Flux.2 [Flex], [Dev], [Pro], & [Max] are live for Generation!
FLUX.2 [Flex] is the next leap in the FLUX model family, delivering unprecedented image quality and creative flexibility. FLUX.2 is a state-of-the-art performance image generation model with top of the line prompt following, visual quality, image detail and output diversity.
Original Flux.2 [Dev] files: https://huggingface.co/black-forest-labs/FLUX.2-dev
FP8 Quantized from ComfyUI: https://huggingface.co/Comfy-Org/flux2-dev/tree/main
Description
FAQ
Comments (32)
Will this model run on an RTX 4080 Super with 16 GB of VRAM? Has anyone tried running it?
If quantized
This model indeed boasts exceptional controllability: it can almost fully interpret and execute complex prompts while delivering impressive details. However, this comes at the cost of its massive size. Even with a top-tier RTX 5090 graphics card and a quantized version of the model, smooth operation remains difficult, and its inference speed is simply not comparable to that of ZIMAGE. That said, this is still a development version with little to no optimization performed. If a team could optimize it—reduce its size and boost its speed—it would mark a brand-new chapter in the field of AI image generation.
There's their Flux.2 Klein?
"it would mark a brand-new chapter in the field of AI image generation" No, Qwen Edit is better, it follows my prompts better and without schizo ultra censorship
@qek I have no idea why, but when running the FP8 quantized versions of Qwen 2512 and FLUX.2 with 24GB VRAM, FLUX.2 outperforms Qwen 2512 in both generation quality and speed. I’m wondering if NVIDIA has specifically optimized it for this model.
I wouldn't say it's better at complex prompts than Nano Banana or even ZIT. I toyed with it for a bit (using AI aggregators 'cause i can't possibly run THAT on my PC) and in general it handled my prompts worse than either of the other 2.
@MV261 Read my comment "I tried Flex on the site, got an image covered in small RGB noise..."
Klein is missing! Oh my gosh! The 4B model is smaller than Z Image! Apache 2.0 license! But also has schizo censorship again.
Edit: I downloaded distilled Klein 4B, it's prompt adherence appears to be better than Flux.1 Dev Kontext's and worse than Qwen Edit 2511's. But Klein 4B is faster than Qwen Edit Turbo/Lightning and Kontext Schnell. Klein's outputs may be questionable, it can make textures smooth/plastic and alter an image due to the prompt adherence, but it's easy to avoid it
Lol, is that a thumbs up or down? I just downloaded it and am about to put it through its paces. Am I wasting my time?
4B is way faster than 9B, but still not as fast as z-image. But I might keep the workflow setup for 9B. It seems far better than Flux.1 Dev Kontext and I might need the turn around images and combine posibilities.
@egpieper I'm running it now 8b and 4b are close. I HAAAAATE Kontext.
Flux 2 did closer to Nano Banana stuff, and ZImage lacks a lot of lighting/ photography/ advanced prompts (I don't know how much better this will be with the same text encoder)
On to experimenting 😀
9B same, not so good.
Ok, I will try 9B to see. Really, I had better luck with Qwen Image Edit, Klein 4B might just ignore my prompt and do something else or wrong colors, just like Dev Kontext
I ran 9B, everything in fp8, OOM after 1 generation, bad prompt adherence and changed styles when unwanted anyway. Even worse, it turned my image into creepy gore due to my typo. When I use two reference latents and ask to make a duo photo, the character's bodies change, but their clothes and heads are same :?
@qek You've got somethign set up wrong. Are you using a conditioning zero out node on the negative side? It reads the clip still, so you need an empty Clip Text encoder. Also, lower the cfg
@lonecatone23 Wdym? I got gore due to "deathers" instead of "feathers". "conditioning zero out node on the negative side" of course, not to encode useless negative prompt due to CFG=1
@qek It was causing me issues until AI swapped it out, even with a CFG. Bad images. No clue why, just bringing it up. Also, Flux2 Klein has a cfg setting of 3.5 -5.0, not 1.0. D is 1.0
@lonecatone23 I have no problems anymore, I've been using default settings. I might dedicate more time to try various sampler+scheduler combos to try to improve the quality
I tried to do tiled upscaling of images generated by Klein with no latent references, it fails, only makes my images worse, Z Image seems better at it :(
@qek well, are you using it to hi rez fix or upscaling? It will always look worse if you don't hi rez fix it first
@lonecatone23 tiled, not hires. fix. It seems I'd need another model to upscale Flux.2's outputs
@qek I always fall back on DPM++ 2M for some reason for better quality. Seems to work on most model types I used.. As I often use partly denoise of a Latent to gain beter quality on resampling. I found something interesting in ComfyUI that Flux 2 Klein uses a new Latent node called "Empty Flux 2 Latent". I dont really know the diffrence between these latents, but I do inject image manipulation via a ReferenceLatent Node for now. That combines conditioning with another Latent gained from VAE encode of an reference image and returns it as conditioning, then I still use the empty latent on the sampler. It seems the CFG is a big factor in this. CFG 1.0 values the prompt you use more, and CFG 8.0 returns almost all details from the reference image and almost ignoring your prompt. You can chain diffrent ReferenceLatent Nodes, but its already interresting to tinker with just one and set CFG between 1.0 and 8.0 to scale until you reach a desired result.
@egpieper The typical "empty Latent" isn't really empty. It's a noise map. The difference in the basic/ sdxl/ flux 2 is how that noise is laid out or how dense it is.
Both Klein 4b and 9b have a cfg suggestion of 3.5 - 5.0, and Flux conditioning is typically how you adjust for I2I or T2I. all models tend to start overbaking at a cfg of 6.5 or above.
@egpieper Got a fired image with CFG 2.0, I don't use Base, is it for Base? Not distilled
@qek Flux 2Dev is a cfg of 1.0. Just the quants are higher cfgs
Where does the Klein model go? It's not a diffusion model and doesn't register as a GGUF. What folder in Comfyui does it go in and what node does it load with?
Think location is unet folder : there is a json for ComfyUI on this article : https://www.reddit.com/r/comfyui/comments/1qdnqmi/flux2_klein_4b_9b_fast_local_image_editing_and/
@egpieper Awesome. Thanks. It was crashing on me
@lonecatone23 I posted a few images that include my altered ComfyUI workflow inside the images. Give it some minutes before it becomes visible on the page. It was far from fast for me using this 9B version with a 5070ti on 16GB VRAM. Guess I stick with exploring Z-Image that is amazingly fast for me. But got interested because this Flux.2 version also has Kontext capabilities.
@egpieper I just finished updating this. I hate to say it, but the model they have here isn't very good. This does pretty decent,. I added notes where to get the better models
https://civitai.com/models/2213699/flux-2-klein-2d-and-gguf-pro-grade-workflow-high-and-llow-vram
Support for Klein models was added to ComfyUI yesterday, so if you're using latest ComfyUI all you need to do is 1) take any GGUF from Unsloth (https://huggingface.co/unsloth/FLUX.2-klein-9B-GGUF), 2) pick any of the Klein templates in ComfyUI, 3) replace default model loading node with the one from https://github.com/city96/ComfyUI-GGUF.
@MoonRide Thanks. Already got it