Vellium 1.9.5 adds Lite mode to run visual novels on 12B local models

Details

Vellium 1.9.5 introduces Lite mode, which replaces structured JSON generation with plain text markers to make visual novels work reliably on smaller local models such as Gemma 12B.
System prompts are reduced from ~8k to ~1k characters, and text streams directly into the reader without waiting for the full response.
Prompt cache preservation is implemented by appending new turns to previous context, avoiding unnecessary reprocessing that can take over a minute on slower local setups.
Inline image generation allows ComfyUI to start rendering images while text is still streaming, including when ComfyUI runs on another machine.
Character consistency is improved by using compact English visual descriptions and pinned outfits instead of passing entire character cards to the image model.
Backgrounds and sprites mode reuses generated backgrounds and cut-out sprites, with background removal handled via chroma-key processing in-app to avoid loading additional neural networks into VRAM.
The update includes 12 new visual style presets optimized for anime image models, tested with Anima, and integrates with user personas, RAG knowledge collections, and Writer projects.

Key point