#Update
Fixed several issues in the workflow and refined some settings.
Added Add Guide for MMH3
Improved timeline control for reference images and audio.
Added full support for T2VA, I2VA, FL2VA, L2VA, and Ref2VA; added reference video and audio uploads; updated the interface.
Minor workflow adjustments and improvements.
Minimax H3 workflow using Turbo LoRA for the fastest generation while maintaining high quality, with Speedup enabled and 8 sampling steps.
# Required Custom Nodes
Install the following custom nodes using ComfyUI Manager or clone them manually:
- ComfyUI-SolAttn_triton
- ComfyUI-Spectrum-MiniMax-H3
- ComfyUI-KJNodes
- ComfyUI-Custom-Scripts
- ComfyUI-Easy-Use
- ComfyUI-Workflow-Encrypt
- CRT-Nodes
- ComfyUI-VideoHelperSuite
- rgthree-comfy
- ComfyUI-VFI
- Nvidia_RTX_Nodes_ComfyUI
Description
Added full support for T2VA, I2VA, FL2VA, L2VA, and Ref2VA; added reference video and audio uploads; updated the interface.
FAQ
Comments (16)
Did you delete your latest post? i was checking it out live, and then it was gone.
I need to fix the workflow first. I’ll let you know once I’ve uploaded the updated version. Thank you for reporting the issue.
I’ve fixed the workflow and uploaded the updated version. Thank you again for reporting the issue.
@Caphaomuoi np, thanks for the wf!
This workflow is solid and very powerful, but I’ve spotted two issues. First, the "Video Combine 🎥🅥🅗🅢" node you're using is terrible; I suggest replacing it. It generates three files at once—two of which are useless and can't be disabled—and doesn't allow for a preview after the work is done. The second issue is that the "seed" is fixed, so generating repeatedly will just produce the exact same video over and over.
You can right-click the Video Combine node and select “Resume Preview” to preview the result after processing. I disabled it by default to reduce RAM usage, especially for lower-end machines.
As for the fixed seed, you’re right — I’ll change it so repeated generations don’t produce the exact same video.
The Video Combine node also outputs an additional PNG file containing the metadata. The video without audio is usually used as a reference file. Thank you.
I’ve changed the seed mode to Random now
I’ve already made the adjustments myself; I’m just letting you know about the issues I encountered. Although generating three image files and a video at once helps with VRAM usage, it makes organizing the output difficult. That said, your workflow is excellent—I really like it!
@a0978750321122 Thank you for the feedback, I really appreciate it!
You can stop Video Combine from writing extra files in ComfyUI → Settings. Search for VHS, then open:
🎥🅥🅗🅢 › Output
Turn these off:
Save png of first frame for metadata — stops the extra first-frame PNG
Keep required intermediate files after sucessful execution — deletes intermediates after a successful run (including the silent video when audio is muxed)
Both default to on, so extra files stay unless you disable them.
@jayjay881 Yes, most complex nodes have a question mark (?) in the top-right corner. The author provides very clear documentation on how to use the node; people just don’t usually pay attention to it.
I'm getting a weird persistent behavior where the camera pans to the right no matter how much I prompt for a steady shot. Happened across 6-7 different ref2va attempts. Speed and quality are great, but kind of sucks if the subject is almost completely out of frame by the end of the scene.
Solved, I swapped out the text encoder for a different one and the problem went away.
Please share the prompt you’re using, or provide a link to your video so I can take a look.
great workflow, learning a lot!
issue: both turbo lora strengths are labelled ref2v and inside the workflow they are linked to the opposite model they are intended to control
You may have jumped to a conclusion a little too quickly. I checked again, and the cables are actually connected to the correct positions. The workflow uses two separate switching mechanisms: the If/Else Switch controls whether the LoRA is enabled or bypassed, while the Any Switch selects which model branch is ultimately passed to the output. Thank you.
There is also a full screenshot showing the entire outer layer of the workflow in the showcase pictures.