Nunchaku NVFP4 version of the Fluxed Up 5.1 model. Some quality issues, legs seem to get weird sometimes. Can use with the turbo lora and seems to produce decent results with Euler/Simple, 8 steps. Without the turbo lora you can use DPM++ 2M/Beta, 30 steps.
I reduced the smoothing grids to speed it up, I might try it again with more grids when I have more time.
Description
INT4 version as requested. This is also a fast smoothed version, so quality may not be great. I don't have any hardware that'll run int4, nunchaku just crashes when I try, so I can't really test it out. The eval samples after the quant completed looks fine, so should be okay? ... Good luck!
FAQ
Comments (23)
Its fp4, but i have rtx 3070, sad
Yes please int4
What does that mean? I have 3060, this won’t work?
@diego892009410 yes. its work only rtx 40-50
@Gaydevai_paulo
Only 50 series, not 40
I'm still fiddling with quantization settings to speed up the process without losing much quality, but I can see about making an int4 version, too. Smoothing can take up to 40 hours and this is also my gaming PC :).
I should be able to get an in4 version released by this weekend.
I did a quick int4 smooth and quant for now, will post it as soon as I finish converting to nunchaku, merge into a single file, test to make sure it actually works, then generate some test images. probably in like 20 minutes.
@thaddeusk Can you post the recommended settings, like scheduler and sampler for Stability Matrix?
@diego892009410 I don't know what Stability Matrix is. I put what I used in the description, and the images have attached workflows in the metadata, so you should be able to use those as a starting point.
so, I have the model ready to go, but the ComfyUI-Nunchaku node won't let me run an int4 model on my blackwell card, so I'm not sure how to test it first :P. Might just modify the script to ignore that to see how it goes.
edit: Well, that just crashes ComfyUI. I get the same error for the stock F1D int4 model, so maybe it'll work? I'll just post it and let you guys test it out, I guess.
It's posted. I just added a few images from the FP4 run. Let me know if it works.
@thaddeusk It works like a charm at Rtx3060 with Turbo lora is like 15sec for 1024x1024 and 768x1344 (this is best resolution for this model I think). Thanks for share the model.
@flo11ok874 Glad to hear it's working! :)
I have nothing personally against merges, but for your hopes of greater adoption of your models, consider images where the faces are not nearly identical, on a biological (male) level, different faces, even if not quite as hot, still hold greater appeal. (your examples are great though), this is intended to be a helpful constructive posting.
I don't really care if people adopt it, I'm just fiddling with quantizing models using nunchaku. I just copied and pasted prompts from the original model's images to see if I can produce similar images. I figured I'd post it just in case anybody cared to try it out. Thanks, though!
Really appreciate the int4 version. Thanks for your work.
No worries. I couldn't test it on my hardware, how's it working?
@thaddeusk It is working perfectly, and very fast, on my 3080ti. Will post some of my generations later. Thanks again.
@Fit_Wafer9678239 Glad to hear!
@Fit_Wafer9678239 And you no post generation later, now people wonder is it bad gens?
Thank you very much for your quantization version. The effect of int4 is very good. I wonder if you can continue the work and quantize some other NSFW models, such as flux kontext or flux krea
I think the base flux models tend to get official releases from the Nunchaku team, was there a specific fine-tuned version of those that you wanted?