What is this?
A custom ComfyUI node that loads a MiniMax H3 (audio+video DiT) checkpoint by merging selected tensor groups from a second 'overlay' checkpoint onto a 'base' checkpoint, producing a hybrid that ComfyUI treats exactly like a normal loaded model.
A custom ComfyUI node that loads a MiniMax H3 (audio+video DiT) checkpoint by merging selected tensor groups from a second 'overlay' checkpoint onto a 'base' checkpoint, producing a hybrid that ComfyUI treats exactly like a normal loaded model.Is this a standalone app or does it plug into something else?It is a custom node (plugin) for ComfyUI, not a standalone program — it adds a new loader node to ComfyUI's model/loaders category.
It is a custom node (plugin) for ComfyUI, not a standalone program — it adds a new loader node to ComfyUI's model/loaders category.Do I need specific checkpoints to use this?Yes — you need both the MiniMax H3 fl2va and ref2va safetensors checkpoints, since the node specifically merges tensor groups between those two models.
Yes — you need both the MiniMax H3 fl2va and ref2va safetensors checkpoints, since the node specifically merges tensor groups between those two models.Does loading two checkpoints double my memory usage?No — both safetensors files are mmap-backed and streamed tensor-by-tensor, so peak memory stays at roughly one model's worth (~19.5GB for int8 checkpoints), the same as the stock loader.
No — both safetensors files are mmap-backed and streamed tensor-by-tensor, so peak memory stays at roughly one model's worth (~19.5GB for int8 checkpoints), the same as the stock loader.What's the recommended hybrid setup?Use fl2va as the base and overlay only the per-block AdaLN modulation weights from ref2va for blocks 25 through 49 — the author found this balances output quality with retained reference-conditioning ability.
Use fl2va as the base and overlay only the per-block AdaLN modulation weights from ref2va for blocks 25 through 49 — the author found this balances output quality with retained reference-conditioning ability.What does the 'none' preset do?It performs pure base-checkpoint loading, equivalent to ComfyUI's stock UNETLoader, with nothing taken from the overlay checkpoint.
It performs pure base-checkpoint loading, equivalent to ComfyUI's stock UNETLoader, with nothing taken from the overlay checkpoint.Does using this modify my original checkpoint files?No, both safetensors files are opened read-only and are never written to or mutated on disk.
No, both safetensors files are opened read-only and are never written to or mutated on disk.Why do the two MiniMax checkpoints need hybridizing at all?fl2va and ref2va share over 97% identical weights, but fl2va has higher output quality while ref2va alone has a known training-quality issue despite adding multimodal reference conditioning — blending lets you keep fl2va's quality while recovering ref2va's reference capability.
fl2va and ref2va share over 97% identical weights, but fl2va has higher output quality while ref2va alone has a known training-quality issue despite adding multimodal reference conditioning — blending lets you keep fl2va's quality while recovering ref2va's reference capability.What are similar projects?LTX-Video, LTX-2, ComfyUI-Ref2VA-VSA, ComfyUI-YCNodes-MiniMax-H3, CogVideo, SolarWM, LLaDA-Image, FreeVideo
LTX-Video, LTX-2, ComfyUI-Ref2VA-VSA, ComfyUI-YCNodes-MiniMax-H3, CogVideo, SolarWM, LLaDA-Image, FreeVideoWhich topics is it in?model, comfyui, diffusion-models, plugin-system, quantization, multimodal, video-generation, transformer, audio, generative-ai, pytorch, plugin
model, comfyui, diffusion-models, plugin-system, quantization, multimodal, video-generation, transformer, audio, generative-ai, pytorch, plugin