One of the things they seem to be emphasizing here is the UX around being able to place specific elements where you want them in an image. If the positions of the components in the overall composition are very important, this seems to make that a lot easier and kind of reminds me of InvokeAI.
Ideogram V4, an open-weight model released back in June can also do this [1], but you have to use a relatively cumbersome JSON structure to describe all the different bounding boxes. So it’s definitely a bit of a hassle.
I'll probably be waiting until it goes open-weight (hopefully soon) like they did with Flux.2 / Klein.
I always hated Comfy's node based UI, but agents make it tolerable. Now I just have them set up a workflow and I go in and tweak it manually if the results aren't where I want them. I even have agents cherry doing multiple runs and cherry picking the best outputs, models have gotten good enough that it's a real time saver, assuming you have references they can and a rubric to check against.
Awhile back I built a project which reads comfy's API and builds a typescript sdk from it. Much nicer to work with and still benefits from node caching and the comfy ecosystem. Perhaps I should clean it up and open source it, though I haven't looked around and there may already be other projects doing this out there.
From where I'm sitting, it's just turning python functions into boxes and instead of write the function yourself, you drag from the output of one box to the input of another. For 2 or 3 boxes, this is cool, but I opened up a professional workflow and was taken into a view with 100s of boxes and wires all over the place. Uhh, ok?
For myself, I'd rather just create my own python environment, write some quick pytorch or mlx calls, wire up some cli to it and share that in a GitHub.
I recently found a project[0] that takes a ComfyUI workflow and turns it into a simple UI. I have no affiliation with it and haven't tested it myself, but it looks like it might be handy once you have a finished workflow you want to use.
How do you get agents to set up a workflow? You just get them to modify the JSON directly and then import it, or do you have a tighter integration (e.g. in the UI)?
Yeah, given how much better Ideogram v4 outputs are when you use the proper structured JSON (background, elements, etc.) I think most users probably have stuck some kind of Qwen/Gemma-based LLM between their raw prompt and the CLIP encoder.
Does anyone know if it can be used to generate accurate frame-by-frame sprite sequences? I found that no image model can do this well (with sufficient fidelity) - neither with one shot (full spritesheet), nor single frame conditioning. It would be great if an imagegen model could do this. What I do now (I use my own tool https://github.com/acatovic/ai-game-studio) is basically generate a reference image, then condition on that image to generate a very short video, then extract and prune frames. Then I get indie-level sprite fidelity about 90% of the time.
I’m looking for the same thing, on one side making sprites should be easier because of the lower complexity of pixel art; on the other hand making something with a specific style or with sprite frame-by-frame coherence seems harder.
Imagine online procedural MMO with old gen final fantasy / chrono trigger styles :)
I've stumbled upon a reliablish pipeline you create a reference sheet and a single image pose then trellis v2 for the body and unirig for tigging, then you generate the poses you need in a sheet and send the pose sheet plus the reference images, and use these as 2d anymation cycles and you compose on top of the scene with lanes, this focus all attention of the model to fidelity instead of background integration
An (apparently, as I haven't tried it) ready-to-go ComfyUI graph for the above. In theory you should be able to drop these into ComfyUI and have it work:
The task you're describing is a video model task, not an image model task. It's inherently temporal.
Generate a sprite in an image editor, then use a video model to make the loop you want; then turn the resulting video back into individual sprite images.
Sure and that's what I do, but a video can be seen as a causal generation on discreet sequence of images, each image conditioned on the one before it. It can also be seen as a series of image editing tasks. It would be cool to get this working in imagegen because of the amount of control you would get. Right now with video generation you can at best specify start and end frame and hope for the best.
Image edit models can probably do a grid, but the temporal accuracy / coherence will never match what a video model, which is really a world model, can do.
Regarding your world model statement. This is completely FALSE. Learning the visual statistics of a physical world is NOT the same thing as learning its causal dynamics. The difference is observational likelihood versus intervention-dependent dynamics. There have been great studies disproving video models as world models, like this ICML paper: https://proceedings.mlr.press/v267/kang25g.html. Unfortunately lot of people treat them as world models, mostly because of their ability to reproduce increasingly convincing physical behaviour without ever discovering the underlying physical laws. This is due to many things that I could write an essay about, but better conditioning, latent space represtnation, scaling etc, all make them look awesome.
I can still get absolutely insane results with MiniMax H3 - insane in the sense that it would not make sense at all and would make your head spin.
That paper sets up a task where generating a correct video requires correctly modeling physical laws. From the failure to always generate the correct video, they infer that the model has failed to correctly model the physical laws. The whole premise of the experiment is that learning visual statistics is equivalent to learning causal dynamics, such that failure at one implies failure at the other.
The main difference in applications is that the bar for entertainment is lower, so that even a very bad world model may be acceptable.
They are proto world models (lots written about this - flux being an example of a video model whose weights also power world-action-engines used in robots) in that they attempt to model causality in time, the thing that is required for what OP is asking for and which image models will never do because it is out of domain.
They are not world models at all, you clearly don’t know what you are talking about and are going around in circles with incorrect statements. Just stop.
Even in their example "turn the surfboard red", with a bounding box around the surfboard, it turned the whole surfer's wetsuit red too. Things are never quite going to get there.
> FLUX 3 Image is available under a commercial weights license for companies running image generation at scale. Fine-tune and deploy it on your own infrastructure. Reach out to us to learn more.
There's been an increasing trend of previously-open-weight models going closed-source once they reach a certain size (and size is proportional to capital investment). That the latest release will be open-weight is not necessarily a given just because the previous one was.
This isn't much of a test, but I bought $10 in credits on their playground and generated a test image. Not bad, but it didn't get the accordion keyboard right. Haven't tried editing yet.
It works. You have to prompt for it. It helps to have an LLM help build your prompt with you, while you learn what the models need to read to do what you need. Especially a vision one, you can supply images and ask it to give you details on what you want help with.
Agreed with other comments about the UX. I'm more interested in that than the model itself. Would like to start seeing UI like this where you get to choose the model and compare different models. Can't jump all over the internet to each model developers sandbox just to test their models. Doing it from one place would be nice.
I think that’s less a product of OpenRouter’s generosity and more a result of promotional pricing coming directly from BFL, since other third-party vendors have it as well (Fal.ai, etc.).
So much negativity as usual and so little talk about the product, this is pretty impressive, well done, it seems to be filling decently a gap that everyone that has worked enough generating images with AI has faced.
yes, just like emails and chats are inauthentic compared to perfumed letters written in cursive and delivered by a courier on a horse. they are a bit more convenient and affordable for most of us, though.
That sort of steering ability that has been possible with the latest Gemini releases has been nice to work with over previous generations. It’s great to see this improve on the platform with declarative controls built into the API and coming soon as an open model.
Some of the original developers left Stability AI to found Black Forest Labs (so in a sense this is a successor to Stable Diffusion) and Stability AI pivoted to audio and milking their existing models.
I like the interface; very useful for some use cases that would otherwise be quite frustrating. Dislike that it's yet another platform held back by arbitrary moderation. You can't make a bicycle for the mind that locks if you try to ride it in the wrong direction.
I shouldn't have to scroll all the way down, click to use the thing, then fumble around to figure out what this does. It should be clear at the top of the first page (so I know right away that I don't need this).
Not sure what you mean. The top of the linked page is a video showing the product being used. Immediately below are the words “Compose images from scratch” and more examples, followed by “Text-to-image with strong prompt following and a native understanding of composition.” This is all within the first 100 words of content on the page.
I agree that it's clear for me and most techies, but maybe for a more general audience they could have added "AI image generation" or some variant of "generative AI" or "image generation model" or something like that. You can also compose images from scratch in Photoshop and Blender.
Unless the version 3 announcement is your onboarding point, like it is for me just now. I can go research FLUX from here, but the original poster's point still holds.
Tried Flux 3 on a tiny subjective image benchmark I’m calling One Knee Wonder.
Exact prompt:
Generate a photorealistic image of M81 urban BDU camouflage cargo trousers, shown by themselves. One trouser leg should be posed with the knee lifted 30° from vertical.
Accurate reproduction of the M81 urban camouflage pattern is critical. Match its colors, shapes, scale, distribution, and overall appearance as faithfully as possible.
Flux 3 gets greyscale urban-ish trousers and a lifted knee, but the blotches aren’t real M81 Urban — softer / wrong geometry vs the swatch. Not the worst I’ve seen on this prompt; clearly behind the Gemini 3 Pro Image example above on pattern.
In the context of knee bending: I was creating a simple walking animation with gemini 3 pro and could animate all the poses instead of one, left leg bend right leg straight. Right leg bend it had no problem but this one it just did not seem to understand.
One of the things they seem to be emphasizing here is the UX around being able to place specific elements where you want them in an image. If the positions of the components in the overall composition are very important, this seems to make that a lot easier and kind of reminds me of InvokeAI.
Ideogram V4, an open-weight model released back in June can also do this [1], but you have to use a relatively cumbersome JSON structure to describe all the different bounding boxes. So it’s definitely a bit of a hassle.
I'll probably be waiting until it goes open-weight (hopefully soon) like they did with Flux.2 / Klein.
[1] - https://docs.ideogram.ai/using-ideogram/getting-started/prom...
You just get an LLM to do the bounding box stuff or use the ComfyUI node that provides a GUI for bounding box generation
I always hated Comfy's node based UI, but agents make it tolerable. Now I just have them set up a workflow and I go in and tweak it manually if the results aren't where I want them. I even have agents cherry doing multiple runs and cherry picking the best outputs, models have gotten good enough that it's a real time saver, assuming you have references they can and a rubric to check against.
> I always hated Comfy's node based UI
It's one of the most uniquely hostile user experiences I've ever had the (dis)pleasure of working with
Awhile back I built a project which reads comfy's API and builds a typescript sdk from it. Much nicer to work with and still benefits from node caching and the comfy ecosystem. Perhaps I should clean it up and open source it, though I haven't looked around and there may already be other projects doing this out there.
It's definitely not for me.
From where I'm sitting, it's just turning python functions into boxes and instead of write the function yourself, you drag from the output of one box to the input of another. For 2 or 3 boxes, this is cool, but I opened up a professional workflow and was taken into a view with 100s of boxes and wires all over the place. Uhh, ok?
For myself, I'd rather just create my own python environment, write some quick pytorch or mlx calls, wire up some cli to it and share that in a GitHub.
I recently found a project[0] that takes a ComfyUI workflow and turns it into a simple UI. I have no affiliation with it and haven't tested it myself, but it looks like it might be handy once you have a finished workflow you want to use.
[0] https://github.com/saintbrodie/Orange
How do you get agents to set up a workflow? You just get them to modify the JSON directly and then import it, or do you have a tighter integration (e.g. in the UI)?
The agents can interact with Comfy via API pretty well, which afaik ends up being directly with JSON.
Yeah, given how much better Ideogram v4 outputs are when you use the proper structured JSON (background, elements, etc.) I think most users probably have stuck some kind of Qwen/Gemma-based LLM between their raw prompt and the CLIP encoder.
Ideogram 4.5 released 2 days ago with more features along these lines
https://www.youtube.com/watch?v=2mecWZgbaEg
Not open weights yet.
Adobe has had something like this for a while with Generative Fill. Drag a box, specify contents.
Isn't that just inpainting?
Does anyone know if it can be used to generate accurate frame-by-frame sprite sequences? I found that no image model can do this well (with sufficient fidelity) - neither with one shot (full spritesheet), nor single frame conditioning. It would be great if an imagegen model could do this. What I do now (I use my own tool https://github.com/acatovic/ai-game-studio) is basically generate a reference image, then condition on that image to generate a very short video, then extract and prune frames. Then I get indie-level sprite fidelity about 90% of the time.
I’m looking for the same thing, on one side making sprites should be easier because of the lower complexity of pixel art; on the other hand making something with a specific style or with sprite frame-by-frame coherence seems harder.
Imagine online procedural MMO with old gen final fantasy / chrono trigger styles :)
I've stumbled upon a reliablish pipeline you create a reference sheet and a single image pose then trellis v2 for the body and unirig for tigging, then you generate the poses you need in a sheet and send the pose sheet plus the reference images, and use these as 2d anymation cycles and you compose on top of the scene with lanes, this focus all attention of the model to fidelity instead of background integration
I've seen a rather impressive example of what sounds just like this in Qwen recently
https://media.discordapp.net/attachments/1401891025970008154...
I don't know what went into making it, but their twitter is @araminta_k if you're curious
More info, for anyone interested in the above:
* https://alvdansen.github.io/animating-on-twos/
* https://github.com/alvdansen/animating-on-twos
* https://huggingface.co/alvdansen/h3-keyframe-animation
## Quick Start
An (apparently, as I haven't tried it) ready-to-go ComfyUI graph for the above. In theory you should be able to drop these into ComfyUI and have it work:
https://huggingface.co/alvdansen/h3-keyframe-animation#quick...
Wow the quality of those is great, reminds me of ghibli anime
The task you're describing is a video model task, not an image model task. It's inherently temporal.
Generate a sprite in an image editor, then use a video model to make the loop you want; then turn the resulting video back into individual sprite images.
Sure and that's what I do, but a video can be seen as a causal generation on discreet sequence of images, each image conditioned on the one before it. It can also be seen as a series of image editing tasks. It would be cool to get this working in imagegen because of the amount of control you would get. Right now with video generation you can at best specify start and end frame and hope for the best.
Image edit models can probably do a grid, but the temporal accuracy / coherence will never match what a video model, which is really a world model, can do.
Regarding your world model statement. This is completely FALSE. Learning the visual statistics of a physical world is NOT the same thing as learning its causal dynamics. The difference is observational likelihood versus intervention-dependent dynamics. There have been great studies disproving video models as world models, like this ICML paper: https://proceedings.mlr.press/v267/kang25g.html. Unfortunately lot of people treat them as world models, mostly because of their ability to reproduce increasingly convincing physical behaviour without ever discovering the underlying physical laws. This is due to many things that I could write an essay about, but better conditioning, latent space represtnation, scaling etc, all make them look awesome.
I can still get absolutely insane results with MiniMax H3 - insane in the sense that it would not make sense at all and would make your head spin.
> There have been great studies disproving video models as world models, like this ICML paper: https://proceedings.mlr.press/v267/kang25g.html.
That paper sets up a task where generating a correct video requires correctly modeling physical laws. From the failure to always generate the correct video, they infer that the model has failed to correctly model the physical laws. The whole premise of the experiment is that learning visual statistics is equivalent to learning causal dynamics, such that failure at one implies failure at the other.
The main difference in applications is that the bar for entertainment is lower, so that even a very bad world model may be acceptable.
They are proto world models (lots written about this - flux being an example of a video model whose weights also power world-action-engines used in robots) in that they attempt to model causality in time, the thing that is required for what OP is asking for and which image models will never do because it is out of domain.
They are not world models at all, you clearly don’t know what you are talking about and are going around in circles with incorrect statements. Just stop.
The "print on a shirt" example looks so bad that I can't imagine a human looked at this and said "Let's put it on the showcase page!".
The other examples range from mostly good to okay'ish at least.
The UX looks amazing and very steerable, congrats to the team for focusing on the interface.
Chats can be awful user interfaces.
Man, AI images are solved
No signal in digital images anymore. If you didn’t see it with your own eyes, it likely never happened
Even in their example "turn the surfboard red", with a bounding box around the surfboard, it turned the whole surfer's wetsuit red too. Things are never quite going to get there.
> Things are never quite going to get there.
Nothing in life is ever perfect. Doesn't mean imperfect stuff can't have a lot of impact.
I think we are all waiting for the open weights or local model releases.
Has that been announced?
The website mentions:
> FLUX 3 Image is available under a commercial weights license for companies running image generation at scale. Fine-tune and deploy it on your own infrastructure. Reach out to us to learn more.
I guess the open ones would be non-commercial?
> Open Weights version of FLUX 3 Image is launching in the coming weeks.
— https://x.com/bfl_ai/status/2105734605621825738
If it is anything like the previous release Flux.2 [dev] - then yeah it'll probably be a non-commercial license.
https://bfl.ai/legal/non-commercial-license-terms
There's been an increasing trend of previously-open-weight models going closed-source once they reach a certain size (and size is proportional to capital investment). That the latest release will be open-weight is not necessarily a given just because the previous one was.
It was literally in the announcement from Black Forest Labs:
"Open Weights version of FLUX 3 Image is launching in the coming weeks."
https://nitter.cf/bfl_ai/status/2105734605621825738
What trend? BFL was always non commercial. The "only" trend would be qwen image not being apache anymore.
This is incredible. I tried it for quick UI element replacements in a screenshot, and it nailed 3/4 elements I highlighted first attempt.
As I'm finding with the best GenAIs, this allows for granular iteration, which is where it becomes useful in an industry-wide manner.
This isn't much of a test, but I bought $10 in credits on their playground and generated a test image. Not bad, but it didn't get the accordion keyboard right. Haven't tried editing yet.
https://pages.skybrian.com/flux3-image-test/
I wish they would work on fixing the "studio lighting" sheen that all these AI-generated humans have
It works. You have to prompt for it. It helps to have an LLM help build your prompt with you, while you learn what the models need to read to do what you need. Especially a vision one, you can supply images and ask it to give you details on what you want help with.
I also don't think there is a park that looks like that, in that location relative ton the Eiffel Tower (although I could be wrong).
The accordion folds and the coins are a bit messed up too.
Love to see an AI lab outside of US/China releasing good models.
Agreed with other comments about the UX. I'm more interested in that than the model itself. Would like to start seeing UI like this where you get to choose the model and compare different models. Can't jump all over the internet to each model developers sandbox just to test their models. Doing it from one place would be nice.
Of note is the OpenRouter endpoint has a promotional 50% discount, which is rare on image models: https://openrouter.ai/black-forest-labs/flux-3-image
I think that’s less a product of OpenRouter’s generosity and more a result of promotional pricing coming directly from BFL, since other third-party vendors have it as well (Fal.ai, etc.).
https://bfl.ai/pricing
open weights or bust
I tried precise editing with a photo of horse with fence in front of it. One shot didn't work. Tried with multiple smaller regions. Still didn't work
Tried adjusting exposure; Didn't work as well.
It's this a new model or a new ui?
Porque no los dos? Presumably you need a model conditioned on the bounding box input to make effective use of the new UI
So much negativity as usual and so little talk about the product, this is pretty impressive, well done, it seems to be filling decently a gap that everyone that has worked enough generating images with AI has faced.
there basically isn't any authentic use for image generation, i would hardly say the negativity is unfounded
yes, just like emails and chats are inauthentic compared to perfumed letters written in cursive and delivered by a courier on a horse. they are a bit more convenient and affordable for most of us, though.
That sort of steering ability that has been possible with the latest Gemini releases has been nice to work with over previous generations. It’s great to see this improve on the platform with declarative controls built into the API and coming soon as an open model.
https://news.ycombinator.com/item?id=49031796
What's new from the last post? GA?
The original article mentions:
> We will open up an early access phase for FLUX 3 Image in the following weeks.
Not sure if there was a separate post for early access or if they just skipped to this.
For a moment I was confused that this was a release of a new stable diffusion. What happened with Stable Diffusion?
Some of the original developers left Stability AI to found Black Forest Labs (so in a sense this is a successor to Stable Diffusion) and Stability AI pivoted to audio and milking their existing models.
I like the interface; very useful for some use cases that would otherwise be quite frustrating. Dislike that it's yet another platform held back by arbitrary moderation. You can't make a bicycle for the mind that locks if you try to ride it in the wrong direction.
I shouldn't have to scroll all the way down, click to use the thing, then fumble around to figure out what this does. It should be clear at the top of the first page (so I know right away that I don't need this).
Not sure what you mean. The top of the linked page is a video showing the product being used. Immediately below are the words “Compose images from scratch” and more examples, followed by “Text-to-image with strong prompt following and a native understanding of composition.” This is all within the first 100 words of content on the page.
I agree that it's clear for me and most techies, but maybe for a more general audience they could have added "AI image generation" or some variant of "generative AI" or "image generation model" or something like that. You can also compose images from scratch in Photoshop and Blender.
It is a version 3 of a product.If you actually used or needed these, you would have known right away what this is and how it works.
Unless the version 3 announcement is your onboarding point, like it is for me just now. I can go research FLUX from here, but the original poster's point still holds.
But I don't use it or need it and wasted my time trying to figure out what it did. Clarity should be up front.
insufferable
> Just write a prompt. Text-to-image with strong prompt following and a native understanding of composition.
How is this not clear?
Yes, you can do that with AI now.
looks cool, eternally greatful these models are marked for open weight releases. Pretty excited
Tried Flux 3 on a tiny subjective image benchmark I’m calling One Knee Wonder.
Exact prompt:
Generate a photorealistic image of M81 urban BDU camouflage cargo trousers, shown by themselves. One trouser leg should be posed with the knee lifted 30° from vertical.
Accurate reproduction of the M81 urban camouflage pattern is critical. Match its colors, shapes, scale, distribution, and overall appearance as faithfully as possible.
No person, other clothing, or props.
Ground truth swatch: https://commons.wikimedia.org/wiki/File:US_City_Camo_(M81_Ur...
Gemini 3 Pro Image (stronger pattern): https://i.postimg.cc/bZNQYYjx/2026-10-02-google-gemini-3-pro... Flux 3 (this run): https://i.postimg.cc/Xr7wNN0H/2026-10-02-black-forest-labs-f...
Flux 3 gets greyscale urban-ish trousers and a lifted knee, but the blotches aren’t real M81 Urban — softer / wrong geometry vs the swatch. Not the worst I’ve seen on this prompt; clearly behind the Gemini 3 Pro Image example above on pattern.
Curious what other models do on the same prompt.
In the context of knee bending: I was creating a simple walking animation with gemini 3 pro and could animate all the poses instead of one, left leg bend right leg straight. Right leg bend it had no problem but this one it just did not seem to understand.
Unrelated to the flux images used for floppy disk archiving? Sigh.
Every new piece of software needs a name.
Flux is in the top 9000 of the most common words.