V5 Updated
I added sam 3.1
V6 Updated
I added a mask uploading node as well as a onboard mask creating feature that allows you to mask any area right inside comfyui
Original Post
Tutorial:
https://www.reddit.com/r/StableDiffusion/s/gl3I2Bqwzu
v5 update:
https://www.reddit.com/r/StableDiffusion/s/gUhUixgR6w
(I'm copy pasting the reddit post but the link is there because the comments are valuable)
Some guy on civitai made this workflow and i gave him some valid criticism, then he called me a gooner that doesn't know anything about workflows and blocked me. This offended me because I know plenty about workflows.
I'm not going to share his name because he's apparently active on reddit under a similar name but I tried to point out some problems with his workflow. So, instead, I just decided to fix them fueled by pure pettiness.
Here it is.
https://github.com/roycho87/minimax_wf
After dissecting the thing I was able to get many of the features that weren't working in the original to work and I added some features as well like the ability to force audio from video, fps control, and a centralized control panel that handles every feature universally.
Enjoy.
[Workflow Share] MiniMax H3 all-in-one workflow
Sharing my current MiniMax H3 ComfyUI workflow. The main goal was to make H3 easier to use by centralizing the important controls and automating the more annoying reference, continuation, audio, and post-processing routing.
Major features
Centralized control panel for the main H3 generation settings and workflow options.
Multi-reference support — up to 6 image refs, 3 audio refs, and 2 video refs.
Mixed reference types — image, audio, and video references can be used together in the same generation.
First-frame / last-frame control using reference images.
Video continuation with overlap-based stitching back into the original clip.
Continuation-aware audio handling for the source video and newly generated section.
Force Audio from either an audio reference or the embedded audio from a reference video.
Trim generation duration to audio length automatically.
Final latent upscale / refinement pass that can process the completed stitched continuation.
Built-in RIFE frame interpolation.
Sparse-attention / low-VRAM controls, including chunking and attention options.
Multiple LoRA support.
Automatic reference routing based on how many image, audio, and video references you enable.
The main idea is to spend less time manually bypassing, reconnecting, and rerouting parts of the graph whenever you want to switch between reference generation, audio-driven generation, continuation, or final processing.
Load your refs, choose the options you want, prompt, and queue.
Edit: if you get errors when trying the workflow make sure you upload placeholder images. The workflow is designed so you don’t have to bypass anything manually. You just need to use the control panel.
Edit2: V2 is updated and the issue of the final output being the first pass instead of the upscaled has been fixed. I also removed the shift and added a subgraph that you can hook in to use a turbo lora if you want with the recommended shift.
Description
FAQ
Comments (18)
I came in for the description of the workflow, stayed for the tea!
DM me the name of the loser 😂. Will definitely download and take a look at how someone builds a "user-friendly" workflow. Cheers ❤️
check the github
"foxydits_sucks_at_making_workflows_v2"
Any time I try to use seamless continuation I get the same error:
[ERROR] !!! Exception during processing !!! shape mismatch: value tensor of shape [584, 32] cannot be broadcast to indexing result of shape [656, 32]
The error numbers change based on how many overlap frames I seem to pick.
it needs to be 5, 22, 39 etc. i suggest leaving at 22. The video also needs to be long enough. like at least 3 or 5 second duration.
Hello.This is the best wf ive see.. Seedhunter has also great but yours i dont know why.. i can do on my rtx 5070 +32GB ram ddr4 from 0,4mpx with 8 steps + 8 steps upscaler to 768p. On other wf's i get OOM but i dont know what magic trick u used in yours wf.. good job ! also i have a request for your maybe next update to wf.. can u add node "Model Preview override" ? i rly try to do on my own but im too dumb and cant figure out how to install to your wf xD.. thx again for awesome wf.
this is the seedhunter wf but it's fixed which is why you're enjoying it most likely.
seedhunter is not a good wf. i removed override because previews are already available by default through comfyui. if you're not seeing them it's because your settings are wrong.
i'm glad you're enjoying the workflow.
Wait........... this is the easy Workflow?
Offf!!!!^^
I can't even figure out where to put Picture1 xD
Is this reference only???
What's going on :P
Do you not see the massive load image node labeled <picture 1>?
Here’s a tutorial
Lol @ the seedhunter workflow comment
Really enjoy the improvements of the Seed Hunter workflow, good job! But we need a better solution for Upscaling seamless video. The problem: Due to the Upscaling the next video I generate with seamless video function small details and whole video quality changes slightly. Any solution or idea?
Just turn the denoise down. .2 or .25 should be fine for basic details. .4 is rather high, i didn't think about that when I saved the workflow.
If I only do ref2va, should the setting "<Picture 1> is First Frame" be true or false? What does that setting do exactly?
if you're using fl2v model you can enable that to change the picture 1 to first frame and picture 2 to last frame, if you're using ref leave it false
I tried seamless continuation with 2pass. It doesn't work correctly. My clip 1 is done using 1pass 1.5MP. My clip2 uses seamless continuation and 2pass 0.3 -> 1.5 MP. When I put both clips in Davinci Resolve, I notice clip 2 has the image zoomed in closer to the camera than the clip 1. They should seamlessly continue because the dimensions (1664x928) are the same, but they aren't.
if you're rendering at 1.5 MP you can't upscale to a megapixel. you'd have to render at a lower resolution, if you're already at 1.5 MP you don't need to upscale it.