SMD-MiniMax-H3-Ref2VA-8step-ComfyUI

8-step SMD reference-conditioned generation with synchronized audio, using native ComfyUI nodes. Includes the converted LoRA, ready-to-run workflows and reproducible example inputs.

Original checkpoint · SMD code

Generated video with audio · Prompt and inputs

Run

New to ComfyUI? Download this repository with the first three commands below, then follow SETUP.md for the tested Linux/NVIDIA installation. If ComfyUI is already installed, run all four commands using its Python environment. No custom nodes or SMD source checkout are needed.

python -m pip install -U huggingface-hub
hf auth login
hf download zjwfufu/SMD-MiniMax-H3-Ref2VA-8step-ComfyUI --exclude '*.safetensors' --local-dir SMD-MiniMax-H3-Ref2VA-8step-ComfyUI
python SMD-MiniMax-H3-Ref2VA-8step-ComfyUI/download_models.py --comfyui /path/to/ComfyUI

The installer verifies the model files, copies example inputs, and adds the complete workflow to ComfyUI's default user library.

1. Start ComfyUI and open its browser interface (normally http://127.0.0.1:8188).
2. Click Workflows in the left sidebar, expand SMD, and double-click H3-Ref2VA-8step-Compact. Refresh the page if it was already open. Alternatively, drag the downloaded workflow.json file onto the canvas.
3. Confirm the workflow contains Load LoRA and Save Video, then click Run with the supplied settings. This loads the complete graph; no manual node connections are needed.
4. When the job completes, play the result in Save Video and enable sound. The MP4 is also saved under ComfyUI/output/.

The example generates about five seconds of video with audio. After your first successful run, edit the prompt or the seed in RandomNoise. To switch between 4 and 8 steps, install and open the corresponding repository's workflow and LoRA.

Use this repository's complete workflow. Substituting H3 weights into ComfyUI's default image template can fail with errors such as audio_scale missing. See setup and troubleshooting.

The two source images condition subject appearance independently; replace them in the Load Image nodes and keep their order aligned with <Picture 1> and <Picture 2> in the prompt.

Model profiles

Profile	Workflow	Downloads
Compact (default; example above)	workflow.json	42.0 GB
Unpruned BF16 transformer	workflow_bf16.json	101.2 GB

The compact profile uses the official pruned INT8 transformer, NVFP4-AWQ text encoder and INT8 video VAE. The BF16 profile uses an INT8 text encoder and FP16 video VAE. Both use the FP32 audio VAE and the same BF16 SMD LoRA. Shared files are downloaded once and verified by SHA-256; the base files come from Comfy-Org. Select the larger profile with download_models.py --profile bf16 --comfyui /path/to/ComfyUI.

Both profiles were run at 1344 × 768, 124 frames, 24 fps. The compact workflow was tested on an A100 80 GB with a reduced memory budget, using about 24.3 GiB at its observed peak. Physical 24 GB consumer GPUs have not been tested. Test details.

Keep LoRA strength 1, Euler, BasicGuider (no CFG), and the supplied ManualSigmas. The 8-step schedule uses video/audio shifts 12 / 3 and a raw final-interval scale of 0.5. Do not add H3 Lightning or Turbo LoRAs.

The adapter conversion preserves source LoRA values while adapting fused attention and SwiGLU layout. Conversion details · MiniMax H3 Community License.