FastH3 Omni 8-Step Reference to Video - ComfyUI workflow (16 GB-friendly)
From: https://freakyjolly.com/best-fasth3-omni-comfyui-workflow-16gb/

FILE
  FastH3_Omni_8Step_Ref2Video_16GB.json  - ComfyUI UI/graph format. Drag it onto the ComfyUI canvas.
  Based on the official Comfy-Org template "video_fasth3_ominiref_r2v" (workflow_templates repo),
  with lower defaults for 16 GB cards: 0.5 MP (960x544, 16:9), 5 s (124 frames), one reference image.

CUSTOM NODES
  None. Every node is a ComfyUI core node. Update ComfyUI to a current build (late October 2026 or newer)
  so MiniMax H3 Reference to Video, ModelSamplingMiniMaxH3 (MiniMaxH3SigmaShift),
  Model Sparse Attention (BlockSparseAttention), Model Attention Backend, Resolution Selector,
  Math Expression, Create Video and Save Video are available.

MODELS (weights are NOT included; MiniMax-H3 Community License)
  ComfyUI/models/diffusion_models/  (pick ONE and select it in "Load Diffusion Model")
    fasth3_omniref_8step_pruned_int8_convrot.safetensors   22.2 GB  (default in this workflow)
      https://huggingface.co/FastVideo/FastH3-Omni-Comfy/resolve/main/diffusion_models/fasth3_omniref_8step_pruned_int8_convrot.safetensors
    fasth3_omniref_qad_nvfp4_comfy.safetensors              13.8 GB  (NVIDIA Blackwell / RTX 50 series)
      https://huggingface.co/FastVideo/FastH3-Omni-NVFP4-QAD-Comfy/resolve/main/fasth3_omniref_qad_nvfp4_comfy.safetensors
    fasth3_omniref_fp8_comfy.safetensors                    23.0 GB
      https://huggingface.co/FastVideo/FastH3-Omni-FP8-Comfy/resolve/main/fasth3_omniref_fp8_comfy.safetensors

  ComfyUI/models/text_encoders/
    qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors   15.7 GB  (default; accelerated on Blackwell)
      https://huggingface.co/FastVideo/FastH3-Omni-Comfy/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
    qwen3vl_32b_minimax_h3_int8_convrot.safetensors 27.1 GB (for non-Blackwell GPUs)
      https://huggingface.co/FastVideo/FastH3-Omni-Comfy/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors

  ComfyUI/models/vae/
    minimax_h3_video_vae_int8_convrot.safetensors  2.8 GB
      https://huggingface.co/FastVideo/FastH3-Omni-Comfy/resolve/main/vae/minimax_h3_video_vae_int8_convrot.safetensors
    minimax_h3_audio_vae_fp32.safetensors          0.6 GB (required for the audio track)
      https://huggingface.co/FastVideo/FastH3-Omni-Comfy/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors

INPUT
  Put your reference picture in ComfyUI/input/ and pick it in the Load Image node
  (the workflow expects a file named reference_1.png until you change it).
  Refer to it in the prompt as <Picture 1>.

KEEP THESE SETTINGS (model card + official template)
  8 steps, euler sampler, simple scheduler, BasicGuider (guidance 1.0)
  ModelSamplingMiniMaxH3: shift_video 12, shift_audio 3
  Model Sparse Attention: method vsa, keep_percent 10

Output: MP4 with synchronized audio, 24 fps, saved to ComfyUI/output/video/.
