Run a finished H3 clip back through the model. Every part of the frame came back sharper, and the fastest moving part gained 99%, measured on all 124 frames.
It saves two clips, the normal render and the second pass, so you can compare.
Watch the video: https://www.youtube.com/watch?v=0n8uQ6__vzY
What you need
- ComfyUI with MiniMax H3 support
- ComfyUI-MAINodes v1.1.3 or newer (Manager or git clone)
- minimax_h3_fl2va_pruned_int8_convrot.safetensors (19.5 GB) in models/diffusion_models
- qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (14.6 GB) in models/text_encoders
- minimax_h3_video_vae_fp16.safetensors (4.9 GB) in models/vae
- minimax_h3_audio_vae_fp32.safetensors in models/vae (without it the clip has no sound)
- The LightX2V four-step adapter in models/loras (the only new file if you already run H3)
How to use it
- Unzip, put the sample picture into ComfyUI/input.
- Load the workflow and check the five loaders point at your files.
- Queue it once. Compare the two clips it saves.
The step-by-step guide, the video and more free workflows are on my portal.
