Multiic [0.8.19]

→ Tisha’s story event has been completely reworked; u can find it in story navigation under “therapy session (Tisha)”
→ Tisha is now a full sandbox character (woohoo); you can also visit her for a therapy session in sandbox mode
→ use jinja – a new toggle in settings. launches koboldcpp with the –jinja flag, so utility requests (image prompts, clothing, detectors, etc.) use the model’s native template
→ added the ability to generate up to 3 images from a single prompt; the setting has been moved to the “image prompts” section
→ the game is now fully translated into french, german, spanish, chinese, ukrainian, and portuguese (br)
→ 4 sweet new transition effects for background changes and generated images
→ the default text model has been changed to Qwen3.5-9B-The-Defiant-Fable-Uncnr-Heretic-NEO-MAX-MTP-Q4_K_M.gguf
→ “continue” story button on the main screen
→ the hyperwoo minigame has been updated to 0.1.35; completing each level now rewards 20–30⦵
→ Mika (180⦵) and Tisha (135⦵) can now be unlocked in sandbox before they appear in the story
→ new gallery items for Tisha, Finn, Kas, Fiona, Laura, and Dina
→ reworked some of Mika’s weaker story images from the first encounter
→ fixed image prompts (and other small utility bots) for certain models, especially gemma
→ server-lv, previously used to bypass blocks, has been retired (due to sanctions, beget’s latvian server stopped working); all of its functionality has been moved to server-kz
→ changed ua to uk for language AI, in accordance with the ISO 639-1 standard
→ claude-sonnet-5 is now selected correctly from the recommended models
→ fixed several story dead ends
→ redesigned the main screen and context settings. language dropdowns have also been standardized
→ koboldcpp updated to 1.119
→ added rtx 50-series gpu support for local speech synthesis/recognition and stable diffusion; older gtx 10-series cards (1050–1080) are no longer supported and will fall back to cpu. the same will happen if yr nvidia driver is outdated – version 580 or newer is required
→ coqui: torch cu121 → cu130; fixed voice cloning – the reference recording no longer falls back to a preset voice; the voice list also no longer requires restarting the game if u switch to coqui while the game is running
→ qwen-tts: torch cu118 → cu130
→ updated the coqui and qwen links in settings and help
→ whisper and stable diffusion rebuilt with cuda 13.3
→ stable diffusion: updated the engine; added support for wan, ltx, qwen-Image, z-Image, chroma, and new samplers; “hot model” now works differently – when enabled, the model stays loaded in vram; when disabled, it stays in system ram, leaving vram free for other tasks; the build is also smaller now; image generation on nvidia gpu has been fixed on linux
→ code refactoring
→ local text models:

• fixed reasoning, system instructions, and service tokens like user / assistant appearing in character replies
• improved compatibility with Qwen, Mistral, Gemma, and other GGUF model families
• both regular and streaming responses are now more stable
• fixed connection to remote KoboldCpp instances
• bypass eos is now disabled by default; cases where it is actually needed have been reduced to a minimum
• KoboldAI has been deprecated — it hasn’t received updates for two years, and its own developers recommend using KoboldCpp instead

→ image generation models:

• your self-description now reaches the model completely untouched
• more room for descriptions: “describe yourself” — 180 instead of 90, character appearance in character cards — 450 instead of 270
• prompts are no longer blindly truncated. the limit now depends on the model, and if space still runs out, styling gets trimmed first instead of the actual scene: pose, action, and camera angle are never discarded
• scene generation is now more accurate. the model separately determines who is in the frame, what is happening, the shot type, and the lighting — instead of processing it as one continuous line
• fixed counting people in the frame: extra people should no longer appear, and required characters should no longer disappear
• character appearances are preserved better between characters, with less feature mixing and fewer traits from one character leaking onto another. distinctive traits of characters who are not in the frame are now automatically excluded
• the scene location is no longer replaced by the conversation context
• fixed cases where, after a failed model response, service text could end up in the image prompt instead of the actual scene description
• you no longer need to specify your age or gender in “describe yourself” — they’re taken from your date of birth and the main screen settings. if the field already contains your age/gender, duplicates are automatically removed from the final prompt
• new “don’t use” toggle next to the “describe yourself” field: excludes your appearance from generated images and fades out the field. useful if your features start bleeding into the character
• when generating multiple images at once: the first shot is framed however the “director” intended, the second is a close-up, and the third is a wide shot. previously, images in a series mostly differed only in noise/grain and ended up looking very similar; now they’re different shots of the same scene
• the tooltip for the “describe yourself” field has been rewritten and now explains all of the above
• comfyui: added support for PrimitiveStringMultiline / PrimitiveInt nodes for Krea Flux models

→ fixed the layout of the API models section on the settings page
→ added a progress bar when unlocking/locking all gallery items
→ changed Tisha’s starting locations in Sandbox
→ fixed location synchronization in Sandbox between the text and image models, including the office location
→ reworked the way characters are called into a scene from dialogue: this is now handled by a separate bot that knows character card names, understands exactly who was called, and ignores vague requests; duplicated dialogue lines when a character arrives are also gone

! bypass eos is no longer a recommended option. it’s better to disable it, since local models now correctly understand when they should stop generating. with it enabled, the model will keep generating until it reaches the response limit

the new text model replacing Saiga Nemo has fewer parameters (9B vs 12B), but in practice it’s smarter, faster, and better tuned for natural roleplay dialogue

generating up to 3 images from a single prompt will be especially useful with local models at a slower game pace, letting each moment play out more fully and making playthroughs feel more varied