My ComfyUI workflows for using MiniMax H3

@Darksidewalker enjoy your break, thanks so much for this workflow!

@Darksidewalker I figured.

Same issue here!

you need to update KJNodes: https://github.com/kijai/ComfyUI-KJNodes

ComfyUI wont say it needs an update sometimes, so manually git it into the customnodes folder

@mrweaz ComfyUI-KJNodes updated, still MiniMaxH3MemoryEfficientSageAttentionPatch node is missing

@myprivacy27091991221 Same.

@hardwire666 Yeah I’m having problems with the director and director guide nodes

@myprivacy27091991221 yeah, i don't now how to fix that, i am having the same issue

@myprivacy27091991221 in that case I'm guessing you don't have sage-attention installed. Try installing that and see if it works.

@mrweaz thanks it was a couple of things missing on my end, had to use chatgpt to walk me thro cuda tourch and a bunch of other updates... i hope it didn't break my setup for something else lol

only hardcore CivitAI users use that browser and shove it down everyone’s throats, causing missing node errors to appear. I wish people would make separate WF without using that abomination of a browser.

@Darksidewalker Tested your updated node pack. I have to say, your node pack are damn amazing. Still, while I can use time placeholder, would still very much prefer to save the video with simple filename with sequential numbers for specific cases like 'XXX_00001, XXX_00002' like the comfyui default one.

This is the challenge. I have done this. :/

https://www.youtube.com/watch?v=peq4TFOxHW4

This video helped me just now when I ran into your same issue johansen. Verify your cuda, python, and torch version. Find the corresponding triton you need (uninstall and reinstall if you have an older version), and then find the correct sage attention wheel for your setup. Good luck!

Easiest way is to use DaSiWa's comfyui installer for a fresh portable install then just copy over your models folder. it auto-installs everything you need including triton and sageattention. Wish I had known about it way sooner.

ive been testing everything that has been suggested, SageAttn, Easy Cache, Sol-Attn and Spectrum. so far the only method that has given a good speed boost is Spectrum, the others are very minor (seconds) with a loss of quality that doesn't make up for the time you save. Spectrum is the only one worth using so far. ive replaced the EasyCache and SageAttn in the wf and it gens alot faster with minimal quality loss, i suggest everyone to do the same.

@Darksidewalker This is already awesome. Can't wait for the next update!

@Darksidewalker I had the opposite results for some reason, Spectrum was the best result I had gotten. Maybe it's hardware specific? I'll run some more tests with Sage and EasyCache and re test 😂

@Darksidewalker ill have to have another go, i may have been setting something up wrong. thanks for the info :)

they're on ComfyUI KJNodes Nightly :)

@Ertyer Still doesn't work. Even with a manual install using the command line.

@Darksidewalker I can't speak for OP but git clone didn't do it for me. I had to drop to torch 2.11 + CUDA 13 and then manually run

python -m pip install --upgrade "triton-windows>=3.6,<3.7

and

python -m pip install --force-reinstall --no-deps "https://huggingface.co/ussoewwin/Sage-Attention-for-Windows/resolve/main/sageattention-2.2.0%2Bcu130torch2.11.0-cp313-cp313-win_amd64.whl"

Seems to be working now.

git clone helped me. Switching to latest/nightly did not help.

@Ertyer thanks

Just to add onto this, I have been getting garbled/cut off chinese audio at the end of the generation versus the official workflow.

@Darksidewalker Thank you! Love your work as always btw! ❤️

@Darksidewalker I still get extreme harsh sounds instead of anything else :(

@Darksidewalker I figured what was hapenning, when installing the workflow I saved it before installing all nodes and for some reason the audio shift was at 0.01 instead of 4.00. It works like a charm now.

just to make sure, shift video should be at 11 and audio at 4 right?

There was a problem with this; change the scheduler to Simple. You probably have Normal or Beta set up, and I have issues with sound when I use those.

haven't tried ref2vid yet but here is the official prompting guide https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md

Yeah you drop them in, and prompt for <Subject 1> is the person in <Image 1> wearing a fantasy tunic with black hair. Then you would reference <Subject 1> in your prompt following the prompt guide like so:

integrated_multimodal_description: [Shot 1] <Subject 1> is standing in a forest filled with ash, he looks around as the camera pans to the left..

overall_soundscape: ...

non_diegetic_music: ...

@mrweaz cool but how we will see or tag the image? (for multiple images)

@Keithxx you do the same for <Subject 2> etc as long as you tell the model what each subject is it doesn't matter so much which order you place them in. But if you do <Image 1> <Image 2> etc then you need to remember what order you imported them in.

@mrweaz hm i did uploaded all at once, but ill use that way

I was able to use up to 9 images at once in this workflow.

@Acleveralias figured it out it was a ui bug, i changed the mode back and forth, fl then ref and it worked

If you add a video in V+A mode in the video lane as well as an audio in the audio lane, which of them is <Audio 1> ?

@maway the prompting guide on Huggingface states this:

Visual and Audio Tracks from the Same Reference Video

<Video N> and <Audio N> are numbered independently. Each index indicates only the label's order within its own category and does not encode a pairing between the two categories. The same reference video may therefore correspond to <Video 1> and <Audio 2>; different indices do not prevent them from coming from the same source asset.

An ordinary reference video does not create <Audio N> merely because the file contains sound.

An <Audio N> definition primarily states the audio's role and does not have to name the <Video N> it comes from. State the shared source only when needed to remove provenance ambiguity, for example:

<Video 1> is the source video for the target video edit. <Audio 2> is the synchronized audio track of <Video 1> and is reused in the target video.

like i said in a previous comment, as long as you reference what is what, you don't necessarily need to worry about linked numbers.

@mrweaz That doesn't make it clearer for me. In the base workflow, with the base node you are the one responsible of linking the video images to a video node and the audio of the video to an audio node. In this situation its very clear which number I have to refer them to. My question still stand. With this MiniMax H3 Director node, which of the audio land at #1 if I use a plain audio file + a video that has audio.

@maway it doesn't matter, you just need to tell the model what the audio is. but in your example of having a dedicated audio track and a video with an audio clip, it would be <Audio 2>. but really you should just be referencing or telling the model what audio clip is what.

so lets say i dont know what <Audio> tag to use if i have multiple sources, i would tell the model <Audio 1> is the song <Audio 2> is the dialogue from <Video 1> etc.

@Darksidewalker is downloading the GitHub as zip and copying it over the old kj nodes afterwards doing the pip install for the requirements.txt incorrect?

properly cloned it this time using cmd, though still missing MiniMaxH3SigmaShift

@Darksidewalker thanks for the trouble, still i have no MiniMaxH3SigmaShift node, only got the other minimaxh3sageattn from the install, maybe just out of luck

Updated comfyui as a whole and suddenly it appeared

@alopgamers how do you guys even do that? git pull doesnt solve anything

@LuckyCharmEr from the comfyui manager i just hit update all, after the restart it worked

Did you find a proper solution to this?
With updated nodes im stilling getting the error of the missing MiniMaxH3SigmaShift

You either have to update KJNodes from the GitHub manually, or you don't have Sage-attention installed.

Delete Kijai Nodes folder, update comfyui to the latest using .bat, git clone Kijai Nodes from its github, restart comfyui, it should download all the missing things. It's the only thing that worked for me.

@mrweaz Thanks, Sage attention is installed. Worked for LTX

@LuckyCharmEr Thanks a lot. That fixed :)

I just used the global and it was shockingly accurate:

https://civitai.red/images/138752630

you could try and connect a seed generator node to the subgraph and select fixed on generate.

Problem is only with sageattention, everything else works fine, im losing my mind

@Darksidewalker im using this one https://github.com/woct0rdho/SageAttention/releases/tag/v2.2.0-windows.post6

Something is not aligned. You can try asking AI to help. Or, do as I did: install a clean ComfyUI environment just for MiniMax. Easy install (not super fast, but easy) - GitHub - Tavris1/ComfyUI-Easy-Install: Portable ComfyUI installer for Windows, macOS and Linux with EZi Desktop app 🔹 Nvidia GPU support 🔹 Pixaroma Community Edition · GitHub

@ArchieMaser yea im starting to think about fresh install, but its kinda personal already xd

I deleted the memory efficient patch node and the flow has worked fine without out. Updated kjnodes and tried to re-add it with no luck.

I won, if anyone needs, you need to do a proper triton installation, and then put two folders from here https://github.com/woct0rdho/triton-windows#8-special-notes-for-comfyui-with-embeded-python in your python_embeded folder

Yeah getting the same error, tried everything including reinstalling Triton and Sage 2.2, nothing works for this node. Every other workflow using Sage 2.2 works fine, so idk what's up with it, a shame

@biggchungus1337 doesnt seem to be sage directly, just the memory patch. Bypass or remove the patch and sage just fine (At least on mine it did).

@srfuentes99 Ah i see, yeah that worked, and sage seemed to shorten the gen time as it should, thanks!

I'm thinking the Minimax director updated and now nodes aren't connected properly.

Ah, it was in conflict with other node I had installed. Is alright now, thanks.

And if I choose an aspect ratio which is identical to my image, e.g. square 1*1, it starts generating pixelated and distorted output...

Select Custom?

@skipwestcott as I said, it's set to custom (I2V) and 0.65 balanced. And the images that are not 16:9 (as I've tested) come out squished.

@skipwestcott and when I set a certain aspect ratio, they are no longer squished but now all blurry and distorted.

Ltx 2.3 was destroyed within minutes.literally everything "except speed" is better.and ostris working on a 4step lora.... so wait few hours or days ;)

Thanks to @Darksidewalker for creating a workflow.there will be more thinks to come from the community.i dont think anybody will use or train ltx.

@hartweizen whats crazy is with how strong the ref to vid is, you don't even need to train loras anymore.... its... an insane paradigm shift

LTX feels like the first version of Will Smith eating spaghetti compared to this. @Darksidewalker great work as always, man. Having an RTX3070 TI with gen times of 5m with strong prompt adherence and no need of loras is insane.

@nekulover24388 I have never been more motivated to scrounge together $1000 to buy a 5070ti

Not alone

Looking at my gallery of generated videos, I kinda feel like the first frame of every minimaxh3 output is already deformed, as if it's been processed thro something and downgraded in quality. The first frames of all of my ltx, wan videos are crisp and clear (and I am using your workflows for them btw). Could it be an issue with this specific workflow nodes?

so its probably a debateable topic right now as its so early days, but after experimenting since it launched i found that EasyCache dramatically decreases quality (for me at least, guessing its hardware specific) the only actual speedup method that works for myself that gives a good quality output is Spectrum. But Darkside did tests himself and found the opposite. so I'm guessing there's more factors at play. but i would give Spectrum a shot instead of EasyCache and see if it improves for you. again, early days so there's no definitive answer.

EDIT: Others on the Banodoco Discord have said the same thing.

@mrweaz  I disabled EasyCache early on cause it produced like major artefacts, but even with sageattention it messes up a lot of things, esp as I found if the initial image is very crisp like a real photo.

@LuckyCharmEr hmm i know that using IMG2VID for this model can produce artefacts on distant shots anyway, it handles close ups waay better. but anyway, i'm using this workflow but i swapped out the Diffusion Loader to Bob's INT8 Loader, EasyCache to Spectrum and my results are decent.

i have the same specs as u with sd preset takes 3 mins.

edit i "fixed" by adding the nodes from original workflow and connecting the first image to them., then connecting width and height node to the director.

If you don't want to do that on every update, a simpler way is to select CUSTOM aspect and in the backend you can link the Custom_aspect_width and Custom_aspect_Height from th DaSiWa Resolution Scale Calculator node to the frontend so you can easily change the resolution without going in the backend each time.

In the meantime the note node is your friend ;)

agreed, I finally fixed my sageattention that i've been putting off for months for this workflow :P

you can use the starting prompt: How the reference pictures align with the target video — <Picture 1> (from [Shot 1]) aligns with the 5.00-second mark of the target video. or whatever lenght you are using