MiniMax H3 Audio Lip Sync
Turn a single photo into a video of that person singing or speaking your audio — with lip movement that matches almost perfectly. This is a MiniMax H3 image-to-video workflow I put together for audio-driven lip sync: feed it a reference image and a reference audio track, and H3 animates the subject to the audio, mouth movement and all.
✓ 100% free — download the workflow and run it locally.Requirements
- NVIDIA RTX 30XX / 40XX / 50XX GPU
- CUDA-compatible GPU (~24 GB VRAM tested; lower may work with the memory nodes from the base post)
- At least 20 GB free storage
Model downloads
- minimax_h3_ref2va_pruned_int8_convrot.safetensors https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/diffusion_models
- qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/text_encoders
- minimax_h3_audio_vae_fp32.safetensors https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/vae
- Kijai's repo https://huggingface.co/Kijai/MiniMax-H3_comfy/tree/main/loras
Custom nodes
- ComfyUI Manager https://github.com/ltdrdata/ComfyUI-Manager
- ComfyUI-KJNodes https://github.com/kijai/ComfyUI-KJNodes
- rgthree-comfy https://github.com/rgthree/rgthree-comfy
- ComfyUI-VideoHelperSuite (Load Audio + Video Combine) https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite
- ComfyUI Nvidia RTX Nodes (RTX Video Super Resolution) https://github.com/xmarre/ComfyUI-Nvidia-RTX-Nodes
Setup steps
- Download the 4 model file(s) listed below and place each one into ComfyUI/models/diffusion_models, ComfyUI/models/text_encoders, ComfyUI/models/vae.
- Install the 5 custom node(s) listed below via ComfyUI Manager, then fully restart ComfyUI.
- Download the workflow .json from the button above and drag it onto the ComfyUI canvas.
- Open each loader node and confirm the model files are selected — a fresh install often defaults to the wrong one.
- Run the workflow. If a node shows red, update ComfyUI and restart before troubleshooting anything else.
Notes & tips
- What's Different in This Workflow?: Swapped the diffusion model to ref2va. The standard H3 image-to-video model is the first/last-frame (fl2va) build. This workflow uses the ref2va model instead, which is known to perform better for straight image-to-video — it holds the reference image's identity and setting more faithfully, which is exactly what you want when the face has to stay consistent while the mouth moves.; Added an audio input. A Load Audio node feeds your reference track into the generation, with the extra conditioning nodes needed to drive the lip movement from that audio.; Kept the turbo LoRA. It's still built on the lightx2v turbo LoRA, so you get fast 4-8 step generation rather than the full 20.
- Custom Nodes: ComfyUI Manager - https://github.com/ltdrdata/ComfyUI-Manager; ComfyUI-KJNodes - https://github.com/kijai/ComfyUI-KJNodes; rgthree-comfy - https://github.com/rgthree/rgthree-comfy; ComfyUI-VideoHelperSuite (Load Audio + Video Combine) - https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite; ComfyUI Nvidia RTX Nodes (RTX Video Super Resolution) - https://github.com/xmarre/ComfyUI-Nvidia-RTX-Nodes
- Load the provided workflow into ComfyUI and confirm the four models are selected — the ref2va diffusion model, the text encoder, and both VAEs.
- Load your reference image through the Load Image node — a clear, front-facing portrait works best, since the face needs to stay readable while the mouth moves.
- Load your reference audio through the Load Audio (Upload) node. Clean vocals with minimal background noise give the best lip sync.
- Match your duration to your audio length. Set the duration so the video covers your whole clip — H3 snaps to its 17-frame-per-block grid at 24fps, so it may round slightly.
- Match the output aspect ratio to your source image via the Resolution Selector. A mismatch degrades identity preservation.
- Keep step count at 4-8 (this is the turbo LoRA) — 6 is a good balance. Lower is faster but softer; the audio and lip sync hold up better a touch higher.
- Start with a short clip (the example here is 15 seconds) to confirm your settings before committing to a longer render.
- Best results come from good input on both sides: a sharp portrait and clean audio. Noisy audio or a low-quality reference image both show up in the output.
- More MiniMax H3 Workflows: FREE MiniMax H3 Video Generation Workflow & Guide (the base setup + memory guide) - https://www.patreon.com/TheLocalLab/posts/free-minimax-h3-165774435; FREE MiniMax H3 Turbo LoRAs (4-8 step speedup) - https://www.patreon.com/TheLocalLab/posts/free-minimax-h3-166073560
More resources
- https://get.runpod.io/Minimax-H3-ComfyUI
- FREE MiniMax H3 Video Generation Workflow & Guide (the base setup + memory guide) https://www.patreon.com/TheLocalLab/posts/free-minimax-h3-165774435
- FREE MiniMax H3 Turbo LoRAs (4-8 step speedup) https://www.patreon.com/TheLocalLab/posts/free-minimax-h3-166073560
- https://www.youtube.com/@TheLocalLab
- https://www.patreon.com/TheLocalLab
- locallabdigest.com https://locallabdigest.com
- https://www.locallabdigest.com
- Related video — MiniMax H3 video + audio walkthrough https://youtu.be/Idg1PqBcr5c
Want the 1-click version?
Skip the manual model downloads and node installs — get a ready-to-run installer, or unlock all 75+ with Local Lab Pro.