fal announces Flux a new AI image model they claim its reminiscent of Midjourney and its 12B params open weights

[-]

yangguize@reddit

Fal is constantly blocking innocuous images as NSFW. Can run the same prompt thru Comfy. Safety tolerance of 2 is ridiculous. No way to contact support.

Reply

[-]

CheeseRocker@reddit

Just to be clear, Black Forest Labs (https://blackforestlabs.ai/) built the model. Fal is just running it on their inference engine.

Reply

[-]

houmie@reddit

Thanks for clarification. But isn't BlackForestLabs also offering their own API by now? Especially when it comes to the newer Flux1.1. Which one is better do you think?

Reply

[-]

rnosov@reddit

Actual [Huggingface repo](https://huggingface.co/black-forest-labs/FLUX.1-schnell) for the smaller apache2 model. The bigger one is non commercial.

Reply

[-]

pigeon57434@reddit (OP)

I don't get why people on Reddit seem to care so much about commercial use licensing I mean how many people are actually gonna use this stuff for a business or something

Reply

[-]

redAppleCore@reddit

A lot more of us than you’d think, people are already using sdxl models in all sorts of commercial tools

Reply

[-]

Then don't listen to people with questionable reading comprehension. You are allowed to use the outputs of both open weight models as you please, except creating a competitor. > "Outputs. We claim no ownership rights in and to the Outputs. You are solely responsible for the Outputs you generate and their subsequent uses in accordance with this License. You may use Output for any purpose (including for commercial purposes), except as expressly prohibited herein. You may not use the Output to train, fine-tune or distill a model that is competitive with the FLUX.1 [dev] Model."

Reply

[-]

pooplordshitmaster@reddit

that is something different. output licensing is not license to commercially use the model/api for commercial purposes

Reply

[-]

eiva-01@reddit

In other words, you are not permitted to use the output to fine-tune stable diffusion. That's actually a pretty significant restriction, even for hobbyists.

Reply

[-]

ThisGonBHard@reddit

Technically, as long as you never accepted that TOS, nothing is stopping you. The images lack copyright after all.

Reply

[-]

eiva-01@reddit

Yes, but I guess the problem is that if you're using Flux then you're accepting the licence terms, and you're technically bound by the terms regardless of the copyright on the images you generate. So you're obligated to ensure you don't use those images for training should you migrate to another model in the future. Practically impossible to enforce though I think.

Reply

[-]

arthurwolf@reddit

I don't plan to use these for commerce, but I still care about license being non-commercial: 1. It strongly reduces how much the model/project is going to get feedback, forks and contributions. 2. They often claim to be open source, but the definition of open-source explicitely exclude non-commercial. 3. They can create uncertainty and legal risks, even for non commercial users, the definition of «commercial use» can be ambiguous, and users can accidentally violate the license terms if their usage changes over time or indirectly leads to commercial benefit. 4. Limits the potential for the project to be integrated into larger ecosystems or platforms, because many of these have commercial aspects and/or goals. 5. They are also a hinderance to research and academic use because these institutions often have commercial deals/partnerships with for-profit entities and this creates weird grey zones, that they then have to avoid, reducing research. 6. Prevents bundling into software and distros / other projects. I expect that'll become more of an issue as usage evolves. There's more than this too, this is just off the top of my head.

Reply

[-]

silenceimpaired@reddit

Hobbies have a tendency to move toward businesses. Licenses outside of Apache and MIT often have rug pull clauses that give the company opportunity to remove your legal access to their models… whether you are commercial or non-commercial.

Reply

[-]

Severin_Suveren@reddit

Worth mentioning it's the implementation of the image generation tools and model that's covered by the license. Any images produced by the commercial models are yours unless they already infringe on existing works

Reply

[-]

silenceimpaired@reddit

That isn’t clear in my mind looking at the licenses.

Reply

[-]

Severin_Suveren@reddit

The FLUX.1 [dev] Non-Commercial License specifies different usage terms for the model itself and the outputs (images) generated by the model. Here's a breakdown of how this applies to commercial use: **Model Usage:** The FLUX.1 [dev] Model, including its algorithms, software, and any other materials provided, is strictly for non-commercial and non-production use. This means you cannot use the model itself in any commercial context. This includes integrating the model into any system that is used for commercial purposes, such as a product or service that is sold or generates revenue. **Outputs Usage:** The outputs or images generated by the model are not considered derivatives of the model, according to the license. Importantly, the company claims no ownership rights over these outputs. You are allowed to use the outputs for any purpose, including commercial purposes, unless specifically prohibited by other terms in the license. For example, you cannot use these outputs to train, fine-tune, or distill another model that would be competitive with the FLUX.1 [dev] Model. In summary, while you cannot use the FLUX.1 [dev] Model itself for any commercial activities, you are permitted to use the images it generates for commercial purposes, subject to the conditions specified in the license

Reply

[-]

bobzdar@reddit

I mean, isn't using the model to generate images for commercial purposes also a commercial use of the model? Or is the restriction there that you basically can't just wrap the model and offer it as a commercial service or as part of a service?

Reply

[-]

Expensive-Paint-9490@reddit

The second one.

Reply

[-]

silenceimpaired@reddit

Wow. My respect for this company continues to grow. I was thinking in the context of SD3 and just assumed this model followed that. I am just floored by their reasonable approach.

Reply

[-]

rnosov@reddit

Say you mindlessly post an image from the non commercial model to your web site or instagram or such like. You could be potentially be hit with some nasty legal action later on.

Reply

[-]

Sarashana@reddit

That's not very likely to happen, at least not unless new laws are getting passed. The output of generative AI is considered copyright-able in absolutely no jurisdiction I am aware of. "Commercial" use in these licenses generally targets hosting and generation services.

Reply

[-]

silenceimpaired@reddit

Not true… if you look at the SD3 license and the expectations of the company based on their webpages and huggingface posts they expected artists to pay for a commercial license. Many large language models have non-commercial licenses and of the few I’ve asked for clarification on the answer is that output is also expected to be used for non-commercial purposes.

Reply

[-]

-p-e-w-@reddit

That's meaningless. Licenses and contracts are not blank slates that the company is free to fill with whatever clauses they can think of. They operate within a larger legal framework, and courts around the world have made it crystal clear that the output of AI models is not copyrightable. The model creator may claim otherwise, but that doesn't make it so. They have no rights to the output, period.

Reply

[-]

silenceimpaired@reddit

I don’t think the leap from it’s not copyrightable to they can’t control my actions via a contract is as easy as some make it. But I’m not a lawyer so (shrugs)

Reply

[-]

-p-e-w-@reddit

The point is that they can't control the *output*. They can (possibly) sue for breach of license or whatever (which usually requires them to prove damages if they want money), but under no circumstances do they own the output. No one owns the output, it's not a "work" any more than the song of a bird is. And even the idea that those licenses are enforceable at all is shaky at best. AI models are generated by automated programs from data that the model creator doesn't have a license to use. It's quite possible that courts might rule that it is in fact *the model creators* who are in breach of license here. The whole thing relies on FUD to work. The model creators would have to be insane to take anyone to court. If the judge rules against them, their entire business becomes instantly worthless.

Reply

[-]

Sarashana@reddit

Yes, I know about these clauses. I was/am looking forward to see them getting laughed out of the nearest court for it. There was already a precedent setting case in the US, that ruled very clearly that AI models cannot infer copyright on the content they create. The output is literally public domain, because no human was involved in creating it. The only exception is when drastic manual changes are being made to AI output, but there is no ruling I am aware of setting thresholds for much human change is required. That's still legally murky terrain. But even in that case, the copyright would be held by the artist, not the model or whoever made the model. Disclaimer: IANAL

Reply

[-]

alvenestthol@reddit

Commercial licenses function perfectly fine for software that don't create any copyrighted material; if a design company was found using e.g. WinRar without a license, it is perfectly legal for WinRar to sue the company, even if WinRar isn't being directly used to create any of the designs.

Reply

[-]

nmkd@reddit

Instagram is not commercial

Reply

[-]

KrazyKirby99999@reddit

It can be

Reply

[-]

AutomataManifold@reddit

Maybe not for this particular model, but my day job does involve AI, so I do care a bit.

Reply

[-]

xcdesz@reddit

There are a lot of real applications for commercial use that eould make artists more productive. People are thinking aboit background images, 3d textures, adventisements, logos, etc.. One potential major usage I can imagine is image generations for the thousands of animation frames neeced for animated films. You dont need text 2 video like Kling or Sora. You can simulate a scene using generic 3d models, and later replace each frame with a generated image. This gives the studio more control over the direction of outputs.

Reply

[-]

West-Code4642@reddit

many?

Reply

[-]

BillDStrong@reddit

Let's say you are writing a book you intend to publish. Using these models to generate ideas, funny scenes, or even first pass editing is a common thing for editors, but you have to know what you are allowed to do with it first.

Reply

[-]

synn89@reddit

Because if I'm going to spend my time/energy learning and fine tuning a model, I don't want limits on what I can do with the output or what other people can do with the stuff I create.

Reply

[-]

cyan2k@reddit

they wrote for the dev version: ""Outputs. We claim no ownership rights in and to the Outputs. You are solely responsible for the Outputs you generate and their subsequent uses in accordance with this License. You may use Output for any purpose (including for commercial purposes), except as expressly prohibited herein. You may not use the Output to train, fine-tune or distill a model that is competitive with the FLUX.1 [dev] Model." So you may sell your outputs.

Reply

[-]

astrange@reddit

Tbf it's legally questionable if they have any rights over the output of their model. Most people probably don't want to argue it though.

Reply

[-]

Inevitable-Start-653@reddit

I can't tell is it smaller or is it just a base and they are charging for API access to a fine-tune?

Reply

[-]

rnosov@reddit

It states on the model card that it is a distillation of their flagship model so it has be smaller. I don't think they offer any finetunes. I guess the business model is to charge for the API use of the flagship model.

Reply

[-]

Inevitable-Start-653@reddit

Oh good catch, thanks. I wonder if the open source community could train the model into better shape than their flagship model? I'm interested in trying out the base model.

Reply

[-]

BangkokPadang@reddit

There will probably be a feedback loop where people will train the smaller models, and they'll take notice of any interesting techniques or improvements and continue to host the 'best' version of their flagship over time, while working on Flux 2.0 in the background, at which point they may even release the OG flagship (ala how Mistral just released Mistral Large).

Reply

[-]

Inevitable-Start-653@reddit

Interesting hypothesis, makes open source more important in the ai development environment too.

Reply

[-]

daHaus@reddit

Their github repo describes it like this while their site gives the following comparison >We are offering three models: >FLUX.1 \[pro\] the base model, available via API >FLUX.1 \[dev\] guidance-distilled variant >FLUX.1 \[schnell\] guidance and step-distilled variant https://preview.redd.it/aq6fuwf1l2gd1.jpeg?width=1024&format=pjpg&auto=webp&s=5425a56124f0181378e497264151f10d7cb89027

Reply

[-]

Downtown-Case-1755@reddit

> 23.8 GB *sweats in 24GB*

Reply

[-]

onetwomiku@reddit

Cries in 2×22Gb

Reply

[-]

Sisuuu@reddit

Noon question: so multi gpu wont work models bigger then one GPUs memory size?

Reply

[-]

onetwomiku@reddit

No, diffusion models can't be split (afaik). But thankfully flux loads in fp8, so single 22gb 2080ti is (screaming) working just fine :D

Reply

[-]

420Microbiologist@reddit

Can you help me understand this? I utilize the modded 2080's in other pursuits but am confused on how it could run this without the extra 2gb vram

Reply

[-]

molbal@reddit

Increased memory 2080s?

Reply

[-]

onetwomiku@reddit

Yeah, 2080Ti from ali

Reply

[-]

arthurwolf@reddit

How much did they cost you?

Reply

[-]

onetwomiku@reddit

~470 usd each

Reply

[-]

sammcj@reddit

Are the weights actually liked there anywhere? Just looks like a marketing pitch for their API

Reply

[-]

nmkd@reddit

https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main

Reply

[-]

sammcj@reddit

Oh! Beaut, thanks! :)

Reply

[-]

uber-linny@reddit

for a noob ,,, how the hell do I load this within LMSTUDIO ?

Reply

[-]

IlIllIlllIlllIllll@reddit

mfw i own a 3090 and it says "loading in lowvram mode"

Reply

[-]

Ketworld@reddit

Go and watch this, https://youtu.be/stOiAuyVnyQ?si=mxFGVuvfCDz-zdPW He is running a 3090 and he changes a setting so he can use the GPUs full capacity and then runs Flux through comfyui

Reply

[-]

CryptoSpecialAgent@reddit

Mistral Large 2 + Flux + Open interpreter + Any of the open source chat UIs = Chatgpt Plus, completely self hosted, with almost no guardrails... Just a bit of code to chain the LLM to the image gen and the tooling and you've got something you can self host that's highly competitive with gpt-4o and dalle-3 and should be equally user friendly. Its quite amazing how much can change in a week!

Reply

[-]

Such_Advantage_6949@reddit

The only problem is this setup is still abit slow on consumer hardware. If 6090 come along with more vram. That would make it perfect

Reply

[-]

Amythir@reddit

Rumors are that the 5000 series will have less VRAM because Nvidia wants to reserve high vrams for commercial grade cards with the higher price tags

Reply

[-]

Such_Advantage_6949@reddit

It cant be less. Cause alot of the buyers of 4090 is for actually machine learning instead of gamers. I think they probably wont increase for the lower end but 5090 should have more VRAM

Reply

[-]

Expensive-Paint-9490@reddit

My perception of consumer hardware has changed in the last year. Before, I thought that no consumer needed more than a top gaming desktop. Because I couldn't imagine an actual use case. Now with local LLMs I can totally understand consumer using multi-GPU set-ups and even workstation. It's not about money, in the sense that spending 10,000 or even 15,000 on your hobby has always been a thing - think of motorcycles, carbon-fiber bicycles, trals around the globe, and so on.

Reply

[-]

Such_Advantage_6949@reddit

Yes. Fully agree. And if u use AI for work e.g. it easily x2 the productivity. Buy a rich with 2 used 3090 should cost about 2k if you waiting for good deal to come buy, so it is not that expensive like some ppl make it to be. ( about same price as those razor laptop that ppl splurp money on). And 2x3090 pretty much already can handle anything till 70B

Reply

[-]

arthurwolf@reddit

Just wait until we have open-source LLMs with native image output (not using a separate model, right out of the LLM). **That** will be insane. I'll say maximum a year until this happens.

Reply

[-]

vlodia@reddit

Is this same with Huggingface? https://huggingface.co/black-forest-labs/FLUX.1-schnell

Reply

[-]

Hunting-Succcubus@reddit

poor guy's gpu

Reply

[-]

Downtown-Case-1755@reddit

Is it actually all on vram, or is it spilling over to ram? What's your backend? Comfyui? Quantized?

Reply

[-]

IlIllIlllIlllIllll@reddit

comfyui. i didnt quantize anything, although i'm using the 8bit version of the clip-model. vram stays constant at 23.1gb usage, main memory usage spikes up to 25 gb (out of 32). not sure what happens internally though.

Reply

[-]

Downtown-Case-1755@reddit

Quantized to FP8 (with diffusers) the entire model+T5 fits into a 24G card with no cpu offloading, but *just barely*. I can probably run T5 on CPU though, right?

Reply

[-]

Electrical_Crow_2773@reddit

Have you tried disabling shared memory? [https://nvidia.custhelp.com/app/answers/detail/a\_id/5490/\~/system-memory-fallback-for-stable-diffusion](https://nvidia.custhelp.com/app/answers/detail/a_id/5490/~/system-memory-fallback-for-stable-diffusion)

Reply

[-]

IlIllIlllIlllIllll@reddit

no, but isnt that dangerous? what happens if i run out of vram without shared memory?

Reply

[-]

Electrical_Crow_2773@reddit

You only disable it for certain applications, like python executable that runs your model. If you run out of vram, you will just get "cuda out of memory" and the generation will stop. Everything else will still use shared memory, and if the model takes too much space, other programs will move to ram. At least, that's how it worked for me with llama.cpp

Reply

[-]

Downtown-Case-1755@reddit

Yeah, I am just thinking there is no way its not spilling over. I am testing it in diffusers now.

Reply

[-]

Friendlyquid@reddit

I'm a bit of a tech noob. But I got stable diffusion working on my computer half a year ago or so. How can I switch to flux? Is it a different thing or is it similar to stable diffusion?

Reply

[-]

arthurwolf@reddit

Do we know if controlnets would work with this? They are indispensible to my workflow...

Reply

[-]

somethingclassy@reddit

Different architecture than SD

Reply

[-]

SquashFront1303@reddit

This model is damn good with text

Reply

[-]

COAGULOPATH@reddit

Hands, too. They still look a bit "weird" (knuckles too long/short) but an image needs to get pretty busy before it starts generating wrong numbers of fingers.

Reply

[-]

CygnusX1@reddit

[Heads not so much?](https://fal.ai/models/fal-ai/flux?share=e7e98018-fd69-45c0-90bd-1bf4eb275ca3)

Reply

[-]

ShadowDV@reddit

I'll have Snotty beam you up, Sir.

Reply

[-]

Many_Ground8740@reddit

does flux support multi-gpu?

Reply

[-]

Ylsid@reddit

closed-source version only available through API /Sigh/ into the garbage then

Reply

[-]

Ill_Yam_9994@reddit

So what's the difference between the Schnell and Dev? They're both the same 23.8GB. Is Schnell like a Lightning sort of situation?

Reply

[-]

wsippel@reddit

Yes, it's a Lightning/ Turbo model. "Schnell" means fast in German. Dev is the standard small model, Pro is the large, proprietary model.

Reply

[-]

Ill_Yam_9994@reddit

Was messing around with the dev seems pretty good.

Reply

[-]

Hans_Meiser_Koeln@reddit

> open-sourced with a non-commercial license I'm just here to get some downvotes for saying again that the notion of "open source" in this entire space is fucked. Nothing about this is open source.

Reply

[-]

Hipponomics@reddit

100% agree. > I'm just here to get some downvotes This is an insanely tired meme though. Comment would have been better without it. Your statement isn't an unpopular opinion.

Reply

[-]

astrange@reddit

The main freedom of open source is the ability to change the program. It's very difficult to change a compiled program, but not difficult to change a model, just expensive. So I think there's some reasoning behind it.

Reply

[-]

bannedfromreddits@reddit

Please don't stop saying it. The people defending giant corporations perverting the term "open source" are genuine bootlickers.

Reply

[-]

riticalcreader@reddit

Facts

Reply

[-]

tessellation@reddit

forget it, it's a hopeless fight against the mass. also, models do not hallucinate, they confabulate... and a hacker does not crack your passwords... (we all know it's '123456')

Reply

[-]

charlesmccarthyufc@reddit

i put this up on [craftful.ai](http://craftful.ai) for free o try out its soooo good better than midjourney https://preview.redd.it/d2r49paa93gd1.png?width=1024&format=png&auto=webp&s=750b47079829e99f7b699fc14858579472876ed5

Reply

[-]

charlesmccarthyufc@reddit

https://preview.redd.it/lp5zp7cf93gd1.png?width=1024&format=png&auto=webp&s=9b2dd2bb58f9748eaea87fbe710313bbf1ce5f5d

Reply

[-]

ZHName@reddit

How did you get Caucasian people? I tried a few times and it seems to be super biased toward only making everyone dark skinned. Even specifying "man" will get you a black man without variation.

Reply

[-]

CryptoSpecialAgent@reddit

I asked for a photo of Kamala Harris addressing a confederate rally in Alabama, with a Let's Go Brandon banner in the background, and it correctly rendered the scene: the audience members were Caucasian, the speaker looked exactly like the vice president of the United States, and the banner read "let's go brambom" in what looked like the scrawl of an illiterate moron - and I don't think it was a problem with the model because it typically renders far more complicated text perfectly, I think it actually inferred "confederate, Alabama" and "illiterate moron" were semantically similar concepts 😂😂😂😂😂

Reply

[-]

charlesmccarthyufc@reddit

I put blond woman. I have not noticed the bias but I didn't look for it

Reply

[-]

ambient_temp_xeno@reddit

This is terrible.

Reply

[-]

charlesmccarthyufc@reddit

https://preview.redd.it/p179csot93gd1.png?width=2048&format=png&auto=webp&s=2ee2f9a9181015f4709df12f917fd8169c59df5a

Reply

[-]

charlesmccarthyufc@reddit

https://preview.redd.it/1gdokc2q93gd1.png?width=2048&format=png&auto=webp&s=f5645e9e15d59f4742c9e1ba35cd89a757dcada0

Reply

[-]

charlesmccarthyufc@reddit

https://preview.redd.it/xo6762fo93gd1.png?width=1024&format=png&auto=webp&s=44b8a107dde5b55161fed60fb6a7bbb2fc6b60b4

Reply

[-]

charlesmccarthyufc@reddit

https://preview.redd.it/62u3ez0m93gd1.png?width=1024&format=png&auto=webp&s=277f1f81cfbe942553944dfac2a490403d47b3f9

Reply

[-]

charlesmccarthyufc@reddit

https://preview.redd.it/liddwbuh93gd1.png?width=1024&format=png&auto=webp&s=852f6290e7b499de914e849439a55b2930f7b697

Reply

[-]

Weary_Weather_7633@reddit

It's a good news for open-source image generation. I build the Flux pro image generator online. If you are interested with it, please visit: flux-pro .net

Reply

[-]

Lemgon-Ultimate@reddit

https://preview.redd.it/4neu8h68k4gd1.png?width=1024&format=png&auto=webp&s=14c7c3829ab9a22d38e83bb19c136521738638d9 I've tested it and... it's crazy good. Generated me this anime girl out of the box. I know this kind of precision only from SDXL finetunes. Honestly, I'm totally impressed.

Reply

[-]

VoidAlchemy@reddit

Working well on my 3090TI in after following the [ComfyUI](https://comfyanonymous.github.io/ComfyUI_examples/flux/) quick start guide to manually download the HF models and put them into the correct directories. The following tests were using the this default workflow. Uses \~20GB VRAM with the `flux1-dev` model with what I believe is the fp16 weights. Though the debug log spits out `loading in lowvram mode 19047.65499973297` With GPU power capped at 275W with `sudo nvidia-smi -pl 275` getting generation at 1.4s/it so just under 30 seconds for a 20 step image. At full 450W it is \~1.3s/it or 25 sec per image though doesn't seem to pull full power. Does words very well and impressive quality! Has a "softer" feeling than many of the SD models I've tried. Cheers! https://preview.redd.it/6v3pjh2e54gd1.png?width=1024&format=png&auto=webp&s=fc7ba547a21b40d311cda15528a0b4089444f23c

Reply

[-]

Downtown-Case-1755@reddit

Its seems faster in diffusers, but obviously everything is diy there

Reply

[-]

celsowm@reddit

Any place to test it online?

Reply

[-]

dr_lm@reddit

Yessir https://replicate.com/black-forest-labs

Reply

[-]

Calcidiol@reddit

So I see the description of the dev model:"FLUX.1 [dev] guidance-distilled variant" But I don't really understand what's different about it than the "FLUX.1 [schnell]" model; why and how would I want to use the dev variant vs. the "FLUX.1 [schnell]" one? What makes it "dev" (for development of what / how and if so what is the schnell one for?)?

Reply

[-]

AnomalyNexus@reddit

FYI from hn: >hi friends! burkay from fal.ai here. would like to clarify that the model is NOT built by fal. all credit should go to Black Forest Labs (https://blackforestlabs.ai/) which is a new co by the OG stable diffusion team.

Reply

[-]

badgerfish2021@reddit

would be interesting to know how big the pro model is compared to dev, the pro quality and prompt adherence is significantly higher

Reply

[-]

AnticitizenPrime@reddit

[Wow, it's really good with text.](https://i.imgur.com/AzNcuYK.jpeg) I re-generated this image like 10 times and it didn't make an error with the text once. Used this Huggingface demo: https://huggingface.co/spaces/black-forest-labs/FLUX.1-schnell

Reply

[-]

balianone@reddit

Wow, that's really good with anatomy, 5 fingers indeed! However, I still can't draw Balinese correctly, probably due to the limited dataset. My Balinese girl ended up looking Chinese, lol https://imgur.com/a/F37q1Bi

Reply

[-]

tessellation@reddit

The LEGO hand, on the other..

Reply

[-]

121507090301@reddit

Does anyone know if such a model would work with a quantized gguf? I can't run this on vram only but I can just about run this in a lower quant on cpu if that is possible, although the loss of quality could be too much of a deal breaker...

Reply

[-]