Introducing: MINIMAX H3 | EZ WF
UPDATE: The workflow has been completely redone on 08/08/2026. if you previously downloaded and it didn't work for you, try now. i have tested in a fresh install, it works, if you are having issues you most likely don't have the most up to date version on Comfy from the branches. do NOT use the releases version, it will not work. Also be sure to update all of the nodes correctly, pull them from github. Check your console for any missing dependencies like sage-attention etc and install them. i will no longer be responding to any issues regarding node pack issues, you just don't have the updated versions or you have something in your setup that conflicts.
a custom workflow for Minimax's H3 video model. Its not too different to the base one Comfy provides, just has some little tweaks that i like to use, including Sage Attention/Turbo LoRAs and more!
This workflow is my personal workflow that I use for all of my Minimax H3 stuff, it may not be to everyone's taste but it works well for me and I'm sure others may find it useful.
Generated a 5 second clip with 0.6MP at 16:9 in 6:32 on an RTX 3060 12GB/32GB Sys RAM using this WF. base workflow was more than double that time.
The workflow uses the most optimal boosters there are right now (all optional, you can just run without them)
Works on 12GB VRAM + 32 GB System RAM. not tested on anything below.
INCLUDES:
3 Ref Images | 2 Ref Audio | 1 Ref Video
Sage Attention Patch (For Lower VRAM Use)
Block Cache & Sol-Attn with Optimised Settings
Bob INT8 Loader (Faster INT8 Loading)
LTX Refiner
RTX Upscaler
ITV/T2V/RTV Modes
Nodes Needed:
[Bob INT8](https://github.com/BobJohnson24/ComfyUI-INT8-Fast)
[Sol-Att](https://github.com/kijai/ComfyUI-SolAttn_triton)
[KJNodes](https://github.com/kijai/ComfyUI-KJNodes)
[rgthree](https://github.com/rgthree/rgthree-comfy)
[EasyUse](https://github.com/yolain/ComfyUI-Easy-Use)
[VHS](https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite)
[RTX Upscaler](https://github.com/Comfy-Org/Nvidia_RTX_Nodes_ComfyUI)
[LTXVideo](https://github.com/Lightricks/ComfyUI-LTXVideo)
Models Needed:
-- H3 MODELS --
[TE]
-- LTX MODELS --
TURBO LORAS:
[LIGHTx]
Troubleshooting:
I'm missing the LTX Nodes and i get this error, what do i do?
ImportError: cannot import name 'pad' from 'kornia.geometry.transform.pyramid' (G:\StabilityMatrix-win-x64\Data\Packages\ComfyUIClean\venv\Lib\site-packages\kornia\geometry\transform\pyramid.py)-- downgrade Kornia in python packages to 0.8.2
The RTX Nodes are missing, even though i have them installed?Go to your python packages and install: nvidia-vfx
I'm missing MiniMaxH3MemoryEfficientSageAttentionPatchUpdate KJNodes through Github (go to your custom_nodes folder and open a git bash and type: git clone https://github.com/kijai/ComfyUI-KJNodes.git
MAKE SURE YOU ARE UPDATED LIKE SO:
Kofi: https://ko-fi.com/mrweaz
Disclaimer: I hold no liability for any images produced, including their legality, ethical implications, or potential misuse. By using this model, you agree that all responsibility for generated content rests solely with you. Use responsibly and within legal boundaries.
Description
Added RTX Upscaler
Replaced SageAtt and EasyCache.
FAQ
Comments (40)
The Dasiwa link is dead.
надеямся и ждем)
Works fine for me 🤷
a bit of a noob here, but it says it is missing MiniMaxH3MemoryEfficientSageAttentionPatch?
Oh no probs, you need to Update KJNodes manually, you need the latest build. If you have done that and still.persists either update ComfyUI or you don't have sage-attention installed.
Also having this issue, I have manually tried updating the sageattention in the venv to 2.2.0 but even with it updated it, and tried KJNodes I switched branch to master and it was already up to date, sageattention is installed, ComfyUI is on latest version, having no luck
# ComfyUI Error Report ## Error Details - Node ID: 1512:2685 - Node Type: MiniMaxH3MemoryEfficientSageAttentionPatch - Exception Type: RuntimeError - Exception Message: RuntimeError: sageattention is not new enough version or could not determine CUDA architecture, cannot apply MiniMax H3 Memory Efficient Sage Attention Patch.
The workflow works without sage attention on, but it takes a long time like 780 seconds for 0.5megapixels 10 second video to generate I have an RTX-4070
@MrDel from a quick look at that message, you probably dont have CUDA installed, or wrong version.
@mrweaz Hey thanks for the reply, I was kind of following AI assist, but CUDA do I need to install it on my PC, or within the .venv? I have ComfyUI Desktop and from the progress I have made so far its all needed to be done within the virtual environment.
I have investigated the instance and it says it has CUDA 13.0, with PyTorch 2.10.0+cu130 and instance has Python 3.13.12, and sageattention 2.2.0+cu130torch2.10.0
from everything it seems like it is all a version match but just still get that error
@MrDel hmm have you tried to install the KJNodes from the GitHub directly? The error says a version mismatch is present. If that ain't it, your better off asking Kijai about it
@MrDel Make sure the whl that you're installing sage attention with has the correct python version.
For exampe, for your setup, you want somethingl like:
sageattention-2.2.0+cu130torch2.10.0-cp313-cp313-win_amd64.whl
("cp313-cp313" means python 3.13.x)
if you installed a wheel like -
sageattention-2.2.0+cu130torch2.10.0-cp312-cp312-win_amd64.whl
There will be a mismatch, since it wants python 3.12.x and you have 3.13.x installed.
Nice work! Fun to explore h3 with. Is there a way to add first/last frame on top of this? Too often it just ignores prompting for <image 1> to be the first frame.
Yeah this workflow is just for reference, if you want FFLF, change the model to the FL one, replace the MinimaxH3 reference node to the FL one and delete the IMG3/Audio nodes and plug everything back into the FL node. It's a quick fix, but yeah this WF was just for reference mode
@mrweaz Gotcha, thanks. Guess there's no way to combine both nodes yet.
@bugmaker86 there is, but you'll need to use switches and bypassers, I was planning on expanding this WF to make use of both, but a more advanced one dropped right after so I'm not updating this one anymore as it's not as advanced 😂
Id recommend you try the Dasiwa one, it's much better.
@mrweaz I'll check it out. 😉
6 minutes for 10sec at 12fps 320x600p on laptop rtx-A2000- 8GB vram , but loaded the 64GB ddr5 ram to 58GB/64GB! The results are definitely worth waiting for, compared to both LTX and WAN.
throw a RIFE interpolator in there and you are golden
Unable to get the "RTX" nodes working. import failed. Everything is up to date and of course i have an RTX GPU
try git cloning directly into your custom nodes folder, sometimes comfy can fail to import, always good practice to clone manually
@mrweaz That doesn't work for me, same errors, but thanks anyway. I'm going to do without these nodes.
@Error666 ah shame, yeah i have started to skip the RTX upscaler now anyway. hopefully an upscaler comes out for it soon
@mrweaz OK, resolved. I was missing a Python package. If anyone else has the same problem:
For comfyui portable:
.\python_embeded\python.exe -m pip install -U --no-build-isolation nvidia-vfx --index-url https://pypi.nvidia.com
@Yathukku Thanks, but I finally found a solution. I'll watch the video anyway—thanks.
@Error666 Oh nice, I'll have to try that. I've been trying to get that shit working for a WHILE
dude r u using 27GB Size INT8 Text Encoder ?
should i go for GGUF ?
i'm also using RTX 3060 12GB and 48GB System RAM
No GGUF is slower than INT8, INT8 works just fine on a 3060, I am yet to get any oom errors
okay
plz add Frame interpolation node.
i have tried some nodes. and set 15FPS as default for video Generation. but instead of 2x frames, video run time is tripled and FPS is same as i set like 15FPS/s and video running in Slow Motion
@Yathukku the problem with adding frame interpolation is that it will make the video cuts warped, i used to have it on the workflow but it wouldn't play nice. i do not plan on adding it again because of this, but if you want to add it yourself, throw a RIFE node before save video.
To experiment I tried your workflow as well and I get the same artefacts around faces as with DaSiWa... I cannot seem to reproduce something as clean as that archer video you attached, for instance. :_(
what resolution is that vid? and is it upscaled?
@LuckyCharmEr so i generated it in 0.7MP, 20 Steps 9:16 and used the RTX Upscale with deblur off, VSR Mode, Ultra settings and a 2x Scale.
if I'm honest, it depends on how close/far the subject is to the frame. if they are a distance they tend to have artefacts. try frame your characters closer and se if it helps.
i have also modified the Dasiwa WF to my personal needs so i may have something in there that you dont.
Also, having high quality reference images helps a ton. im using images i generated from Krea with my Krea 2 workflow and i have decent results with them, oh and prompts matter too. describe your references.
I have exactly the same problem with any workflow. The ref faces are full of artifacts , but other then the faces everything is perfect.
@cucurucu so after talking to someone on discord about it earlier, as they had bad gens over and over with the same settings, it was all down to the quality of reference images and the prompt, make sure you are anchoring your references correctly. Id suggest looking at the Minimax H3 reference prompt guide on their HF repo.
Or try different reference images. Mine was a bigger file I made from Krea 2 so it was nice and sharp
@mrweaz can you please provide a good promt used with ref images ? To compare what I am missing
@cucurucu subject_definitions: <Subject 1> is the young woman with dark long wavy hair, green eyes and large smile whose appearance and facial features come from <Picture 1>.
@mrweaz strange, that's a simple description - I've been doing this and then a more detailed descriptions and still the artifacts. Did this worked for you ?
@cucurucu so for the archer video, I had a high quality image as a reference, without her face, the video made the face on its own. It started as a side view image. Maybe that has something to do with it?