Ref2VA VSA Cuts MiniMax H3 Video Time

The clever part is the attention-layout patch, not another model checkpoint; it makes reference-conditioned H3 video substantially more practical locally. The speedup is compelling, but the 9× figure depends on a specific four-step workflow, GPU, and CUDA kernel stack.

• –Keeps text, image-reference, and audio-reference tokens dense while pruning video-only attention tiles.
• –README reports approximately 13.5 GB peak VRAM, though the documented setup still targets 24 GB-class GPUs.
• –Comparisons against 50-step native H3 are useful end-to-end, but not directly equivalent in sampling configuration.
• –Requires current ComfyUI H3 support, comfy-kitchen CUDA Sol-Attention, model weights, and a compatible NVIDIA GPU.
• –Apache-2.0 licensing makes this a useful building block for local video-generation workflows.