This is an SVDQuant int4 conversion of my CreArt-Ultimate Hyper Flux.1_Dev model for Nunchaku.
It was converted with Deepcompressor at Runpod using an A40.
It increases rendering speed by 3x.
You can use it with 10 steps without having to use Lora Turbo.
But 12 steps and turbo lora with strenght 0.2 give best result.
Work only on comfyui with the Nunchaku nodes
This int4 version does not work with rtx 5000
Description
svdq-int4-CreArt_Ultimate Version merged for nunchaku 0.3.0
FAQ
Comments (63)
I don't know how to use this...most of the time you have to have the 2 .json files and tranformer blocks and unquantized layers.safetensors file. this is just one file.....do i just treat it like a checkpoint and try it?
@pychobj2001741 you need to install the new version of nunchaku (0.3.0), the new loader load single files
Has anyone updated to Nunchaku 0.30 yet? Is there a way to update to it or do we need to uninstall 0.20 and treat the 0.30 as a fresh install?
That's what I did, as Nunchaku 0.20 broke in one of the updates. Delete the nunchaku folder, uninstall nunchaku on the cmd box and reinstall.
The V3 is even better. I've got a bit puzzled that the file was a "Safetensors" and not a zipped file - but it worked flawlessly.
Yes the new loader uses a single safetensors file
I'm curious, do the other Nunchaku converted models (0.20 and prior) still work once you have updated to 0.30?
@BilboTaggins Yes the 0.2.0 models (multi files) work with 0.3.0 version
How much video memory and time did it take to convert without Evaluation Baselines Preparation?
@Retyz A40 with 48gb vram and about 20 hours to convert (smoothing and low branch calibration), setup with minimal quality
I've unpacked the fails into diffusion_models but sample pictures contains strange data for workflow - some NaNs, some strange schedulers or samplers. I've changed it and it works but this is little strange. You could add some description how to use this model...
You need to install Nunchaku, https://github.com/mit-han-lab/ComfyUI-nunchaku
https://github.com/mit-han-lab/nunchaku
i know it sounds wierd buut can u make svdq version of hidream full ?
To my knowledge no, because actualy deepcompressor only supports flux(dev, schnell), pixart or sana
@jice ohh ok ,jib mix flux also has that
I know this is not directly related to the model but nunchaku is a FUCKING mess. I tried to use it with Comfiu, and it's always giving me missing nodes errors, even if I manually install them with Git or using Manager. Is there a way to use it without having a college degree?
Have you installed the python module that corresponds to your installation?
https://github.com/mit-han-lab/nunchaku/releases/tag/v0.3.0
jice They have a point. I've managed to get a working install but it's a massive pain in the arse.
Super simple...https://www.youtube.com/watch?v=eATIu4lkOl0
official help release,
https://nunchaku.tech/docs/ComfyUI-nunchaku/get_started/installation.html
and i have installed it in one click by latest scripts, one is for wheel installation and other is workflow. and its working.
i ma having issue with feet and hand fingers. is there a way to fix that?
The hands are no worse than at any other checkpoint. Give me your prompt so I can try it.
my nunchaaku loader not detectingthe nunchaku v3 model, but works with svd version
svd version ?
@jice SVDQuant-int4-CreArt this 5.3 GB version works, but Nunchaku v3 - 6GB version not detected by NunchakuFluxDiT Loader
@imsostupidatu102 You need to install the last version of the wheel and the last version of the node before use the single file.
Have you noticed that when using Nunchaku to load this int4 model for generating high-resolution images, you can see horizontal stripes in the generated images upon close inspection? Is there any way to reduce this issue? I don’t see this problem when using the flux dev int4 model.
It's a problem with all compressed models. Some are worse than others.
This only show up on a very small percent compared to other models.
I think the quality of this one is really good after trying a bunch of INT4 models
EulerA bong_tangent 13steps, no loras
Usually when you upscale the image the stripes disappear, try different upscale models
@r600 Thank you for your answer, I understand now! I also really like this model, and the stripe issue was something I just discovered. I found that increasing the number of steps helps add more detail and alleviates the stripe problem. By the way, what is bong_tangent?
@jice I tested several upscalers (such as 4x_NMKD-Siax_200k, 4x-UltraSharp, 4xFFHQDAT), but none of them alleviated the stripe issue I mentioned. Currently, I found a relatively effective method is to use the original flux dev model for repainting with low denoise value.
@j0ey In reality, stripes often appear in models that have been merged with Loras. Your repainting solution seems really effective. Thanks for sharing.
there is a nice solution for all models in the workflow here: https://civitai.com/articles/16849
@Optimum Now, you can use my new node for remove banding artifacts
https://github.com/tritant/ComfyUI_Remove_Banding_Artifacts
@jice yes, it's great! Thank you. I exposed the denoise setting and I'm finding all kinds of uses for it.
Optimum XL model with very low denoise value. ultra-light solution always looks great!
This version 0.3.0 is really good. I'm really impressed.
thx
How do I find the installation method and workflow?
You can download the png file and drag it to comfyui
Nice! I would really like to see a commercial free Schnell version. 😊
I don't work on schnell versions
@jice no worries. I have moved on to other models. Thanks for your work!
please make blackwell rtx 50series Model. Thank you
+1
+1, also keep up the great work. Love the models, hypersdxlv1.2 is still great.
+1
I can't make an fp4 version, it costs too much to rent an rtx 5000
@jice How many hours does it take?
@gonztok approximately 20h
Good question
unbelievable so good, 0.3.0 is great, will we see more trained nunchaku models, are you working on more,
civitai needs nunchaku under filters, sad there is only 2 models on civitAi for nunchaku
Not at the moment
The model is divided into two files. How do you load them in Comfy? Nunchaku node can only load one file.
This is the old version of nunchaku who use two files, for use with the last version of nunchaku, download version 0.3 of this model
I'm using Nunchaku INT4 CreArtModel model and I have a problem with infamous Flux grid artifacts when using img2img (any denoise value) without other Loras. However your Creart Lora with original Flux-dev does not produce the grid effect and is clean.
@vladulidlo You can use https://github.com/tritant/ComfyUI_Remove_Banding_Artifacts for removing artifact (search jice in manager)
@jice Nice, I will try, thanks!
Thanks but where is the fp4 version?
fp4 version does not exist
Can someone help me upload the model? I don't understand why it's compressed :( Or if you could provide a workflow, I could replicate it and start from there. Thanks in advance.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.