Open-source Live2D avatar pipeline – 80-95% cost reduction using Gemini (gist.github.com)

NanoLive2D is an open-source pipeline that turns text descriptions (or optional reference images) into Live2D-ready clothing textures in 3–5 seconds using Gemini 2.0/2.5 (Nano Banana). It integrates with Live2D (PIXI.js v7 + Cubism SDK) to apply generated textures to existing rigs at 60+ FPS on phones, supports real-time Q&A via a connected knowledge base, and handles speech→text→KB lookup→text→speech with lip sync. The result: setup that once took 4–5 weeks and $50K–$200K can now be done in less than a day with estimated 80–95% cost savings and no heavy GPU hosting for texture generation. Technically, the system solves a key challenge—image models often hallucinate whole characters and corrupt texture atlases—by using a two-stage flow: optional reference analysis (gemini-2.0-flash-exp) for style inference, then a constrained “Nano Banana” text→texture step that enforces the exact texture layout and only modifies specified torso/clothing regions. Example prompts explicitly instruct the model to keep hair, face, limbs, and atlas positions unchanged, producing a new texture sheet that fits the rig. The stack includes 26 motion files, sample model JSON, and REST API calls to Gemini/GBase; it’s ideal for rapid avatar customization (not for cinematic 3D or photoreal rendering).