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.
what we did at fal is take the model and run it on our inference engine optimized to run these kinds of models really really fast. feel free to give it a shot on the playgrounds. https://fal.ai/models/fal-ai/flux/dev
The playground is a drag. After accepting being forced to sign up, attach my GitHub, and hand over my email address, I entered the desired prompt and waited with anticipation.. Only to see a black screen and how much it's going to cost per megapixel.
Bummer. After seeing what was generated in the blog post I was excited to try it! Now feeling disappointed.
My go-to test for these tools so far has been the seven horned, seven eyed lamb mentioned in the Book of Revelation. Every tool I've tried has failed at this task.
> A Gary Larsen, "Far Side" comic of a racoon disguising itself by wearing a fedora and long trench coat. The raccoon's face is mostly hidden by the fedora. There are extra paws sticking out of the front of the trench coat from between the buttons, suggesting that the racoon is in fact a stack of several raccoons.
Every human I've ever described this to has no problem picturing what I mean. It's a classic comic trope. AIs still struggle.
A rough rule of thumb is that if a text-generator AI model of some size would struggle to understand your sentence, then an image-generator model a couple of times the size or even bigger would also struggle.
The intelligence just doesn't "fit" in there.
Personally I'm curious to see what would happen if someone burnt $100M of compute time on training a truly enormous image generator model, something the same-ish size as GPT4...
This gets interesting. One approach that I've used with image generation before is to find an image of the sort that I want, and have Dall-e describe it... and then modify the prompt that it provides to be one with the elements that I want.
The image shows an imaginative, whimsical illustration of a character composed of two parts. The upper part features a man dressed in a long, elegant gray coat, wearing a bowler hat and round sunglasses, with a sophisticated white polka-dot ascot tie. His face has a subtle smile. The lower part of the character transitions seamlessly into a smaller figure of a cat, appearing to wear striped pants, with its tail visible. The entire character combines human and feline elements, creating a surreal, anthropomorphic appearance. The illustration is in black and white, emphasizing a stylized, cartoon-like design.
The image captures a whimsical and secretive scene featuring three dwarves stacked in a totem formation, each attempting to conceal their nature under a large brown cloak. The top dwarf has a bright, cheerful expression and blond hair, holding the cloak wide to mimic wings, and is dressed in black armor adorned with teal gems and matching earrings. The middle dwarf displays a fierce expression, sporting a bushy orange beard, and is also clad in similar dark armor with teal embellishments. The bottom dwarf, an older figure with a long white beard, is adorned in a royal dark outfit with gold accents and a small crown, clasping a glowing white orb. This trio of dwarves, each with distinctive fantasy armor, unites in a playful attempt to disguise their stature and nature, adding an element of adventure and mystery to the scene.
Working off of that idea of the totem formation ... "Create an image featuring three children in a totem pole formation that are trying to conceal their nature in a single oversized trench coat."
I suspect the orange beard came from the previous part in the session. But that might be an approach to take in trying to describe it in a way that can be used.
Current generation image generators don’t understand text like instructions as you’re trying to do, describing an object then placing it then setting the scene.
It’s more like a giant telescope of many lenses (the latents from the prompts) and you’re adjusting the lenses to bring a possible reality of many into focus.
Sure. Normally I try a few variants, but "lamb with seven horns" was what I tried when I made that post.
For what it's worth, I've previously asked in the Stable Diffusion Discord server for help generating a "lamb with seven horns and seven eyes" but the members there were also unsuccessful.
I mean, you can use a fork to make whipped cream, but it won't be easy and it's not the right tool for the job. Does that mean that the fork is useless?
I never said it was useless, just that it fails at this specific problem. One of my complaints with many of these image generation tools is that there's not much communication as to what should be expected from them, nor do they explain the areas where they're expected to succeed or fail.
Recently Claude began to allow generation of SVG drawings, and asking it to draw a unicorn and later add extra tails or horns worked correctly.
A fork exists in physical space and it's pretty intuitive to understand what it can do. These models exist within digital space and are incredibly opaque by comparison.
Here's the screenshot [0] that was shared with me. It's obviously pretty basic, but Claude understood the correct location for where the horns and tails should be located. This looks like a clear iterative improvement over older models.
You also might want to "clarify" that it is not open source (and neither are any of the other "open source" models). If you want to call it something, try "open weights", although the usage restrictions make even that a HUGE FUCKING STRETCH.
Also, everybody should remember that these models are not copyrightable and you should never agree to any license for them...
When I read "open source" i thought they actually are doing open source instead of "open weights" this time. Surely they would expect to be called out on hackernews if they label it incorrectly...
A personal bugbear is the AI fascination with calling themselves open source, virtue signalling I guess. Open weights is exactly right. Source code and arguably more important datasets are both required to replicate the work, which is more in the spirit of open source (and science). I think Meta is especially egregious here, given their history.
Never underestimate the value of getting hordes of unpaid workers to refine your product. (See also React, others)
It's certainly not true that models are not copyrightable; databases have copyright protection if creativity was involved in creating them.
That said, I don't think outputs of the model are derivative works of it, any more than the model is a derivative of its training data, so it's not clear to me they can actually enforce what you do with them.
I'm no IP lawyer, but I've always thought that copyright put "requirements" on the artefact (i.e the threshold of originality), not the process.
In my jurisdiction we have database rights, meaning that you get IP protections for the artefact based on the work put into the process. For example a database of distances between adress pairs or something is probably not copyrightable, but can be protected under database rights if enough work was done to compile the data.
EDIT: Saw in another place in thread speaking about the https://en.wikipedia.org/wiki/Sweat_of_the_brow doctrine, relates to Database rights. (Neither of which notably are not applicable in the U.S)
These models are a product of much more creativity than simply a list of phone numbers in a phone book. I don't see how they wouldn't meet the modicum of creativity required for US copyright protection.
The software that creates the model is the product of creativity. The model itself is the product of mechanically applying that software to datasets that are (a) assembled with minimal, if any creativity, and (b) definitely not assembled with any eye to the specific form of the resulting model. The whole point is to get the software to form the model without you having to worry about what the result is going to look like. So you can't turn around and claim that the model is a creative work because of the choice of training data.
The only thing that's really specified about the model itself is its architecture, which is (1) dictated by function, and (2) usually deeply stereotyped.
> mechanically applying that software to datasets that are (a) assembled with minimal, if any creativity, and (b) definitely not assembled with any eye to the specific form of the resulting model.
Fair enough, but those datasets are also primarily copyrighted material. If the software here merely transforms the input material (which I agree it does), then the output is a derivative work.
But the pitch-shifted song is still recognizably a creative work. It has identifiable, humanly comprehensible forms of all the original creative elements that Swift originally put into it (plus I guess a de minimis amount of extra creativity from the choice to pitch shift it).
If I take a string of data from a true hardware RNG, XOR it with a Taylor Swift song, and throw away the original random stream, is the resulting fundamentally random bit string still a derivative work of the song? As with the ML model, you can't recognize the song in it. And as with at least some training examples in the inputs of most ML models, you can't recover the song from it either.
It feels like the test for whether X is derivative for copyright purposes should include some kind of attention to whether X is a creative work at all. Maybe not, but then what test do you use?
I do recognize the possibility that the models might not themselves be eligible for copyright as independent works, yet still infringe copyright in the training inputs. It seems messy, but not impossible.
... and as I said elsewhere, it's also messy that while you generally can't recover every training input from the model, you can usually recover something very close to some of the training inputs.
> If I take a string of data from a true hardware RNG, XOR it with a Taylor Swift song, and throw away the original random stream, is the resulting fundamentally random bit string still a derivative work of the song? As with the ML model, you can't recognize the song in it. And as with at least some training examples in the inputs of most ML models, you can't recover the song from it either.
It's not a copy of it, and when you distribute it you're not distributing the original. So it's not a derivative for copyright purposes.
It can still be a derivative for other legal purposes. Judges don't appreciate it when you do funny math tricks like that and will see through them.
> It feels like the test for whether X is derivative for copyright purposes should include some kind of attention to whether X is a creative work at all. Maybe not, but then what test do you use?
Yes, that's how US copyright law works. (well sort of…)
Being a transformative work of something makes it less of a copy of it, the more transformed it is, since it falls under fair use exemptions or is clearly a different category of thing.
If a model was a derivative of its training data, then Google snippets/thumbnails would be derivatives of its search results and would be illegal too. Unless you wrote a new law to specifically allow them.
In other countries (Germany, Japan) fair use is weaker, but model training has laws specifically making it legal in certain circumstances, and presumably so do Google snippets.
> It's not a copy of it, and when you distribute it you're not distributing the original.
A compressed (or normally encrypted) version wouldn't be a copy that way, either, but I would still absolutely go down for distributing it. The difference is that the compression can be reversed to recover the original. Even lossy compression would create such a close derivative that nobody would probably even bother to make the distinction.
You're right that "math games" don't work in the law, but that cuts both ways. If you do something that truly makes the original unrecoverable and in fact undetectable, and if nothing salient to the legal issues at hand about the new version derives from the original, then judges are going to "see through" the "math trick" of pretending that it is a derivative.
> then Google snippets/thumbnails would be derivatives of its search results
Thumbnails are legally derivative works, in the US and probably most other places. In the US, they're protected by the fair use defense, and in other places they're protected by whatever carveouts those places have. But that doesn't mean they're not derivative works.
In fact, if I remember the US "taxonomy" correctly, thumbnails are infringing. It's just that certain kinds of ingfringement are accepted because they're fair use.
If thumbnails weren't derivative works at all, then the question of fair use wouldn't arise, because there can be no infringement to begin with if the putatively infringing work isn't either derivative or a direct copy.
Where thumbnails are different from ML models is that they're clearly works of authorship. In a thumbnail, you can directly see many of the elements that the author put into the original image it's derived from.
The questions are (a) whether ML models are works of authorship to begin with (I say they're not), and (b) whether something that's not a work of authorship can still be a derivative work for purposes of copyright infringment (I'm not sure about that).
So far as I know, neither one is the subject of either explicit legislation or definitive precedent in most of the world, including the US.
There's a lot of effort involved in the training runs too, and you might be able to get away with saying the ops engineers made creative choices too (of which checkpoints were good etc.)
Since it costs millions to produce one of these models, it's not just taking the software and running it to compile them.
I get the sentiment, but one of their models, albeit the worst one, is licensed under Apache without usage restrictions. The source to run the models is also open source.
It would be nice here if you give some examples of what you call open source model. Please ;) Because the impression is that these things do not exist, it's just a dream which does not deserve such a nice term..
As far as I know, none have been released. And it doesn't even really make sense, because, as I said, the models aren't copyrightable to begin with and therefore aren't licensable either.
However, plenty of open source software exists. The fact that open source models don't exist doesn't excuse attempts to falsely claim the prestige of the phrase "open source".
I can tell you a secret. What you call 'open source' models are impossible. Because massive randomness is a part of training process. They are not reproducible. Having everything you cannot even tell if the given model was trained on the given dataset. Copyright is a different thing.
And a bad news, what's coming is even worst. Those will be the whole things with self awareness and personal experience. They can be copied, but not reproduced. More over, it's hard or almost impossible to detect if something undeclared was planted in their 'minds'.
All together means 'open source' model in strict interpretation is a myth, great idea which happen to be not. Like Turing test.
> However, plenty of open source software exists.
Attempt to switch topic detected.
PS: as for that massive downvote, I even wasn't rude, don't care. This account will be abandoned soon regardless, like all before and after.
You are wrong about that. It's a file with numbers. Which makes it a database or dataset and very much protected by copyright. That's why licenses are needed. For the phone book, things like open street maps, and indeed AI models.
> The fact that open source models don't exist
The fact that many people (myself included) routinely download and use models distributed under OSI approved licenses (Apache V2, MIT, etc.) makes that statement verifiably wrong. And yes, I do check the license of stuff that I use as I work with companies that care about such matters.
> You are wrong about that. It's a file with numbers. Which makes it a database or dataset and very much protected by copyright. That's why licenses are needed. For the phone book, things like open street maps, and indeed AI models.
This is only true in jurisdictions that follow the sweat of the brow doctrine, where effort alone without creativity is considered enough for copyright. In other places, such as the USA, collections of facts are not copyrightable and a minimal amount of creativity is required for something to qualify as copyrightable. The phone book is an example that is often used, actually, to demonstrate the difference.
> Which makes it a database or dataset and very much protected by copyright.
Not every collection of numbers is a database, and a database is not the same thing as a dataset.
Databases have limited copyright-like protection in some places. Under TRIPS, that extends to only databases that are "creative by virtue of the selection or arrangement of their contents" or something along those lines. In the US they talk specifically about curation.
ML models do not meet either requirement by any reasonable interpretation.
> The fact that many people (myself included) routinely download and use models distributed under OSI approved licenses (Apache V2, MIT, etc.) makes that statement verifiably wrong.
The "source code" of an ML model is most reasonably interpreted as including all of the training data, which are never, ever available.
Now you know better.
[On edit: By the way, the people creating these works had better hope they're outside copyright, because if not, each one of them is a derivative work of (at least some large and almost impossible to identify subset of) its training data, so they need licenses from all the copyright holders of that training material, which few of them have or can get.]
If we stop unnecessarily anthropomorphizing software, I think it is plainly obvious these are derivative works. You take the training material, run it through a piece of software, and it produces an output based on that input. Just because the black box in the middle is big and fancy doesn't mean that somehow the output isn't a result of the input.
However, transformativeness is a factor in whether or not there is a fair-use exception for the derivative work. And these models are highly transformative, so this is a strong argument for their fair-use.
"Fair use" is pretty much entirely a US concept, and similar concepts in other countries aren't isomorphic to it.
The model does have a radically different form from its inputs. So you could easily imagine that being "transformative enough" for US fair use. A lot of the other fair use elements look pretty easy to apply, too. Although there's still the question of whether all the intermediate copies you made to create the model were fair use...
In fact, I'll even concede that a court could find that a model wasn't a derivative work of its inputs to begin with, and not even have to get to the fair use question. The argument would be that the model doesn't actually reproduce any of the creative elements of any particular training input.
I do think a finding like that would be a much bigger stretch than a finding that the model was copyrightable. I could easily see a world where the model was found derivative but was not found copyrightable. And it's actually not clear to me at all that the model has to be copyrightable to infringe the copyright in something else, so that's another mess.
Somewhat related, even if the model itself isn't infringing, it's definitely possible to have most models create outputs that are very similar to (some specific examples in) their training data... in ways that obviously aren't transformative. Outputs that might compete with the original training data and otherwise fail to be fair use. So even if the model is in the clear, users might still have to watch out.
I agree, but that can't happen with the vast majority of these models because they're trained on unlicensed data so they can't slap an open source license on the training data and distribute it.
I've decided to draw my personal line at Open Source Initiative compliance for the license they release the model itself under.
I respect the opinion that it's not truly open source unless they release the training data as well, but I've decided not to make that part of my own personal litmus test here.
My reasoning is that knowing something is "open source" helps me decide what I legally can or cannot do with it when building my own software. Not having access to the training data downs affect my legal rights, it just affects my ability to recompile myself. And I don't have millions of dollars of GPUs so that isn't so important to me, personally.
> that can't happen with the vast majority of these models because they're trained on unlicensed data
Tough beans? There's lots of actual software that can't be open source because it embeds stuff with incompatible restrictions, but nobody tries to redefine "open source" because of that.
... and, on a vaguely similar-flavored note, you'd better hope that the models you're using end up found to be noninfringing or fair use or something with respect to those "unlicensed data", because otherwise you're in a world of hurt. It's actually a lot easier to argue that the models aren't copyrightable than it is to argue that they're not derivative of the input.
> I've decided to draw my personal line at Open Source Initiative compliance for the license they release the model itself under.
You're allowed to draw your personal line about what you'll use anywhere you want, but that doesn't mean that you should try to redefine "open source" or support anybody who does.
There was a looong distracting thread a month ago about something similar, niche language, might have been Julia, had a package with the same name as $NEW_THING.
I hope this one doesn't stir as much discussion. It has 4000 stars, there isnt a large mass of people who view the world through the lens of "Flux is ML library". No one will end up in a "who is on first?" discussion because of it. If this line of argument is held sacrosanct, it ends up in an infinite loop until everyone gives up and starts using UUIDs.
i would give them a break, so many things exist in the tech sector that being completely original is basically impossible, unless you name your thing something nonsensical
also search engines are context aware, if your search history is full of julia questions, it will know what you're searching for
It would be nice to understand limits of the free tier. I couldn't find that anywhere. I see pricing, but I'm generating images without swiping my credit card.
If it's unlimited or "throttled for abuse," say that. Right now, I don't know if I can try it six times or experiment to my heart's desire.
Congrats Burkay - the model is very impressive. One area I’d like to see improved in a flux v2 is knowledge of artist styles. Flux cannot respond to requests asking for paintings in the style of David Hockney, Norman Rockwell, Edgar Degas, — it seems to have no fine art training at all.
I’d bet that fine art training would further improve the compositional skills of the model, plus it would open up a range of uses that are (to me at least) a bit more interesting than just illustrations.
It's "just" another diffusion model, although a very good one. Those people are probably in there even if its text encoder doesn't know about them. So you can find them with textual inversion.
>Flux cannot respond to requests asking for paintings in the style of David Hockney, Norman Rockwell
Does it respond to any names? I noticed SD3 removed all names to prevent recreating famous people but as a side effect lost the very powerful ability to infer styles from artist names too.
thanks for hosting the model! i created an account to try it out, you started emailing me with “important notice: low account balance - action required” and now it seems like there’s no way for me to unsubscribe or delete my account. is that the case? thanks!
It is very fast and very good at rendering text, and appears to have a text encoder such that the model can handle both text and positioning much better: https://x.com/minimaxir/status/1819041076872908894
The logo has the same exact copyrighted typography as the real Vanity Fair logo. I've also reproduced the same-copyrighted-typography with other brands with identical composition as copyrighted images. Just asking it "Vanity Fair cover story about Shrek" at a 3:2 ratio gives it a composition identical to a Vanity Fair cover very consistently (subject is in front of logo typography partially obscuring it)
The image linked has a traditional www watermark in the lower-left as well. Even something innocous as a "Super Mario 64" prompt shows a copyright watermark: https://x.com/minimaxir/status/1819093418246631855
If the training data includes a public blog post which has a screenshot of a vanity fair piece?
It's like GRRM complaining that LLMs can reproduce chunks of text from his books "they fed my novels into it" Oh yeah? It's definitely not all the parts of your book quoted in millions of places online, including several dedicated wiki style sites? That wouldn't be it, right?
Just to be clear: you're comparing the collapse of the creative restrictions which the state has cleverly branded "intellectual property" to... the holocaust?
Of all of the instances on HN of Godwin's law playing out that I've ever seen, this one is the new cake-taker.
This is like the fifth time I see someone paraphrasing Niemöller in an ai context, and it's exhausting. It's also near impossible to take the paraphraser seriously.
More to the point, AI is a tool. I could just as well infringe on vanity fair IP using ms-paint. Someone more artistic than me could make a oil-on-canvas copy of their logo too.
Or, to turn your own annoying "argument" against you:
First they came for AI models, and I did not speak out, because I wasn't using them. Then they came for Photoshop, and I did not speak out, because I had never learned to use it. Then they came for for oil and canvas, and now there are no art forms left for me.
Nobody at all is "coming for" fashion magazines, but you sure seem to be "coming for" AI. Whether you have any power or not is besides point.
Whether you are paraphrasing or referencing to a famous confessional poem dealing with the Holocaust, the only reasonable interpretation is that you're comparing with the Holocaust. Even if you were unaware of the phrases origins, that's how anyone who does know where it comes from will interpret it. See other comments drawing the same conclusion for reference.
Again. Ai is a tool. It can produce illegal material, just like a pencil can, or a brush with oil and canvas. How are they different? They are not.
You don't need an A100, you can get a used 32GB V100 for $2K-$3K. It's probably the absolute best bang-for-buck inference GPU at the moment. Not for speed but just the fact that there are models you can actually fit on it that you can't fit on a gaming card, and as long as you can fit the model, it is still lightyears better than CPU inference.
Much slower memory and limited parallelism. Gpu ÷- 8k pr more cuda cores vs +-16 on regular cpu. Less mem swapping between operations. Gpu much much faster.
Well, I was wondering about bias in the model, so I entered "a president" as the prompt. Looks like it has a bias alright, but it's even more specific than I expected...
Schnell is definitely worse in quality, although still impressive (it gets text right). Dev is the really good one that arguably outperforms the new Midjourney 6.1
What's the difference between pro and dev? Is the pro one also 12B parameters? Are the example images on the site (the patagonia guy, lego and the beach potato) generated with dev or pro?
I think they are mainly -dev and -schnell. Both models are 12B. -pro is the most powerful and raw, -dev is guidance distilled version of it and -schnell is step distilled version (where you can get pretty good results with 2-8 steps).
something about pro must be better than dev or it wouldn't be made API-only, but what exactly, how does guidance distilling affect pro it and what quality remains in dev?
I think they may have turned on the gating some time after this was submitted to HackerNews. Earlier this morning I definitely ran the model several times without signing in at all (not via GitHub, not via anything). But now it says "Sign in to run".
Tested it using prompts from ideogram (login walled) which has great prompt adherence. Flux generated very very good images. I have been playing with ideogram but i don't want their filters and want to have a similar powerful system running locally.
If this runs locally, this is very very close to that in terms of both image quality and prompt adherence.
> A captivating and artistic illustration of four distinct creative quarters, each representing a unique aspect of creativity. In the top left, a writer with a quill and inkpot is depicted, showcasing their struggle with the text "THE STRUGGLE IS NOT REAL 1: WRITER". The scene is comically portrayed, highlighting the writer's creative challenges. In the top right, a figure labeled "THE STRUGGLE IS NOT REAL 2: COPY ||PASTER" is accompanied by a humorous comic drawing that satirically demonstrates their approach. In the bottom left, "THE STRUGGLE IS NOT REAL 3: THE RETRIER" features a character retrieving items, complete with an entertaining comic illustration. Lastly, in the bottom right, a remixer, identified as "THE STRUGGLE IS NOT REAL 4: THE REMI
Otherwise, the quality is great. I stopped using stable diffusion long time ago, the tools and tech around it became very messy, its not fun anymore. Been using ideogram for fun but I want something like ideogram that I can run locally without any filters. This is looking perfect so far.
whenever I see a new model I always see if it can do engineering diagrams (e.g. "two square boxes at a distance of 3.5mm"), still no dice on this one. https://x.com/seveibar/status/1819081632575611279
Would love to see an AI company attack engineering diagrams head on, my current hunch is that they just aren't in the training dataset (I'm very tempted to make a synthetic dataset/benchmark)
It'll probably come suddenly. It has been fascinating to me watching the journey from Stable Diffusion 1 to 3. SD1 was a very crude model, where putting a word in the prompt might or might not add representations of the word to the image. Eg, using the word "hat" somewhere in the prompt might do literally nothing or suddenly there were hats everywhere. The context of the word didn't mean much to SD1.
SD2 was more consistent about the word appearing in the image. "hat" would add hats more reliably. Context started to matter a little bit.
SD3 seems to be getting a lot better at the idea of scene composition, so now specific entities can be prompted to wear hats. Not perfect, but noticeably improved from SD2.
Extrapolating from that, we're still a few generations from being able to describe things with the precision of an engineering diagram - but we're heading in the right direction at a rapid clip. I doubt there needs to be any specialist work yet, just time and the improvement of general purpose models.
Can’t you get this done via an LLM and have it generate code for mermaid or D2 or something? I’ve been fiddling around with that a bit in order to create flowcharts and datamodels, and I’m pretty sure I’ve seen at least one of those languages handle absolute positioning of object.
I have likewise been utterly unable to get it to generate images that look like preliminary rapid pencil sketches. Suggestions by experienced prompters welcome!
>> Would love to see an AI company attack engineering diagrams head on, my current hunch is that they just aren't in the training dataset (I'm very tempted to make a synthetic dataset/benchmark)
That seems like a good use for a speech driven assistant that know how to use PC desktop software. Just talk to a CAD program and say what you want. This seems like a long way off but could be very useful.
You're not. I'm surprised at their selections because neither the cooking one nor the beach one adhere to the prompt in very well, and that first one only does because it prompt largely avoids much detail altogether. Overall, the announcement gives the sense that it can make pretty pictures but not very precise ones.
Well, that's nothing new, but it doesn't matter to dedicated users because they don't control it just by typing in text prompts. They use ComfyUI, which is a node editor.
No directly. but it encourages iteration on the same seed and then on specific details rather than just trying different prompts on different seeds from scratch over and over
Sounds to me like it's an issue with their VLM captions creating very "pretty" but not actually useful captions. Like one of the example image prompts includes this absolute garbage:
> Convey compassion and altruism through scene details.
The quality is difficult to judge consistently as there's variants among seed with the same prompt. And then there's the problem of cherry picked examples making the news. So I'm building a community gallery to generate Pro images for free, hope this at least increases the sample size https://fluxpro.art/
I have seen a lot of promises made by diffusion models.
This is in a whole different world. I legitimately feel bad for the people still a StabilityAI.
The playground testing is really something else!
The licensing model isn’t bad, although I would like to see them promise to open up their old closed source models under Apache when they release new API versions.
The prompt adherence and the breadth of topics it seems to know without a finetune and without any LORAs, is really amazing.
Vast majority of comparisons aren't really putting these new models through their paces.
The best prompt adherence on the market right now BY FAR is DALL-E 3 but it still falls down on more complicated concepts and obviously is hugely censored - though weirdly significantly less censored if you hit their API directly.
I quickly mocked up a few weird/complex prompts and did some side-by-side comparisons with Flux and DALL-E 3. Flux is impressive and significantly performant particularly since both the dev/shnell models have been confirmed by Black Forest to be runnable via ComfyUI.
I did put them through pro/dev as well just to be safe. The quality changes and you can play with guidance (cranking it all the way to 10) but it doesn't make a significant difference for these prompts from what I could tell.
Several iterations and these were the best I got out of schnell, dev and pro respectively for the following prompt:
"a fantasy creature with the body of a dragon and a beachball for a head, hybrid, best quality, shadows and lighting, fantasy illustration muted"
How long until nsfw fine tunes? Don’t pretend like it’s not on all of y’all’s minds, since over half of all the models on Civit.ai are NSFW. That’s what folks in the real world actually do with these models.
Mmmh, trying my recent test prompts, still pretty shit. F.e. whereas midjourney or SD do not have a problem to create a pencil sketch, with this model (pro), it always looks more like a black and white photograph or digital illustration or render. It is also like all the others apparently not able to follow instructions on the position of characters. (i.e. X and Y are turned away from each other).
It does provide information. Regardless of whether they use a post-inference filter, we now know that the model itself was trained on and can produce NSFW content. Compare this to SD3 which produces a noise pattern if you request naked bodies.
(Also you can download the model itself to check the local behaviour without extra filters. Unfortunately I don't have time to do it right now, but I'd love to know)
Hey, great work over at fal.ai to run this on your infrastructure and for building in a free $2 in credits to try before buying. For those thinking of running this at home, I'll save you the trouble. Black Forest Flux did not run easily on my Apple Silicon MacBook at this time. (Please let me know if you have gotten this to run for you on similar hardware.) Specifically, it falls back to using CPU which is very slow. Changing device to 'mps' causes error "BFloat16 is not supported on MPS"
Photo of teen girl in a ski mask making an origami swan in a barn. There is caption on the bottom of the image: "EAT DRUGS" in yellow font. In the background there is a framed photo of obama
The way someone explained it to me is that text-to-image models are essentially just de-noisers.
They train them by taking an image with a label, ie, "cat", and then adding some noise to it, run a training step, add more noise, run another step, and so on until the image is total (or near total) noise and still being told it's a cat.
Then, when you want to generate "cat", you start with noise, and it finds a cat in the noise and cancels some of the noise repeatedly. If you're able to watch an image get generated, sometimes you'll even see two cats on top of each other, but one ends up fading away.
Turns out, these denoisers don't require that many parameters, and if your resulting image has a few pixels that are just a tiny bit off color, you won't even notice.
I wonder if the key behind the quality of the MidJourney models, and this models, is less about size + architecture and more about the quality of images trained on.
It looks like this is the case for LLMs, that the training quality of the data has a significant impact on the output quality of the model, which makes sense.
So the real magic is in designing a system to curate that high quality data.
Midjourney unquestionably has heavy data set curation and uses RLHF from users.
You don't have to speculate on this as you can see that custom models for SDXL for instance perform vastly better than vanilla SDXL at the same number of parameters. It's all data set and tagging.
That is technically true, but when the base model is wasting parameter information on poorly tagged, watermarked stock art and other garbage images, it's not really a meaningful distinction. Better data makes for better models, nobody cares about how well a model outputs trash.
Ok, but you're severely misrepresenting the importance of things. Base SDXL is a fine model. Base SDXL is going to be much better than a materially smaller model that you've retrained with "good data".
It's the quality of the image text pair not the image alone but midjourney is not a model it's a suite of models that work in conjunction. They have an llm in the front to optimize the user prompts, they use SAM models, controlnet models for poses that are in high demand and so much more. That's why you can't really compare foundation models anymore because there are none.
No, it’s definitely the size. Tiny LLMs are shit. Stable Diffusion 3’s problem is not that that its training set was wildly different, it’s that it’s just too small (because the one released so far is not the full size).
You can get better results with better data, for sure. And better architecture, for sure. But raw size is really important the difference in quality for models, all else held equal, is HUGE and obvious if you play with them.
I would agree - midjourney is getting a free labour since many of their generations are not in secret mode (require pro/mega subscription) so prompts and outputs are visible to everyone. Midjourney rewards users to rating those generations. I wouldn't be surprised if there are some bots on their discord that are scraping those data for training their own models.
I enter an elaborate prompt, press "Sign in to Run", sign in with my GH, get taken back to the previous page and my prompt text has reset to some default with no way to get back what I entered before.
Complete and utter UX/first impression fail. I had no desire to actualy try the model after this.
Is the architecture outlined anywhere? Any publications or word on if they will publish something in the future? To be fair to them, they seemed to have launched this company today so I doubt they have a lot of time right now. Or maybe I just missed it?
I don't have anything to compare it to as I'm not that familiar with other diffusion models in the first place. I was kind of hoping to read the key changes they made to the diffusion architecture and how they collected and curated their dataset. I'd assume their are also using LAION but I wonder if they are doing anything to filter out low quality images (separate from what LAION atheistic already does). Or maybe if they have their own dataset.
Holy crap this is amazing. I saw an image with a prompt on reddit and didn't believe it was generated imaged. I thought it must be joke that people are sharing non-generated images in the thread.
> Photo of Criminal in a ski mask making a phone call in front of a store. There is caption on the bottom of the image: "It's time to Counter the Strike...". There is a red arrow pointing towards the caption. The red arrow is from a Red circle which has an image of Halo Master Chief in it.
Some of the images I generated using schnell model with 8-10 steps using this prompt. https://imgur.com/a/3mM9tKf
I'm really impressed at its ability to output pixel art sprites. Maybe the best general-purpose model I've seen capable of that. In many cases its better than purpose-built models.
These venture funded startups keep releasing models for free without a business model in sight. I am all for open source but worry it is not sustainable long term.
So I'm forced to signup and give my email for a supposed trial, only to be immediately told by email that I have a "Low Account Balance - Action Required"? Seriously?
Someone on Reddit did promptless img2img in comfy by passing an image into vae decode and then thru the schnell model for a kind of a refiner, with great results
Try something like "Photo of...", "as photography" or "photorealistic". You can even specify the camera model and lens/exposure settings. You can find these in metadata of your phone photos for example.
Thanks for the reply buddy. But I am still not able to understand how camera model and lens/exposure settings can be used to make the photo real.
Lets, say that you took an image of a flower in a garden and ai has also generated an image of the same flower. When we see these pics side by side we find a lot of difference between them. Origin of my question was "how can we minimize this difference ?". How we can tell the machine that the more the magnitude of a certain parameter the more real it is, not sure if camera settings could help in this case.
Works great as is right now, I can see some workflows being affected or having to wait for an update, but even those can do with some temporary workarounds (like having to load another model for later inpainting steps).
So if you're wanting to experiment and have a 24GB card, have at it!
Yeah I mean like controlnet / ipadapter / animateddiff / in painting stuff
I don’t feel like base models are super useful. Most real use cases depend on being able to iterate on consistent outputs imo.
I have had a very bad experience trying to use other models to modify images but I mostly do anime shit and maybe styles are less consistently embedded into language for those models
"Photo of Criminal in a ski mask making a phone call in front of a store. There is caption on the bottom of the image: "It's time to Counter the Strike...". There is a red arrow pointing towards the caption. The red arrow is from a Red circle which has an image of Halo Master Chief in it."
I think they may have turned on the gating some time after this was submitted to HackerNews. Earlier this morning I definitely ran the model several times without signing in at all (not via GitHub, not via anything). But now it says "Sign in to run".
what we did at fal is take the model and run it on our inference engine optimized to run these kinds of models really really fast. feel free to give it a shot on the playgrounds. https://fal.ai/models/fal-ai/flux/dev