Main features:
- MiniMax H3 VDN workflow for high-quality second-pass product showcase hero shots.
- Uses the OpenVDN/vdn-minimax-h3 model path with a low 8-NFE pass and a high 4-NFE refinement pass.
- Includes learned latent 2x upscaling toward approximately 1920x1088 output.
- Uses separate video and audio VAE components and preserves or audits generated audio according to the connected H3 nodes.
- Supports image references, a source video input, frame/FPS inspection, trimming, and an approximately 1080p 24 fps AV output.
Suggested workflow:
1. Load the source video and reference images required by the graph.
2. Run the low-resolution VDN pass, then apply the learned latent upscale.
3. Run the high-quality VDN refinement pass and the connected audio conditioning/reconciliation chain.
4. Review the decoded preview and save the final AV result.
Dependencies and limitations:
- Requires the MiniMax H3 ComfyUI custom nodes and the referenced H3 VAE, CLIP, UNET, and latent-upscaler files.
- The graph is configured around the supplied model filenames and a 24 fps, approximately 1080p output path; changing these assets may require node adjustments.
- Any bypass or side branch not connected to the final output is not treated as an enabled feature.
RunningHub:
https://www.runninghub.ai/zh-cn/post/2107778471853326337?inviteCode=rh-v1111
Bilibili:
https://www.bilibili.com/video/BV1AsHy6nEPR/
RunningHub registration:
https://www.runninghub.ai/?inviteCode=rh-v1111