-reasoning, gpt-5.4, gpt-5.3-chat-latest (censored gpt)
→ fixed the layering in chess
→ fixed track switching in the sandbox, so it no longer breaks the in-game event queue or causes duplicate messages; it now works in “radio” mode by default — when one track ends, the next one starts playing

when publishing to custom contexts for the first time, you’ll need to verify your email by entering a code → then come up with a login and password. i hope the backend behaves itself and everyone who wants to share will be able to do it

0.6.61: Kas, qwen-tts, coque-tts, koboldcpp, music, refactoring, bugfix

→ new male character — Kasper (Kas)
→ local speech synthesis via qwen
→ in the win and win-local builds, for convenience, added a check for the required coqui and qwen tts libraries and download links in settings
→ koboldcpp for images — added gpu/cpu selection; for gpu, cuda is prioritized, if not found it launches vulkan; cpu is not recommended, but in rare cases can be useful
→ confetti effect in key story moments
→ new music
→ fixed a critical message history bug; by default, history is now limited to 18 messages (excluding system contexts) and the slider works correctly. the maximum value was increased from 40 to 50 (that’s a lot, and not all models can handle that, so if your local model starts struggling or api costs get too high, lower it to 27–32)
→ koboldcpp: now the image and text models don’t start together unless it’s selected in settings; set the launch priority for the image model — gpu/cpu; updated to 1.108.1
→ added info that Ava for sandbox unlocks as you progress through the story; the unlock was moved closer to the beginning of the part where Ava comes to, in story navigation this is the point — “in the park with Ava”
→ updated whisper (local speech recognition); the model was changed to the small and fast ggml-small-q5_1
→ improved local speech synthesis from coqui (xtts-v2); fixed behavior where punctuation marks were read incorrectly and bugged out as ‘point’; pauses became more natural
→ rebuilt the main screen layout; the custom character moved into a button in the bottom right corner (click the icon — selection screen; hover and click the picture — quick jump to the character)
→ improved snow performance
→ fixed determining the current relationship status with a character for models via koboldcpp
→ slightly increased the length for the image prompt builder bot
→ improved text model error messages: instead of an uninformative “crash” window — a short message
→ refactoring and code optimization; small ui/ux improvements

Kas is only in sandbox for now; he’ll appear in the story later. he currently has much less visual content than the other characters, but he got a well-developed context

qwen3-tts: creating a voice from a description is still in development; for now, cloning and preset voices are available

message history: because of a bug, all models except gemini were limited to 12 messages of history, thx to everyone who reported the issue

whisper: changed support for nvidia 50 series so the build doesn’t carry extra weight, but it still should work on new gpus and fall back to cpu as a last resort; if not — drop a comment pls

0.6.45: comfyui, story, endpoints, bugfix

→ comfyui support
→ rework: the very beginning of the game up until arriving at the house; day 2 city walk together with jo
→ the story now has arrow-based “navigation” while walking
→ the branch “if you fall asleep on a bench in the park with ava” has been merged into other branches
→ hints in settings got one more state — instead of generating hints on every reply, it adds a button that generates them on click
→ locations: added a lecture hall/classroom; updated the city, beach, park, apartments; redrew the bad ones
→ https connections are allowed; custom endpoints are no longer blocked; fixed the server-side password/version check
→ sandbox: fixed loading of autosaves and saves — they now load correctly and no longer wipe the log and message history
→ the “triumphant” return of copy/paste buttons for input fields
→ fixed values in the story editor that were being applied incorrectly
→ fixed the position of the speech recognition button
→ claude anthropic: fixed hint generation — the empty message “: ” before generation no longer appears
→ fixed the fireplace with fiona on android
→ updated the help section
→ koboldcpp updated to 1.107.1
→ minor ui tweaks

the hints button appears in the sandbox on open after the second reply

comfyui:
in the in-game help section there’s a guide with screenshots; quick important bits:
all settings are done in comfyui — for changes to affect requests from the game, you need to update the workflow json file via export api

• in comfyui settings, select port 8188 and put “*” in the “enable cors header..” field
• the game sends prompts to comfyui, so in your workflow it’s important to rename the fields to pos_prompt and neg_prompt (the game looks for fields starting with pos and neg)
• save json via file → export (api) and select it in the game settings

workflow examples are in resources/comfyui:
wf-mltc-base.json – minimal scenario
wf-mltc-face.json – with face replacement from a reference
i don’t have much experience with comfyui yet, so i’m not planning a more detailed guide for now — but it’s a popular tool and there’s lots of info out there. if something’s unclear, i recommend asking a neural net (chatgpt or any other thinking model), ideally with screenshots or (even more convenient) by sharing the exported json
share your workflows in the discord

have fun •ᴗ•

0.6.34: contexts, kobold for images, gemini-3, fireplace, endpoint, bugfixes

→ rewrote the system contexts for all characters: personalities now have more depth, detail, and individuality; optimized — ~10–15% shorter
→ koboldcpp: now supports images too; if koboldcpp is selected in settings but the model isn’t running, it auto-starts the model on request; updated to 1.105.4
→ safe place with Fiona by the fireplace (ask to go to the fireplace)
→ added gemini-3 models
→ added a visual effect for connecting to Ava’s chip — now randomly chosen from two variants; + a button to disable the effect
→ the save button now works even in places with a text input; on save it generates a summary (what happened over the last n messages, where n comes from the “message history” setting) with the character currently on screen
→ custom endpoint
→ fixed a desktop bug where hints couldn’t be clicked
→ desktop: adjusted the area around the text input so you can scroll long replies and the invisible wall doesn’t block 2.5d/3d
→ chatgpt-5: fixed bots (who’s speaking, summary, and quick replies) that relied on an older api revision
→ fixed remote tunnels for koboldcpp

koboldcpp for images:
it can run alongside the text model, it’s more flexible and cooler than stablediff cli, but: out of the box it only supports a single lora and applies it a bit weirdly, so when using koboldcpp it only takes the lora with the highest strength from your settings — the rest are ignored; and kobold also clamps the image size. if i manage to add multi-lora support and remove the size limits, koboldcpp will become the main local tool in the game, and stablediff will be retired to avoid dragging extra gigabytes into the build.
remote (ngrok/cloudflare) for images will be added later
upd: for the standard ilustmix model (and many others), you need to specify the vae file resources/stablediff/sdxl.vae.safetensors in the game settings

save-button summaries:
story and sandbox summaries are separated. for now it only works when the “message history” setting is enabled, and it needs more testing — if no issues show up, it’ll be available in the next release for everyone (no password) and enabled by default •ᴗ•

custom endpoint:
https only, still needs testing; depending on cors security policies it might not work for you — tell me which provider failed and i’ll add it to the exceptions in the next release

0.6.23: promptchan, story tweaks, gallery

→ image generation via api (no censorship): added support for the promptchan service
→ story updates, prologue: fiona’s shower visuals changed; a small variation in the initial conversation with fiona; reworked dina’s night branch (if you take a nap for a couple of hours and go to the kitchen). ava’s story images updated. dina’s system context rewritten (now closer to sandbox-style)
→ sandbox: updated images for ava and laura; updated warm clothing for finn, jo, and dina
→ sound on successful image generation
→ new elements in the gallery
→ general image generation settings moved into a separate block
→ story navigation added to a collapsible block
→ for local image generation, maximum sizes increased to 2048×1536
→ windows: fixed 3d vrm
→ links switched to boosty instead of patreon (what a bunch of assholes, grrr)
→ minor ui/ux tweaks
→ win-local default loras: hyperdetailed illustration and sexy details
→ a promptchan guide with links added to the help section

promptchan: before the local model appeared in the game, i used them for explicit visuals. by default, the illustration xl+ style is set with the artsy 2 filter; also, the hyper filter is good, in my opinion.
tokens without a subscription give access to regular styles; xl styles require a subscription to their not-so-cheap service, but if you cancel, in the best adobe traditions they’ll offer a huge discount for a month.
on android, requests have to go through a proxy due to cors and preflight checks; it should work the same as on desktop, maybe just a tiny bit slower.
if u can recommend another api-based image generation service without censorship, pls let me know — i’ll try to add it

0.6.11: journal, story content, bugfixes

→ story content, branches: morning at Josephine’s apartment (if u stay the night with her and Ava); morning at Fiona’s
→ journal: added buttons for determining your current relationship status with characters, plus animated images; story-choice editor
→ new trigger for when a character relationship level changes
→ kitchen location
→ 2d: images now dim only while the text-input field is in focus; 2.5d/3d: fixed field appearance for multiline input
→ android: fixed the room-navigation block in story mode when moving around the apartment; fixed copy/paste menu
→ quests/tasks: after receiving a reward, the task no longer remains in the list; slight redesign
→ redrawn poor-quality character images and winter start-screen backgrounds
→ replaced vrm models in the story part
→ optimized the location-detection bot (now shorter)
→ story image generation now correctly works with narrator text
→ updated trigger logic: if multiple triggers appear in one message, all of them now fire instead of just the first
→ nvl mode: fixed button display, text field, and a bug preventing image generation
→ minor ui tweaks
→ fixed elevenlabs tts
→ koboldcpp updated to 1.103
→ fixed a 2.5d menu bug when using koboldcpp
→ win-local: default lora updated to cryosistyle v3
→ base, person, visionary passwords changed

if you’re here for the story continuation — i’d recommend waiting a bit; the new sections are still small and will keep changing for a while

0.5.95: 2.5d, 3d, clean context, vroid, bugfixes

→ 2.5d: the angle shifts with mouse movement/touch, zoom works via scroll/pinch gestures; a render-quality slider was added to settings, default is 0.9
→ 3d: every character now has their own vrm model instead of the crappy readyplayer ones + undressed variants
→ added another state that clears emoji triggers from the game. “remove sexy context and personality” now also removes the short character info, leaving only game triggers
→ added gpt-5.1-chat-latest, gpt-5.1 and claude-sonnet-4-5-20250929 (still the most “humanlike”, top 1) to the recommended models
→ if a location is changed via the menu and it’s the same type as the current one, the bot won’t be triggered
→ fixed chatgpt behavior when using the extended model picker
→ small ui/ux tweaks
→ redrawn several failed images, especially faces; optimized backgrounds and redrew the weak ones
→ for 2.5d, added alternative static high-detail images for animated backgrounds and characters
→ fixed task display when hovering over notes
→ improved Ava’s chip visualization performance; if an image is generated, the visualization won’t be shown anymore

3d: Finn has a kinda scuffed model (i couldn’t find the dick, so he’s in shorts) and Ava’s model is also meh; the rest are better but still not at the quality i want, though closer. i’ll keep improving and periodically updating the models. if u wanna try vroid studio yourself — all original .vroid files are here, feel free to use, edit, remix, make your own

2.5d: lots of room for improvement; i still haven’t managed to get rid of the staircase artifacts, and i’m sure the parallax for far background layers can be made more pronounced

2.5d and 3d don’t work with two characters yet
on android, 2.5d intentionally reduces quality by 0.2 from the chosen value so it won’t lag

———

0.5.79: thoughts, quests, api models, bugfixes

→ character thoughts in a separate animated block (feels more alive and aesthetic to me)
→ quest/task system
→ 3 settings moved to experimental and now available starting from the base tier
→ windows, linux, mac: added live lists of all available api models for gemini, openrouter, openai, anthropic, grok, cohere, groq; the “recommended models” toggle has been retired
→ all characters are less rambling by default, but will be verbose if asked or when the situation calls for it
→ added items to the gallery
→ fixed a bug in the dialog history where the narrator’s text got “split” in streaming mode
→ improved censorship bypass for api models when forming the prompt for local image generation
→ fixed a bug with english language for quick replies
→ redesign of the “erase” part of a character’s context: added color markers, new phrasing, and an explanation in the help section
→ fixed a claude issue when two characters are on screen with streaming enabled
→ fixed the story intro animation for character reveals (bottom-up)
→ default number of messages in history increased to 18
→ tamagotchi memory message counts reduced from (45, 90, 135, 180, 225, 270) to (27, 45, 90, 135); 45 is recommended — higher values will be too expensive for api models, and locals won’t handle them
→ increased scrollbar thumb width in locations and the journal
→ optimized gallery images/videos and character pictures
→ koboldcpp updated to 1.100.1
→ fixed the nvidia/amd toggle display on entering the game
→ thoughts now hide when clicked and reappear on the next click
→ in win-local, the default text model saiga_nemo_12b was switched to a lower quantization Q4_K_M (didn’t notice much loss in answer quality, but it runs a bit faster)
→ optimized images, videos, and styles (scss)
→ fixed koboldcpp streaming output bug
→ ui adjusted, minor ui/ux tweaks
→ added setting “don’t limit ai response length” — disables the instruction that restricts ai verbosity

———

0.5.62: image generation, help/guide, story rewrites, tamagotchi, bugfixes

→ overhauled the help section (the most newcomer-friendly version so far)

image generation:
→ optimized the negative prompt (turns out overly long ones can lead to less crisp details)
→ improved auto-prompt writing: now if there are no people in the scene, the bot will focus on something else; especially handy in dnd modes; got a bit better when there are 3 characters in frame
→ now considers the last 3 messages instead of two
→ updated the default model to ilustmix_v9 (quantized to q8 — smaller size with no visible quality loss)
→ fixed auto image generation via api for grok and openai (the “every n messages” setting is taken from local generation); added a size parameter for openai

tamagotchi:
→ if the cursor doesn’t move for more than 5s, mika stops tracking
→ reacts if you stare at her chest (5 clicks)
→ Mika can now generate images; uses sd settings from the main game, but the size is fixed at 768×1024

→ in the main game, fiona’s 3d was updated (using the tamagotchi model)
→ in the story, the fiona changing scene was rewritten — it’s now vrm 3d
→ added an airport location

→ android: added loadable lists of all available api models for gemini, openrouter, openai, anthropic, grok, cohere, groq (for now only on android, other platforms will get it in the next update), the “recommended models” toggle has been retired

→ polished the very start of the game: added airport sounds and images, improved the quality of older images, and introduced a new kind of animated text
→ better gemini error handling: real model errors instead of a placeholder
→ reworked info dialogs; added a message when there’s no active text model, and a speech setup prompt when 3d is enabled by the story (shown once)
→ fixed the link in settings for testing voice on openai.fm
→ new icon for the koboldcpp console (the old one was obscenely bold)
→ android: fixed selection and caret in the multi-line text input
→ koboldcpp updated to 1.99.4
→ redesigned the kobold local models section in settings
→ local tts: fixed language/accent — now correctly taken from the ‘language ai’ setting

———

0.5.49: tamagotchi, save fix, bugfix, refactoring

→ tamagotchi (windows)
– Mika (agent) learned to set timers, just say/type: Mika, set a timer for n minutes/seconds/hours
– tracks the cursor
– fixed panel hiding and the drag button
– models: Fiona – red, club, strawberry, fishnet dresses; frog dress; skirt; stockings; shorts & top; in latex; no clothes. and sexy bunny
– added breast physics (not perfect yet, but god it’s so fun to work on this)
– pdf manual
→ chibi anger animation for dina when she gets mad
→ branch with Dina and Ava
→ story navigation through rooms of the location (dina’s apartment)
→ browser speech recognition moved to local server (i was dumb and didn’t know it was possible)
→ save fix*
→ koboldcpp updated to 1.98.1
→ code optimized and refactored
→ font selection fixed

*on android an old-branches bug might still remain, but I couldn’t reproduce it, so it’s probably fine
and thanks to everyone who reported the problem

0.5.31: gpt-5, hot model, bugfix

→ story mode now supports image generation – prompts are now built to match the current location and characters on screen
→ new pool location
→ redesigned location screen with captions
→ added claude-opus-4.1 (noticeably different from claude-sonnet-3.7, but costs way too much – 5× more, use with caution)
→ support for gpt-5, gpt-5-chat, gpt-5-mini, gpt-5-nano
→ sd:
– improved prompt generation for images
– sd server (hot model) now auto-restarts up to 3 times
– sd doesn’t like long prompts – for auto-generation they’re now forcibly trimmed to 540 chars to avoid crashes
– max length for negative/user prompts limited to 270 and 180 chars
– settings now have a “self description” field
– fields for manually editing clothing and location
– fixed a bug where the default LoRA was applied at game start, requiring opening settings to load the chosen ones
→ local speech synthesis chunk size increased from 200 to 450 chars
→ koboldcpp updated to 1.97.4
→ fixed openrouter model switching – no restart needed when changing from the popup
→ added cuda support in sd for linux (except 50x cards), fixed img_gen error, appImage size greatly reduced, resources for koboldcpp, stablediff and whisper moved into a separate folder
→ fixed “invalid mode txt2img” error for mac

gpt-5:
temperature doesn’t work for new models – it’s stuck at 1, and top_p (which was 0.9 in-game) no longer has any effect. censorship hits hard on the “thinking” models (gpt-5, gpt-5-mini, gpt-5-nano), but can be partially bypassed by continuing sexualized plots with them after starting on gemini / claude-3.7 / gpt-4.1.
gpt-5-chat is better – censorship is about the same as gpt-4.1, maybe slightly more vanilla

sd: for better accuracy, raise cfg-scale to 8-9 and use extra LoRAs sparingly – only ones that affect overall style. LoRAs with baked-in poses/characters reduce relevance

———

0.5.22: local tts, spectrum analyzer, hot sd, gallery

→ windows: local tts
→ local image gen: → added local server launch so the model stays in memory, no idea how stable it is, so both ways are left for now → test flux model support
→ spectro analyzer for openai, google, and local tts
→ updated gallery
→ new music
→ nvidia/amd switch (instead of auto detection)
→ desktop: a tiny ‘easter egg’ on the main screen (5 clicks on the version)
→ koboldcpp updated to 1.96.2
→ empty input field now hides after 3 sec
→ tweaked system contexts, so chars should be a bit less horny (just a bit)
→ help section updated
→ comic mode fixed
→ openrouter: fixed prompt gen for image

0.5.13

bugfix, ngrok, ui-ux
→ stable diffusion:
– changed folder structure: files are now distributed into cuda, darwin, vulkan folders, models in the models folder
– default model converted to gguf: ~1-2 seconds faster on gpu and about 10% faster on cpu, build is 2.7 gb smaller (yay)
– added support for amd cards: app auto-detects the card and uses vulkan instead of cuda
– improved location detection and memory
– updated loras in the win-local build
– rebuilt for mac and amd cards; increased graph sizes
→ koboldcpp:
– updated to version 1.95.1
– koboldcpp remote: added ngrok support as an alternative to cloudflare
– fixed bots: for detecting outfit changes and location; to generate image prompts
→ added rounded mplus 1c font as default
→ added photo studio location
→ desktop: added button to hide character and text, hotkey shift+h
→ added grok 4 model (still unstable but interesting in places)
→ lora: tweaked system prompt for image generation
→ input field improved for easier entry of large multiline text
→ adaptive font sizing adjusted: text now fits more often without scrolling
→ help section updated
→ quick menu now shows only icons
→ updated icons for saving images and phone
→ settings: tabs now hide when clicking anywhere on the tab bar, not just the icon
→ fixed save-image button in nvl mode
→ removed gemini recommendation text from main screen
→ suggestions (quick replies): now use the last 4 utterances instead of the entire dialogue
→ openrouter: fixed error when selecting models from dropdown
→ cosmos: added cosmosrp-3.0 and cosmosrp-3.5 models; removed deprecated ones
→ anthropic claude: fixed empty “: ” message error with small bots; suggestions (quick replies) now work
→ cohere: added suggestions support
→ rewritten location detection bot: now triggers immediately after the trigger, not on the next “turn”

0.5.04: local image generation, stable diffusion, LoRA, gemini 2.5

→ local stable diffusion and LoRA models
→ button for saving images
→ added models: gemini-2.5-pro and gemini-2.5-flash, gemini-1.5-flash-002 removed
→ koboldcpp updated to version 1.94.2
→ person: a prompt field for the local image model has been added to custom characters
→ android: fixed fiona export to telegram bot
→ fixed a weird issue with gemini models where the last user message was duplicated
→ game commands like “we moved to…, react appropriately” are now aesthetically hidden

patch for amd gpu – unpack the archive into …\resources\stablediff\ and overwrite existing files – required for local image generation

the main points have been included in the in-game help section;
below are the things I forgot to mention:

→ for the default iLustMix model, use clip-skip = 2. if u get a black image instead of a picture, it means your selected model needs a different clip-skip value – try setting it to 1 or 2. I haven’t seen value 3 used, but I left it just in case

→ custom character settings now have an appearance field
(context for text models will be added later, there are some structure issues)

→ write custom character prompts and those from the quick generation field (button in the bottom right) strictly in english, stable diffusion doesn’t work with other languages

→ the clothing detection bot uses the same triggers already present in the game
→ the image prompt bot only looks at the last two messages from u and the character

→ on mac, there might be issues with image generation
if u’re playing on a mac, pls take part in this open poll (unfortunately, I can’t test it on a real device myself)

———

0.4.92: openrouter, Fiona catgirl, telegram bot, ui-ux, saiga nemo, bugfix

→ all remaining openrouter models added
→ Fiona got a catgirl outfit, updated failed images, system contexts rewritten and optimized
→ person: telegram bot for Fiona
→ a few new images for Jo
→ gallery: thumbnails added so it won’t lag on weak devices, unlock function fixed and optimized
→ many settings items are now collapsible, looks tidier, and hopefully this minimalism will scare off fewer new users •ᴗ•
→ win-local: default local model changed to saiga_nemo_12b.Q5_K_M
→ linux: the game now looks for gguf models in the resources/koboldcpp folder, ~ like on windows
→ android: fixed animation for the 6-sided cube
→ koboldcpp updated to version 1.93.2
→ help section updated
→ changed sounds on menu items and characters on the main screen
→ fixed and optimized animation on the start screen, UI elements adjusted
→ changed animation for scene transitions in the story mode (I really didn’t like the fade from black, so now there’s even more blur), smooth transitions for video inserts

telegram bot:
in Fiona’s user contexts, the “export to telegram” button creates a json file which u need to send to the opened bot in telegram.
security: uploaded json profiles and the bot’s chat history are encrypted before saving using the fernet scheme (AES-128 + HMAC-SHA-256).
as a result, even if someone steals json files from the server, without the key they’ll only see base64-garbage, not yr secret tokens.
in the future, the plan is to link the telegram bot history with the game, so u can seamlessly switch between the bot and the game while keeping the chat history.
if people like it, other characters might get bots too

saiga nemo:
I’ll keep changing the model until we find the best one •ᴗ•
it really seems multilingual – should understand and write basic spanish, russian, german, french, italian. english is about on par with gemma and deepcogito

0.4.81: local speech recognition, Dina, log editing, openrouter, bugfix

→ log: editing and deleting entries, button for quick scroll to bottom, when opening the log the latest entries are always visible
→ added support for openrouter. free models: qwen/qwen3-30b-a3b:free, tngtech/deepseek-r1t-chimera:free, meta-llama/llama-4-scout:free, meta-llama/llama-4-maverick:free, deepseek/deepseek-chat:free, deepseek/deepseek-chat-v3-0324:free
→ Dina: new images, start locations changed, now u can give her cunnilingus, squeeze her breasts, + images of blowjob
→ added models: claude-sonnet-4, claude-opus-4, claude-3-5-haiku-20241022, command-r7b-12-2024, gpt-4.1-mini; deprecated gpt-4-0613, gpt-4o-mini and gpt-3.5-turbo (16k version remains)
→ local speech recognition model (windows, linux)
→ now speech recognition settings are saved; at game start it launches the chosen solution (browser/local)
→ internal ru proxy: lv proxy server added (Latvia)
→ default local model is now deepcogito_cogito-v1-preview-qwen-14B-Q4_K_M, the list of recommended models has been updated
→ fixed location behavior: now a random one from the saved set is chosen instead of the first
→ fixed ui-ux animations and minor interface bugs
→ koboldcpp updated to version 1.92.1
→ now for windows, if u have a cuda 12 gpu u can put koboldcpp_cu12.exe into resources/koboldcpp (it will be slightly faster); the game will use it instead of standard koboldcpp.exe
→ default context trimming for local models increased to 9900 characters
→ changed the passwords for base, person, visionary

log: now u can edit dialogue history – entries for both character and yourself – which can help fight api censorship of large models and also just to have the history exactly as you want

openrouter currently without streaming output, as not all their models support it and the syntax is a bit different; I’ll add streaming later

internal lv proxy (ru lang): unfortunately it didn’t help for gemini since it’s also blocked in Latvia, but works fine for chatgpt and claude. thus for gemini the internal proxy works only through server 2 (sf). if that is blocked for you, the only option currently is your own vpn

local speech recognition: uses model ggml-large-v3-turbo-q5_0. it works a bit differently than the browser one: it can’t stream (but this is solvable; I’ll improve it), it waits until recording ends, so without a gpu it may be slow with high delay. to speed it up, you can try replacing the local model file in resources/whisper with a simpler model (e.g. ggml-tiny-q8_0.bin), just rename it to ggml-large-v3-turbo-q5_0.bin so the game picks it up

0.4.56: cosmo girls, spaceship, redesign, dnd space odyssey, bugfix
→ starship location
→ cosmo girl outfit for Jo, Dina, Fiona, and Lora; cosmo guy outfit for Finn
→ person: dnd scenario – space odyssey
→ koboldcpp updated to version 1.89
→ fixed bug where a duplicate character already on screen could be added
→ redesigned the model-selection block in settings – now more compact and slightly clearer
→ ru-lang proxy: the second kz server for anthropic models now works correctly

0.4.50: grok‑3, gpt‑4.1, hotfix, chibi Fiona, locations, Dina + Ava
→ added api models: grok‑3 and gpt‑4.1
→ added small chibi animations for Fiona’s surprise and irritation
→ new sandbox locations: sofa, bedroom, beach house
→ nvl mode: fixed text rendering for claude
→ fixed critical instruction‑injection bugs in the story for english lang and the female protagonist. because injections weren’t triggering, the english narrative was clearly patchy and in some places made certain branches inaccessible.
→ koboldcpp updated to version 1.88
→ fixed the in‑story phone‑close button
→ expanded Dina + Ava route if u choose to sleep over at Dina’s;
→ added branches for sleeping alone
→ Jo — a few new images and updates to previously unsuccessful ones
→ now characters always know which location they are in; the last used location is remembered and loaded into the system context. The favorite location can now be disabled by clicking the icon again (if history saving is enabled, I recommend disabling the favorite location)
→ added locations: police station and prison
→ added model command-a-03-2025 from cohere
→ cohere: fixed bots for in-game triggers
→ improved logic for deciding whether warm winter clothing is needed — should reduce false positives

0.4.41: music control, history saving, bug fixes
→ new images in the gallery
→ music control in the sandbox
→ plot: a small scene in the park with Ava (if u find the bench)
→ person: history saving in the sandbox
→ android: fixed an issue with koboldcpp-remote
→ android: fixed an issue with settings import
→ koboldcpp updated to version 1.87.2
→ fixed the long-standing issue with input windows that, under certain circumstances, could overlap each other
→ now clearing history also clears the message log of the selected character
→ optimized instructions for summarization (every n messages)

0.4.35: story branch, images by grok
→ added image generation by grok (option in settings)
→ story: added a branch with a police station (upon exiting the park)
→ when generating an image, the character now becomes smaller and shifts to the right
→ images for Ava
→ fixed error when the local model is selected
→ added proper error/notification output during image generation

0.4.27: story & bugfix

→ continuation of the story
→ animation on the start screen
→ micro-animation and clock sound after some time has passed
→ a few images for Jo
→ added the gemini-2.0-flash-lite model, discontinued gemini-1.5-pro-exp-0827, gemini-1.5-flash-latest, and gemini-1.5-pro-latest
→ koboldcpp version updated to 1.86 (supports gemma3)
→ fixed an error where closing the game and reopening it within 5 seconds caused saves and settings to disappear.
when using local models via koboldcpp, it is still recommended to wait 2-3 seconds before reopening the game
→ added information that a game restart is required when importing settings and save files.
→ fixed the music in the storyline.
→ changing the AI language no longer requires a game restart.
→ the sandbox setting that allows the character to speak first no longer requires a game restart.

0.4.05: gallery, mini-game, bugfix

→ gallery: added a multitude of images, now to open them you need to use ⦵, which can be farmed by winning mini-games: chess, save the princess and collecting macrodata
→ Jo, Laura, Fiona and Ava: now they can go anal (trigger and images)
→ Laura and Jo masturbate (trigger and images)
→ a gallery option has been added to the quick menu
→ mini-game with macrodata (the very same ones collected by the workers from the show – severance), available from the phone or by asking Ava
→ added model gemini-2.0-flash
→ groq: added model llama-3.3-70b-versatile, deprecated llama-3.2-90b-vision-preview
→ added park locations
→ fixed bug with english context when adding a second character
→ koboldcpp updated to version 1.84.2
→ removed the button for activating 3d from the story section
→ apk: fixed voice selection for elevenlabs
→ fixed bug with the car, when the location did not change after the car
→ person and base passwords changed

for “save the princess” the most points are given – 40
for chess (losing is almost impossible) – 10
macrodata – random from 1 to 18

macrodata: the idea was spotted and swiped from the expanses of codepen
for everyone who hasn’t watched the series (severance) – I highly recommend it

0.3.97: speech elevenlabs, 3d lipsync demo
→ speech synthesis from elevenlabs (also on apk)
→ sandbox: demo of 3d characters with lipsync
→ when first opening the game, the language selection for ai is now set correctly
→ selection of cosmosrp-pro and chatgpt-4/4o models is open and no longer requires a game subscription (without person password)
→ updated list of starting models on the welcome screen and in settings (if there is no local model in the game folder)

3d:
one of those ideas that came almost immediately but was put on the back burner my knowledge in 3d is quite superficial and not enough to create a model myself, so I used the free very limited editor ready player me, models are boring without any hint of sexuality, but together with lipsync demonstrate what I want to do

plan: find a suitable 3d artist on the marketplace and order a character based on refs from the game (later we will decide who will be first)

approximate level of the model that I want

if u know something that can help – write ) the model should be compatible with ready player me + arkit, oculus blend shapes

with the 3d avatar, only speech synthesis from google and elevenlabs works so far,
fps locked at 45 frames

speech synthesis from elevenlabs in my brief view is top: a large selection of voices, trial, slightly cheaper than openai, and no worse in terms of level

the continuation of the story is still in development. unfortunately, plans have shifted somewhat (there were many important bug fixes). the new deadline for continuing the plot is the end of february to the beginning of march

0.3.92: bugfix, locations, eos
→ sandbox: when changing locations from the menu, characters comment on the movement. you can set a favorite/start location for each character
→ added bypass_eos setting
→ updated koboldcpp to version 1.82.4
→ updated help section (hoping it’s more user-friendly for new users)
→ fixed auto-save: the slider now also affects automatic dialog trimming. the recommended value is set to 18, and the maximum reduced to 40
→ fixed steam activation in the shower when selecting locations from the menu
→ fixed AI language selection in dark theme
→ winter theme extended until February 13
→ fixed “rescue the princess” minigame (my crooked hands forgot to close a parenthesis in the code, damn it)
→ sandbox response options setting now works correctly with generated answer variants (it was broken and worked the opposite way, hehe)
→ story: now, at the end of a short series dedicated to the communication part, if a character asks a question, the answer is not replaced with an ellipsis but is fully displayed with the note “the question remains hanging in the air.”
→ koboldcpp logs are now always trimmed to the last 360 lines
→ when opening the sandbox, the “what happened before” button now correctly appears only if the character’s history was saved
→ fixed the order of save/load slots
→ fixed dice flying off the table in DnD: now there is an option to re-roll

bypass_eos – use this if the local model outputs “EOS token triggered” too frequently and inappropriately

there was a bug with auto-save – a hardcoded value of 20, which when a higher value was set in settings (now also affects trimming), was simply ignored and saving did not work

0.3.86: phone, ai language, bugfix, mac
→ phone call mechanics in the sandbox. it’s for the storyline part, and in the sandbox it’s used for testing, adds a bit more immersion
→ added ai language selection (Français, Español, Português, Deutsch, Italiano, Українська, Slovenský, العربية, Türkçe, 한국어, ภาษาไทย, Polski, 日本語, Tiếng Việt, ქართული, Հայերեն)
→ fixed the die for dnd modes to a d20, added sound, and made some visual adjustments to the buttons
→ if a local model is selected, the game no longer requires a network connection
→ fixed version checks and password checks when offline; in some cases—especially on smartphones—they could cause the app to freeze
→ person and base passwords changed

ai language: this only affects the text from chatbots. because the system context is written in english, sometimes the characters may slip into english, but in the sandbox this happens much less often, and it’s easier to get them back into the chosen language

the phone mechanic is new and may be buggy on some models

mac: for now, without support for the local model—I’ll try to integrate it next week
on my friend’s macbook, everything launched without additional permissions, but it might not be the same for u.
the prblem: apple requires the app to be signed, and it costs an additional $100 per year for that (greedy). without a signature, u might need to grant extra permissions in the settings to run the application

0.3.80: local model + koboldcpp
→ integrated koboldcpp with local model*
→ desktop: added welcome screen with choice of local model or via api
→ in the sandbox added location selection (upper left corner)
→ updated failed Ava images
→ redesign of selection buttons
→ reworked path to one of the sexual scenes with Dina in the plot
→ now if koboldcpp is selected and koboldcpp-remote tab kobold in settings is open by default
→ minimum screen size in windowed mode increased from 640×360 to 900×500
→ improved bots determining location changes and arrival of second character. fixed work on local models via koboldcpp
→ maximum temperature value lowered to 1.4, recommended to 0.72
→ fixed a bug with version checking
→ in the storyline the earthquake animation has been corrected (and possibly some others)
→ replaced music in the narrative event with fiona in the shower
koboldcpp:
→ now upon closing the game processes koboldcpp and electron terminate correctly and do not remain in the task manager
→ fixed premature token generation termination (error eos token triggered! id:2)
→ parameters adjusted
→ fixed the behavior of suggestions and generated options
→ maximum response length reduced to 630 tokens, recommended value up to 190
→ added a button to shut down koboldcpp
→ added information about approximate system requirements on the start screen and in settings (if there are no models in the folder)
→ added custom triggers for stopping

*now for windows there will be three options: standard (with koboldcpp), -local (with local model) and -lite (without koboldcpp). for linux u will need to download the model separately and select it in settings (later I will figure out the configuration and make it the same as on windows with the folder resources/koboldcpp)

windows: koboldcpp with models is located in the folder resources/koboldcpp, all gguf models from this folder the game will load on the start screen and tab in settings

windows/linux: a model from any folder can be loaded from the settings tab by clicking on “select model”

the local model where I stopped – Gemma-2-Ataraxy-v4d-9B.i1-Q4_K_M – by feeling ~ level of chatgpt-3.5, a bit weak, but much better than cosmosrp handles triggers-emojis, I think it should work fine for most users

among the models I tested, the one that might be best suited for the game—but is more demanding—is ChatWaifu_v1.4-Q4_K_M-GGUF. (If there are no models in the resources/koboldcpp folder, it will display system requirements and a link on the main screen and in the settings)