ifroyou/minimax-h3-ref2va-sketch-p15b

MiniMax-H3 Ref2VA — P15B sketch polish

This public Space is a project-specific execution route for the Frieren P15B visual polish stage. It keeps the exact MiniMax-H3 ref2va / transformer_ref generation path and the pinned adapter:

• —adapter repository: Inner-Reflections/MiniMax-H3-Looping-Sketch-Anime
• —adapter file: minimax_h3_looping_sketch_anime_v1.safetensors
• —adapter revision: 9c88fbc800ea87d745137f1b637c08aa1a5e3bd6
• —default strength: 0.85
• —generation steps: 28
• —output audio: disabled

The adapter was serialized by the FL2VA/Kohya training route as lora_unet_* keys. This Space converts those keys before loading them into the live Ref2VA transformer: fused QKV B matrices are split into to_q, to_k, and to_v, and the two MLP groups are mapped to the MiniMax-H3 SwiGLU modules ff.net.0.proj and ff.net.2. The converter is structural. For the locked P15B route, the verified 0.85 delta is then fused into the CPU-resident Ref2VA transformer weights with PEFT safe merge; the adapter wrappers are unloaded before the first GPU worker starts.

At startup, the Space downloads the pinned adapter and injects it into the actual transformer_ref model, then removes the temporary adapter. The hidden p15b_preflight endpoint returns PASS only when all 600 converted tensors load into the real model. A generation request is not accepted as P15B-ready when this gate fails.

The P15B GPU route has a second admission proof that runs inside the allocated worker. It records the actual CUDA device name and total VRAM and refuses to denoise unless the allocation clears an 80 GiB lower bound for xlarge. This keeps the public runtime label from being mistaken for evidence of the per-request VRAM allocation. The final response includes this allocation receipt together with the fused-adapter execution mode.

The route accepts ordered reference images and one reference video. It is intended for the P15B runner, which supplies the locked START image, END image, identity/costume references, the P15A base clip as the motion reference, and the scene prompt. Project media is not stored in this public repository.

This Space is public because the user explicitly selected public visibility. The repository contains route code and provenance only; project endpoint images, clips, and prompts remain outside the Space repository.