If you have a specific style in mind you could train your own LoRA for it. If you just want to generate random weird stuff, I've had that happen when I stack too many style LoRAs on top of each other, so you could try that. And if you're more technically minded, you could experiment with injecting latent noise. Couldn't find the exact example I was thinking of, but this article might provide a starting point.

Impressive, if it's true. Would be interesting to see. So far I've seen flux 2, which would require way more GPU! Han I have, and queen image which seems on par with flux 1. This could get me to finally up my pipeline

Oh crap this is a huge problem that they might have solved! I'll be trying this out!

Wonder how useful this'll be for adding additional language support to existing fonts . So many fonts out there can't use bcus only support latin characters . But ahhhhh (vectoris|tweak)ing each one's gunna be some work !

Holy shit that's amazing! 10 sec on a consumer gpu in 4g vram! I might decide to onboard this to the horde...

Yeah.

gotcha, so this downloads its own custom quantization models?

Yeah, it's a different form of quantization. I think it's supposed to look a little better than conventional quants, but it still has all the tradeoffs of a quant.

What are the tradeoffs? I can't believe a 50% to 75% RAM reduction is free. Is this like quantization?

The license is Apache 2.0, but I don't know how much VRAM it takes.

Do we know it's license and the vram requirements?

Might be interesting to infer optimal dimensions. Glancing at the source, looks like that doesn't presently happen.

The BF16 one is like 48GB+ of VRAM, but you can get the GGUF ones to run on 8GB VRAM. To telling how bad that'll look.

Ooof, 20G file for the FP8.

Do you know how much VRAM each of these needs?

Yeah, it's pretty normal within the Linux community not to recommend Snap 😉

Snap isnt recommended it says.

Uh, Snap package or compile entire GIMP yourself isn't very convenient for all the people who aren't within Ubuntu's ecosystem...

I will add, for anyone who has not played with it, that I am pretty impressed with what little I've seen of FLUX.2. If you can run it, it's got some pretty impressive output.

The downside is that it's memory-hungry and compute-hungry. On a Framework Desktop, I can run the FP8 release with about 54 GB of VRAM. It takes about 290 seconds

a bit under 5 minutes

to generate a 20 iteration, 1280x720 image, which is much more compute-heavy than any other image generation model that I've run locally.

I was just thinking how it's becoming cumbersome to manage all of these pipelines and different dependencies, models, etc.

Have only tried a couple textures yet, nothing fancy, but holy shit, rare Ubisoft W.

Some models such as Z-Image Turbo exhibit low seed variance

Honestly, that seems like a problem with the model. If you want similar images, that's really the point of the seed.

Thanks!

Thanks for explaining!

It's basically when you use a larger model to train a smaller one. You use a dataset of data generated by the teacher model and ground truth data to train the student model, and by some strange alchemy I don't quite understand you get a much smaller model that resembles the teacher model.

It's really hard training on a distilled model without breaking it, so people prefer models undistilled whenever possible. Without the teacher model, distilled models are basically cripple-ware.

A distilled model is a more lightweight version of a full model which can run in fewer steps at slightly reduced quality.

Z-image-turbo is a distilled model, and the full version of the model will be released soon.

This post is referring to someone attempting to somehow undo the distillation to make an approximation of the full model, I guess. Which is basically pointless because as I said, the actual full model will release soon.