FLUX.1 Kontext
bfl.ai
bfl.ai
It's pretty good: quality of the generated images is similar to that of GPT-4o image generation if you were using it for simple image-to-image generations. Generation is speedy at about ~4 seconds per generation.
Prompt engineering outside of the examples used on this page is a little fussy and I suspect will evolve over time. Changing styles or specific aspects does indeed work, but the more specific you get, the more it tends to ignore the specifics.
May I ask on which GPU & VRAM?
edit: oh unless you just meant through huggingface's UI
Something to be said about distributors like Replicate etc that are adding an exponent to the impact of these model releases
Llama 4 was another recent case where they explicitly worked with downstream distributors to get it working Day 1.
Replicate has a much bigger model selection. But for every model that's on both, FAL is pretty much "Replicate but faster". I believe pricing is pretty similar.
They'll have one of the victors, whoever it is. Maybe multiple.
If something's not as fast let me know and we can fix it. ben@replicate.com
Totally frank and possibly awkward question, you don't have to answer: how do you feel about a16z investing in everyone in this space?
They invested in you.
They're investing in your direct competitors (Fal, et al.)
They're picking your downmarket and upmarket (Krea, et al.)
They're picking consumer (Viggle, et al.), which could lift away the value.
They're picking the foundation models you consume. (Black Forest Labs, Hedra, et al.)
They're even picking the actual consumers themselves. (Promise, et al.)
They're doing this at Series A and beyond.
Do you think they'll try to encourage dog-fooding or consolidation?
The reason I ask is because I'm building adjacent or at a tangent to some of this, and I wonder if a16z is "all full up" or competitive within the portfolio. (If you can answer in private, my email is [my username] at gmail, and I'd be incredibly grateful to hear your thoughts.)
Beyond that, how are you feeling? This is a whirlwind of a sector to be in. There's a new model every week it seems.
Kudos on keeping up the pace! Keep at it!
I'm deep in this space and feel really good about FLUX.1 Kontext. It fills a much-needed gap, and it makes sure that OpenAI / Google aren't the runaway victors of images and video.
Prior to gpt-image-1, the biggest problems in images were:
- prompt adherence
- generation quality
- instructiveness (eg. "put the sign above the second door")
- consistency of styles, characters, settings, etc.
- deliberate and exact intentional posing of characters and set pieces
- compositing different images or layers together
- relighting
OpenAI's gpt-image-1 solved all of these things with a single multimodal model. You could throw out ComfyUI and all the other pre-AI garbage and work directly with the model itself. It was magic.
Unfortunately, gpt-image-1 is ridiculously slow, insanely expensive, highly censored (you can't use a lot of copyrighted characters or celebrities, and a lot of totally SFW prompts are blocked). It can't be fine tuned, so you're suck with the "ChatGPT style" and (as is called by the community) the "piss filter" (perpetually yellowish images).
And the biggest problem with gpt-image-1 is because it puts image and text tokens in the same space to manipulate, it can't retain the exact precise pixel-precise structure of reference images. Because of that, it cannot function as an inpainting/outpainting model whatsoever. You can't use it to edit existing images if the original image mattered.
Even with those flaws, gpt-image-1 was a million times better than Flux, ComfyUI, and all the other ball of wax hacks we've built up. Given the expense of training gpt-image-1, I was worried that nobody else would be able to afford to train the competition and that OpenAI would win the space forever. We'd be left with only hyperscalers of AI building these models. And it would suck if Google and OpenAI were the only providers of tools for artists.
Black Forest Labs just proved that wrong in a big way! While this model doesn't do everything as well as gpt-image-1, it's within the same order of magnitude. And it's ridiculously fast (10x faster) and cheap (10x cheaper).
Kontext isn't as instructive as gpt-image-1. You can't give it multiple pictures and ask it to copy characters from one image into the pose of another image. You can't have it follow complex compositing requests. But it's close, and that makes it immediately useful. It fills a much-needed gap in the space.
Black Forest Labs did the right thing by developing this instead of a video model. We need much more innovation in the image model space, and we need more gaps to be filled:
- Fast
- Truly multimodal like gpt-image-1
- Instructive
- Posing built into the model. No ControlNet hacks.
- References built into the model. No IPAdapter, no required character/style LoRAs, etc.
- Ability to address objects, characters, mannequins, etc. for deletion / insertion.
- Ability to pull sources from across multiple images with or without "innovation" / change to their pixels.
- Fine-tunable (so we can get higher quality and precision)
Please build it, Black Forest Labs.
All of those feature requests stated, Kontext is a great model. I'm going to be learning it over the next weeks.
Keep at it, BFL. Don't let OpenAI win. This model rocks.
Now let's hope Kling or Runway (or, better, someone who does open weights -- BFL!) develops a Veo 3 competitor.
I need my AI actors to "Meisner", and so far only Veo 3 comes close.
Thanks for the detailed info
Im building a web based paint/image editor with ai inpainting etc
and this is going to be a great model to use price wise and capability wise
completely agree so happy its not any one of these big co’s controlling the whole space!
OpenAI models are expensive to train because it’s beneficial for OpenAI models to be expensive and there is no incentive to optimize when they’re gonna run in a server farm anyway.
Probably a bunch of teams never bothered trying to replicate Dall-E 1+2 because the training run cost millions, yet SD1.5 showed us comparable tech can run on a home computer and be trained from scratch for thousands or fine tuned for cents.
Glad to see this release. It also puts more pressure onto OpenAI to make their model less lobotomized and to increase its output quality. This is good for everyone.
I haven't played around with from-scratch generation, so I'm not sure which is best if you're trying to generate an image just from a prompt. But in terms of image-to-image via a prompt, it feels like FLUX is noticeably better.
It generated (via a prompt) an image of a space ship landing on a remote planet.
I asked an edit, "The ship itself should be more colourful and a larger part of the image".
And it replaced the space-ship with a container vessel.
It had the chat history, it should have understood I still wanted a space-ship, but it dropped the relevant context for what I was trying to achieve.
https://replicate.com/flux-kontext-apps
I've thrown half a dozen pictures of myself at it and it just completely replaced me with somebody else. To be fair, the final headshot does look very professional.
> Preserve Intentionally
> Specify what should stay the same: “while keeping the same facial features”
> Use “maintain the original composition” to preserve layout
> For background changes: “Change the background to a beach while keeping the person in the exact same position”
So while the marketing seems to paint a picture that it'll preserve things automatically, and kind of understand exactly what you want changed, it doesn't seem like that's the full truth. You need to instead be very specific about what you want to preserve.
I liked keeping Flux 1.D around just to have a nice baseline for local GenAI capabilities.
https://genai-showdown.specr.net
Incidentally, we did add the newest release of Hunyuan's Image 2.0 model but as expected of a real-time model it scores rather poorly.
EDIT: In fairness to Black Forest Labs this model definitely seems to be more focused on editing capabilities to refine and iterate on existing images rather than on strict text-to-image creation.
A knight with a sword in hand stands with his back to us, facing down an army. He holds his shield above his head to protect himself from the rain of arrows shot by archers visible in the rear.
Adding it would also provide a fair assessment for a leading open source model.
The site is a great idea and features very interesting prompts. :)
So they can take it and run with it. (No contributing back either).
Open-weights, distilled variant of Kontext, our most advanced generative image editing model. Coming soon" is what they say on https://bfl.ai/models/flux-kontext
On HN, generally, people are more into technical discussion and/or productizing this stuff. Here, it seems declasse to mention the gooner angle, it's usually euphemized as intense reactions about refusing to download it involving the words "censor"
That's not the same as a diffusion model.
Here is a post about the difference that seems right at first glance: https://diffusionflow.github.io/
THIS MODEL ROCKS!
It's no gpt-image-1, but it's ridiculously close.
There isn't going to be a moat in images or video. I was so worried Google and OpenAI would win creative forever. Not so. Anyone can build these.
First time native image gen was introduced in Gemini 1.5 Flash if I'm not wrong, and then OpenAI was released for 4o which took over the internet by Ghibli Art.
We have been getting good quality images from almost all image generators like Midjourney, OpenAI and other providers, but the thing that made it special was true "multimodal" nature of it. Here's what I mean
When you used to ask chatgpt to create an image, it will rephrase that prompt and internally send that prompt to Dalle, similarly gemini would send it to Imagen which were diffusion models and they had little to know context in your next response about what's there in the previous image
In native image generation, it understands Audio, Text and even Image tokens in the same model and need not to rely on diffusion models internally, I don't think both Openai and google has released how they've trained it but my guess is that it's partially auto-regressive and diffusion but not sure about it
The people behind flux are the authors of stable diffusion paper that dates back to 2022.
Openai initially had dallee but stable diffusion was a massive improvement on dallee.
Then openai inspired itself from stable diffusion for gpt image
I imagine this is for anti-jailbreak moderation reasons, which is understandable
I spent two days trying to train a LoRa customization on top of Flux 1 dev on Windows with my RTX 4090 but can’t make it work and I don’t know how deep into this topic and python library I need to study. Are there scripts kiddies in this game or only experts ?
Sometimes behind patreon if some YouTuber
I was able to run this script to train a Lora myself without spending any time learning the underlying python libraries.
So you probably better off using Ai-ToolKit https://github.com/ostris/ai-toolkit
Windows is mostly the issue, to really take advantage, you will need linux.
The main thing is having 1. Good images with adequate captions and 2. Knowing what settings to use.
Number 2 is much harder because there’s a lot of bad information out there and the people who train a ton of Loras aren’t usually keen to share. Still, the various programs usually have some defaults that should be acceptable.
I like that they are testing face and scene coherence with iterated edits -- major pain point for 4o and other models.
Investigators will love this for “enhance”. ;)
I guess my main point is "this is where you draw the line? at a mostly accurate reconstruction of a partial of someone's face?" this was science fiction a few years ago. Training the model to accept two images (which it can, just not for explicit purposes of reconstructing (although it learns that too )) seems like a very task-specific, downstream way to handle this issue. This field is now about robust, general ways to emerge intelligent behavior not task specific models.
sure you could tell AI to remove the snow and some face will be revealed, but who is to say it's accurate? that's why traditionally you have a reference input.
As I stated a few times, the model HAS SUPPORT FOR MULTIPLE IMAGES. The article here doesn't try your very specific reference-image-benchmark but that doesn't mean you can't do it yourself - and it also doesn't imply there's anything wrong with the article or BFL - they're merely presenting a common usecase - not defining how the model should be used.
I notice American text2image models tend to generate less attractive and more darker skinned humans where as Chinese text2image generate attractive and more light skinned humans.
I think this is another area where Chinese AI models shine.
This seems entirely subjective to me.
> where as Chinese text2image generate attractive and more light skinned humans.
Are you saying they have chosen Asian traits that Asian beauty standards fetishize that in the West wouldn't be taken seriously at all? ;) There is no ground truth here that would be more correct one way or the other.
Of course, given the sensitivity of the topic it is arguably somewhat inappropriate to make such observations without sufficient effort to clarify the precise meaning.
> Chinese text2image generate attractive and more light skinned humans. > I think this is another area where Chinese AI models shine.
That is racism. There simply is no other way to classify it.
This particular woman looks Vietnamese to me, but I agree nothing about her appearance looks like anyone's fashion I know. But I only know California ABGs so that doesn't mean much.
I do have a "works on my machine"* :) -- prompt "Model F keyboard", all settings disabled, on the smaller model, seems to have substantially more than no idea: https://imgur.com/a/32pV6Sp
(Google Images comparison included to show in-the-wild "Model F keyboard", which may differ from my/your expected distribution)
* my machine, being, https://playground.bfl.ai/ (I have no affiliation with BFL)
https://commons.wikimedia.org/wiki/Category:IBM_Model_F_Keyb...
Specifying "IBM Model F keyboard" and placing it in quotation marks improves the search. But the front page of the search is tip-of-the-iceberg compared to whatever the model's scrapers ingested.
Eventually you may hit trademark protections. Reproducing a brand-name keyboard may be as difficult as simulating a celebrity's likeness.
I'm not even sure what my friends look like on Facebook, so it's not clear how an AI model would reproduce a brand-name keyboard design on request.
Another way of looking at it is, insistence on complete verisimilitude in an image generator is fundamentally in error.
I would argue, even undesirable. I don't want to live in a world where a 45 year old keyboard that was only out for 4 years is readily imitated in every microscopic detail.
I also find myself frustrated, and asking myself why.
First thought that jumps in: it's very clear that it is in error to say the model has no idea, modulo there's some independent run that's dramatically different from the only one offered in this thread.
Second thought: if we're doing "the image generators don't get details right", it would seem to be there a lot simpler examples than OPs, and it is better expressed that way - I assume it wasn't expressed that way because it sounds like dull conversation, but it doesn't have to be!
Third thought as to why I feel frustrated: I feel like I wasted time here - no other demos showing it's anywhere close to "no idea", its completely unclear to me whats distinctive about a IBM Model F Keyboard, and the the wikipedia images are worse than Google's AFAICT.
There are different sorts of details though, and the distinctions are both useful and interesting to understanding the state of the art. If "man drinking coke" produces someone with 6 fingers holding a glass of water that's completely different from producing someone with 5 fingers holding a can of pepsi.
Notice that none of the images in your example got the function key placement correct. Clearly the model knows what a relatively modern keyboard is, and it even has some concept of a vaguely retro looking mechanical keyboard. However indeed I'm inclined to agree with OP that it has approximately zero idea what an "IBM model F" keyboard is. I'm not sure that's a failure of the model though - as you point out, it's an ancient and fairly obscure product.
Then the law is broken. Monetizing someone's likeness is an issue. Utilizing trademarked characteristics to promote your own product without permission is an issue. It's the downstream actions of the user that are the issue, not the ML model itself.
Models regurgitating copyrighted material verbatim is of course an entirely separate issue.
> Then the law is broken.
> Utilizing trademarked characteristics to promote your own product without permission is an issue.
It sounds like you agree with parent that if your product reproduces trademark characteristics, it is utilizing trademarked characteristics. Just about at what layer you don't have responsibility. And the layer that has responsibility is the one that profits unjustly from the AI.
I'm interested if there's an argument for saying only the 2nd party user of the 1st party AI model, selling AI model output, to a 3rd party, is intuitively unfair.
I can't think of one. e.g. Disney launches some new cartoon or whatever. 1st party Openmetagoog, trains on it to make my "Video Episode Generator" product. Now, Openmetagoogs Community Pages are full of 30m video episodes made by their image generator. They didn't make them, nor do they promote them. Inuitively, Openmetagoog a competitor for manufacturing my IP, and that is also intuitively wrong. Your analysis would have us charge the users for sharing the output.
I wouldn't agree with that, no. To my mind "utilizing" generally requires intent at least in the context we're discussing here (ie moral or legal obligations). I'd remind you that the entire point of trademark is (approximately) to prevent brand confusion within the market.
> Your analysis would have us charge the users for sharing the output.
Precisely. I see it as both a matter of intent and concrete damages. Creating something (pencil, diffusion model, camera, etc) that could possibly be used in a manner that violates the law is not a problem. It is the end user violating the law that is at fault.
Imagine an online community that uses blender to create disney knockoffs and shares them publicly. Blender is not at fault and the creation of the knockoffs themselves (ie in private) is not the issue either. It's the part where the users proceed to publicly share them that poses the problem.
> They didn't make them, nor do they promote them.
By the same logic youtube neither creates nor promotes pirated content that gets uploaded. We have DMCA takedown notices for dealing with precisely this issue.
> Inuitively, Openmetagoog a competitor for manufacturing my IP, and that is also intuitively wrong.
Let's be clear about the distinction between trademark and copyright here. Outputting a verbatim copy is indeed a problem. Outputting a likeness is not, but an end user could certainly proceed to (mis)use that output in a manner that is.
Intent matters here. A product whose primary purpose is IP infringement is entirely different from one whose purpose is general but could potentially be used to infringe.
Disclaimer: I work for BFL
As with almost any AI release though, unless it’s open weights, I don’t care. The strengths and weaknesses of these models are apparent when you run them locally.