You can use the FP8 version! It's only checks notes... 35 GB.

And somehow it works on 24 GB, at least fp8.

@theally I guess I'll have to buy a SSD only to fit the checkpoint shrug

Available on HuggingFace Hub

@qek I'm not sure what that is?

@ggorebama853 ?
https://huggingface.co/Comfy-Org/flux2-dev ?

You get 50 free Generations here: https://playground.bfl.ai/

@qek you also need 64GB-96GB of Vram to run the full model.

@qek Thank you!

@J1B Thank you! 

That's what always happens with computers, when bill Gates said "640K RAM ought to be enough for anybody" in 1981, he was wrong.

Plus text encoder 36Gb :)

@J1B But now Moore's Law has failed, and as computer prices keep rising while keeping up with computing power demands becomes harder, many models can only run on server clusters. The most useful software on PC nowadays is the web browser.

USE GGUF

@love123654 It's not moore's law you should blame. The current silicon is plenty fast. It's greed we are battling. Nvidia and AMD do NOT want you to have more vram. They don't want to canabalise the sales of their high end hardware. Because there is no other competition in the space we all suffer. It's not a law of scaling. It's just greed.

@love123654 in 1981 the base model IBM 5150 cost $5,600 (and upto $20,000 for a higher end models) in today's money, you could buy a very nice AI capable PC with a RTX 5090 GPU for that much.

@J1B With limited advancements in chip manufacturing processes, performance improvements in chips like the 5090 are largely achieved through increasing chip size.

@Keroro_Gunso The greed of these hardware companies is certainly a significant factor, but hardware manufacturing costs are also clearly increasing. Meanwhile, the growth rate of computational power and GPU memory requirements for models far exceeds the improvement in PC performance. I believe this is a joint attack by hardware and model companies on offline generation, but I believe our open-source community can counter it.

@love123654  I hope.

Thanks. On a 3090 24GB, Q4 loads "completely" in comfy, anything above loads "partially", however it doesn't seem to noticeably affect generation speed. Which is about 200 seconds for 1024x1200, as reported by comfy.

Only Q2 loads entirely on my 5060ti (16GB), but the system ram offloading for higher quants seems to work very cleanly, adding about 20 seconds per gen (230->250). I've tested Q3 and Q4, and as long as you have enough system ram there's no speed difference between the two. So presumably you can use the highest quant you can stuff into your system ram with no further drawbacks (still downloading Q6 to verify!).

2026 FLUX-3 256gb video ram

Good luck 🤑

It actually is "not safe for work". I, for example, have some work to do and instead I'm sitting here feeding silly prompts into this...

I ran 2-bit quants to try. 768x768, 20 steps, Prompt executed in 237.49 seconds. Became slower after adding ReferenceLatent, does it happen with other users?

Все хотят вкусно жрать и слаще спать,

Yes, it gets slower with each reference image. I'm getting generation times between 200 and 500 seconds on 3090 with Q8 version. It seems to need at least 30 steps to deliver. A slow beast, but so far it looks good.

It's a terrible model.

It still gives butt chins, triple chins. Flux Krea is just a fine-tune

Они не слаются, нет, чтобы изменить архитектуру, они клепают модели под 60 гигов, так там ещё остались проблемы с генерацией отдаленных объектов, как они были уродливыми, так и остались))))

How many steps did you do? I had body issues with extra arms and fingers but with 20 steps or less. They cleared out when I increased steps to 30. At least in 3 cases I encountered so far. I'll also need to figure how guidance works, from what I seen the range is 1.5 to 15. So probably set it half way for starters.

@Stardeaf 30 steps. Cfg was 5 or 6. I got a lot of bad images. This is best I've got.

@SaiWeb Was "DJ." the whole prompt?

It seems they use a LLM as TE.

There are allready 2 loras.

你说得很对

Quality of image can be discuss yes, but prompt adherence is very impressive. I m running some tests and the coherence of the final image related to the details prompted are stunning.

@Svengali75 What kind of adherence are you referring to? If it could handle comic panels like Qwen, that would be somewhat acceptable, but I believe this is far from sufficient. For a model of such a massive size, its quality should not be this blurry and lacking in detail. Its performance in realistic styles is even inferior to SD 1.5, which is unacceptable, not to mention the large number of anatomical errors it still exhibits.

@karl1688 Hmm... I'm not getting any of that so far. Maybe you do something wrong? Give us some details of the failed gens.

@Stardeaf I tested it using the same realistic photography prompt. As an ultra-large model with 32 billion parameters, its performance is even worse than a fine-tuned SD1.5 model. The skin in portraits still looks extremely plasticky, distortion is severe, limbs and fingers lack detail, and the image style is actually a regression compared to Krea. In summary, its generation quality is simply too poor. I can tolerate the slow speed, but this quality is unacceptable. Look at Z Image; that is the performance an advanced model should deliver.

Using AI Toolkit

We are talking about civit, realistic time frame for adding lora training for Flux 2 will be release of Flux69

@Shrekman17 We have it available for internal testing :) We'll open it up when we're satisfied that it works. It's going to be costly though; it's very resource intensive.

@theally Sounds nice anyway

@theally bar for my trust to civit at this point lays in hell,
i would like to be finally surprised,
but for over ~6 months the only useful for me thing added was Sora 2,
Actually Civit might be one of the best places to use Sora for now

I ran it and it took 10 GB of VRAM (used a 2-bit quant)

@qek 600 секунд генерировал одну картинку?))

У меня 5090 32gb, завелась с fp8, суть в том что flux2 не должна была работать, а вылетать с нехваткой, но по всей видимости они встроили что то вроде block swap, потому что видео память что при 1024х1024 забита почти полностью, что при 2048х2048... еще удивило жор оперативки, например: mistral_3_small_flux2_bf16.safetensors, сжирает все 92gb ram... С mistral_3_small_flux2_fp8.safetensors проблем нет, но все равно в пределах 64gb ram нужно иметь.

768x768, 20 steps, no reference latents, ~240 seconds

Уверен что сделают nunchaku версию и будет быстрее :)

@devold5000 Mistral is a family of LLM. So in fact you use hybrid of txt2img + LLM (Qwen also uses its own LLM). If someone replace LLM with something lightweight TE the size and memory consumption could be smaller.

у меня генерация картинки на 5080 30 шагов в фулл хд, юни пс занимаем около 250-280 секунд. Ну долговато, это на q8_0 модели. Ты можешь скачать мой воркфлоу с флакс2 и парить себе мозги касаемо времени генерации

@devold5000 Потребление памяти зависит от разрешения изображения практически линейно. У меня 4080S и 96Gb RAM. 1Mp - 6.5 s/it; 1.5Mp - 11s/it, 77%RAM; 3.5Mp -28s/it, 83% RAM в режиме Img2Img.

And large

I agree

I didn't have that impression. Actually I think it's much better than F1. But it needs a lot of steps to deliver. I think 30 is minimum. And maybe you need to figure out how to prompt it.

@Stardeaf F1 is a low bar. I don't think I've ever seen a good realistic image of a person with just F1. Even with a lora it's just bad. I never got the hype for Flux. It's finicky, huge model sizes, slow generations, and bad output. This new model is even slower and worse. And I thought Pony 7 was a disappointing model... Flux just proved they're not capable of improving their model as long as they're limiting their datasets and censoring.

@LetTheBassDrop I agree F1 is not great to put it politely. I've been late to tinker with it but so far results I could get were kind of disappointing, which I attribute to my lack of experience in handling it. But I don't get the hate F2 is getting here. I think it's mind blowing. It's not just about the image 'realism' or whatever, but the scope of concepts it understands. It's taking a step forward to a hybrid with language model.

Z-image turbo can fix it with a second pass of 0.65-0.85 denoise

@Primaveri Z-image turbo really good, but it needs a second pass in SDXL models to be realistic. Сreate an image in F2, refine it in Z-image turbo and then refine it in SDXL-I passed this quest. The result is good, but not worth the time. I hope that Z-image-base will fix the situation.

what r u talking ? link is right there. fp8 quantized from comfyui. click on split files and text encoders !

The text encoding folder in the link contains 01of10, 02of10, 03of10... Which of 10of10 is the text-encoded file? Do all of them need to be downloaded? If you want to download all of them, you can only choose one safetensors in ComfyUI

@sunweixi1993786 They said NO

https://huggingface.co/Comfy-Org/flux2-dev/tree/main/split_files/text_encoders

@ak002 I've seen the URL you sent. The file shows "small". Has it been deleted?

@ak002 Hello, the text_encoders you shared are the abrided version marked "small" made by ComfyUI, not the complete version of text_encoders from the studio!

👉 GGUF

My 4090 handles the fp8 version quite well. Just takes a bit of time...

@musigreg I can only run Q2

Must be your settings, my 4090 runs it just fine. Is your output image larger than 2K? That might be your problem.

I saw a news item saying that Nvidia has been working with ComfyUI to make this model run on RTX. You need to update your ComfyUI installation.

download my workflow for flux2, download gguf q8_0 model - it will better than fp8 for a quality, and you get 250-280 seconds generation ksampler time in 30 steps and 1920x1080 resolution. Good look!

@NapoInfr It uses the tensor cores? Is it why I am getting a 4k image with 20 steps (2 days ago) in just 80-100 seconds on 5090?

@mkDaniel gguf models never use tensor cores. only fp16, fp8 and fp4. 80-100 Seconds in 20 steps with 4k image on 5090 is very-very nice result

nvidia payed them to make it that big so that you consider buying pro card

@frfromg paid*

I'm a forge webui user and well.. I downloaded the model and the minstral 3.2 for text encoding and then I downloaded the VAE file for flux2 and put them all where they go.. I get an error that says I don't have clip or dict_ and I didn't see anywhere that says forge webui is not compatible with flux2, I also did not see anywhere that said I needed clip. oh well, I'll wait a bit longer before trying to use it again.

I can only run 2-bit quants

Yeah, I do not even try FP16 on 5090.

So i just seen clip needs to be set on CPU

worse than SD 1*

If you have at least 12k buzz, try the Civitai trainer

I can train on a 32GB with FP8. It is something like 10s/it but IT IS training.

Same as training FLUX1 on 12GB

No Schnell, there will be Klein instead
"Do we have other options for 12GB Vram"
For Flux2? Use more RAM than VRAM

@qek how to use more RAM ? In the VAE node? OR is there another way?

@skechtup 1. Some arguments, compare:
--highvram: Max performance, doesn’t unload models once loaded.
--normalvram or unset: Standard operation.
--lowvram: Saves VRAM but slows things down.
--novram: Barely uses VRAM.
2. The Load CLIP has the device option, it can be CPU, it uses RAM. VAELoader KJ also has the option

@qek i've tried all, it does not speed things up. When is the Klein version expected ?

12k buzz*?

@qek yeah 12k not 10 :/

Mistral-Small-3 is a standard model made by MistralAI, right? There are also Medium and Large models, but they are commercial and provided via API. Only Small is released as an open model.

@aueki4g467 It seems the OP didn't get how to get the text encoder for Flux2

@aueki4g467 So, is the text_encoders (bf16 and fp8) of "small" the complete version?

@sunweixi1993786 Yes, it is for Flux.2

Heavily censored a model unable to gen porn, such a joke

No doubt. I invested heavily, went ahead and got a 5090 FE, to pair it with my 12900 K, 64 GBS of Corsair Dominator Platinum DDR5 6800 mega transfers RAM, a 4 TB Sabrent Rocket 5 , Gen 5 NVME, And I used my old 3080 12GB as a secondary, and even got the high dollar motherboard the Z790 Dark Hero, figuring OK that should get me set up, then they're like... OK here hold my beer.

@qek yep, fuhked up, They don't realize how much they're holding their own model back from developing. 

@WWG1WGA17 yeah limitation has reach its peak

@WWG1WGA17 I just got a 5090. Was not expecting any models on CivitAI to be off the table! lol damn

@WWG1WGA17 Crazy, I can only run it in 3 bits. It seems the devs really hate us and made it as big and as censored as possible. But, apparently, Qwen Image Edit isn't like that, it wasn't made to be gross

@lucidzachary473 You can run a quantized version on a 5090 for sure. I'm running a Q8 version with a 3090.

No Flux.2 Max, but added GPT Image 1.5 and Seedream 4.5

@qek Max is live for generation now :)

@theally I see

I got noisy images, you can see my comment