@RalFinger I think it can`t be nsfw

@RalFinger  Does it work in Forge? Does it run on an RTX 3060 Ti?

you need new PC ;)

@0l1v1aR0551 What ? You'll not make a 4GB wf ?

@Brewce here you go: https://civitai.com/models/2168426/flux2-low-vram-workflow

Even more reason to upgrade from my 5080 to a 5090

@delta45424155 or pay for cloud services

@delta45424155 actually a 5070Ti and a 5080 would be the bare minimum to have some optimal speeds even with high Quantized versions.

@Lancerx My 5080 already feels slow. I couldn't imagine trying to use a 5070ti.

Hopefully some speed loras come soon. 2 minutes to gen a single 1024x1024 w/ 20 steps on 5090. Yikes.

@delta45424155 5090 won't make a difference, you'd need 6000 with 96GB VRAM

@Stardeaf I def wouldn't spend 10k for flux. I would if wan released a model that could continue without degradation after 5 second tho.

@delta45424155 Yeah, it's a bad time to upgrade anything anyway. Prices are nuts. I was considering upgrading 3090 to 5090, but hell, it's 7x what I paid for 3090, and only 8 GB of VRAM gain. I'll wait till the bubble pops.

@Stardeaf I paid 1,400 on day one release for my 5080. Would have gotten a 5090 but microcenter was out.

@delta45424155 What currency?

@Stardeaf usd

@delta45424155 ye but both have 16GB, which is usually the more important bit. Obviously the additional 15-20% raw power performance is welcomed :)

Nope.. but I guess they may have some "Artistic Nudity" data.

No. Expect it to be even more censored than F1. I doubt many LORAS will be trained for this model given it's insane size and restrictions.

<--- SAD

I tried on HF spaces and nope, for edition and for creation is SFW

Well do I care at this moment.. I don't.. I am loving Z-Image Turbo.. screw Flux.2 for now.. I made four workflows for Z-Image Turbo (Txt2Img, Img2Img, Inpainting & Outpainting) that will be able to use LORAs (if there are any for Z-Image Turbo) and all of them can do nearly as good result as Flux.2 with no Quantization under 30 seconds even on poor 8 GB (possibly 6 GB ?) GPUs.. I will upload these soon on my profile and share.

I couldn't even get the fp8 version to run the other day on 16gb and 48gb of system memory, one run took almost 8 minutes and the second run locked up my comp from disk swap

@FloatsYourStoat well on the other hand Z-Image Turbo can almost get you similar results without quantization under 30 seconds even on poor ass 8 GB GPUs, you have to try to believe it. If you feel like trying it on ComfyUI I got the workflows -

Z-Image Turbo Text-To-Image with LORAs Workflow

https://civitai.com/models/2172979

Z-Image Turbo Image-To-Image with LORAs Workflow

https://civitai.com/models/2173008

Z-Image Turbo Inpainting with LORAs Workflow

https://civitai.com/models/2173031

Z-Image Turbo Outpainting with LORAs Workflow

https://civitai.com/models/2173055

@FloatsYourStoat make sure to set the clip to CPU

@sarcastictofu I have tried out z-image, after my attempts with flux.2 failed. It is quite impressive speed wise, a mere 15.5 seconds per run. I did not however know it does in/outpainting

@2thecurve That could be where it went wrong for me, I tried running it all on gpu as I had done for flux.1, I may give that a try + the Q6_K a bit later

@FloatsYourStoat well they have not yet released the Z-Image Edit yet but even the turbo model can do quick and decent job, not as good as some Flux inpainting models but definitely better than many SDXL 1.0 models.. it can handle long texts well, almost 8 or 9 out of 10 times yet doing it under 30 seconds even on $h!tty 8 GB GPU.. which is crazy!!

@sarcastictofu I really can't wait for the base of it to come out, once loras land for it I think sdxl derived checkpoints will finally have some actual competition. As it stands it really struggles with some concepts, especially anthro stuff

@sarcastictofu holy crap! Thanks for turning me on to Z-Image Turbo, wow! That's a good model and fast, even the bf16 version of it. I really want Flux.2 to be fast because of its text encoder and advanced prompting support, but it's great to have so many new models coming out.

@HackAfterDark there is still hope for flux.2 to be reasonably fast, nunchaku is amazing. It takes them quite a while to release their versions of the models but they are crazy fast, especially fp4

@FloatsYourStoat I and get the gguf version working yet. I do have everything downloaded

@HackAfterDark how does it do with text in images

I ran it with 8GB VRAM + 32GB RAM. Of course is slow (~5 minutes, 1088 x 1440), but you can try it.

@promptkid honestly if this is too slow for you, give flux.2 klein a try its way faster, especially the low step (non base) version

@FloatsYourStoat Yeah, it was just a test. Klein 9B fp8 generates excellent images in 4-8 steps, 10~20 seconds.

@promptkid I have a 6gb GPU with 16 RAM. If it take on average 5 mins to generate a image, then I don't mind waiting.

Q8 works, if you don't mind 'partial' loading thing Comfy does. Q4 if you want to load 'fully'. Q4 is marginally faster, but there seem to be noticeable quality drop - I only did a couple of tests.

Q5_K_S

@ErikLee thank you

@RalFinger At least you tried quanting them, I'm sure many people will still find a lot of use for them. As for my 12gb's of VRAM sufferage, welp

Q7 if someone makes it should be able to run on a 5090 i'd guess, i'd think that should be good enough

The base model is 32B parameters... A bigger model has supposedly a higher capacity for problem solving... But at this size of model the amount of training required is insane in order to get a better result than flux 1 dev.

It has insanely great prompt adherence, very good quality, and edit capabilities with character consistency that not only beats all open source models, but not rarely Nano Banana Pro as well.
But it's over-censored and very slow model.

@yorgash How does this compare to Chroma?

The size is massive for the quality you get lmao you're right.