People on reddit claim it's bad at generating images of people, apparently because they removed anything close to nsfw from the training set. Anyone already tried it?
I've been running it through tests on my local machine. The Text-to-Image component is VERY poor. It's not a good sign when the stock ComfyUI Flux3 Klein template with built-in config/prompts produces SD 1.5 levels of half-melted faces and various other body horror.
Flux 2 Klein 9b also had anatomical issues as well, but this one feels practically lobotomized.
I honestly don’t know what the hell BFL was thinking with this release. There’s idle conjecture that they may have been too heavy-handed when removing even the slightest risqué images from the training data, and as a result the final product has such a grossly poor understanding of basic human anatomy that no amount of LoRAs is going to be able to patch it up.
It’s bafflingly worse than other comparably sized ones: Qwen-Image 2.1 7B, Z-Image Turbo 6B which was released practically a year ago, even its own predecessor Flux 2 Klein 9b.
Even if we concede that the Klein variants are intended more as editor models, the results I've seen on Reddit aren't that impressive either.
I feel like they just handed their collective heads on a platter over to the team over at Krea especially since Krea 3 addresses the major limitation of Krea 2 which was strictly text-to-image only.
They don’t just filter out images in the dataset, but also have said they do targeted training and RL to neuter the model’s capability’s NSFW capability.
Safetymaxxing at the expense of benign output quality is sad to see.
European edit model cannot make it easy to nudify people… they lobotomize it, it’s trash, Mission Accomplished!
Another reason that China is pulling ahead daily. They can leave in some anatomy and not knee cap their own product, leave in IP, and give away the model so as to undercut API/credit sales of US and EU models.
Finally… the Neo Puritan death march is having a direct financial toll.
I gave Flux 3 Klein my typical human test that has never failed on newer image models and it did fail; in addition to lower quality, the fingers were definitely not the correct proportions.
Whether the cause is due to more safeguards against NSFW content is hard to confirm technically; it was more an intuition back in the Stable Diffusion XL days but I suspect modern training paradigms are better at correlative inference.
i think its instructive how the ancient sculptors and artists used to have to draw nudes in order to understand the human body. perhaps our image models need this
It isn't just the ancient sculptors and artists (unless 20 years ago is ancient :-P), when i went to an art school there were nude drawing lessons there too.
I've tried it some, only locally. I rarely use APIs, so can't say what the official inference API results look like. It correctly followed some prompts that almost every model I've ever tried performed poorly at, but failed in strange ways on prompts that other models handle with ease.
At high resolutions, structural integrity on people can definitely fall apart like older models, where you get one person turning into a long spaghetti spanning across the frame. It might be that you need to do grounding by generating some kind of scene reference at its stable resolutions first then having it use that for the larger resolution like older models too.
They say that their prompt "upscaler" model or system prompt are essentially necessary for the model, but my results have been hit or miss with that claim. Some basic prompts are better, some upscaled prompts are better. I think that if you gain an intuition for how the model understands things, you might get better results by default by hand crafting prompts specifically for it.
Some of the stylized art I'm getting out of it at higher resolutions are really nice and I prefer them over other models.
All of these models seem to have their own character in terms of system requirements, performance, prompt adherence, artistic outputs and so on. None of the models I have tried are a complete homerun, so maybe in this case the model feels a little inconsistent in some ways, but it's a model I will probably keep around.
It's editing capabilities have been really solid, but I've been disappointed with its ability to do finer text details (a niche use case for my own purposes). There's also a lot of reports of body horror outputs, but I haven't encountered any myself.
Still though, it's a lot like the porsche mission S vs the tesla roadster situation with the weights release. Moss Image dropped which looks far better at the moment, leading to a "why Flux?" question in the community.
Looks like the people over /r/SD seem to like MOSS (i'm not up to date on the image stuff) but what i notice is that it is Apache2 licensed when every other openweight model seems to have some weird license (from a quick check Qwen-Image even downright forbids any commercial use at all).
Yeah, unfortunately, a lot of them have been moving to their own bespoke, non-commercial licenses (Ideogram, Krea, Flux, Qwen 2, etc).
Annoyingly, the original Qwen-Image was actually Apache 2 before they switched over unfortunately.
In addition to Moss, Boogu Image 0.1 is a unified model which came out in June and is kind of being slept on right now, is actually pretty decent and falls under the Apache 2.0 license. It scored surprisingly high on my GenAI benchmark (comparable to Krea 2 and Ideogram 4).
Looks like MOSS was made by some Chinese university as part of an AI education/research program so i guess they don't have to care about funding as much as the other labs :-P
By memorizing the fact that the sun rises in the East and sets in the West, you no longer have to make complex predictions and draw arcs every day. By memorizing the fact that the Earth rotates, and that generally most bodies in solar systems do, you can extrapolate this fact to any planet or solar system.
This example is a rough comparison to what actually happens, but it does attempt to demonstrate what compression and knowledge are, and how patterns, symmetries, and extrapolation/interpolation can make describing/modeling a complex system much more efficient.
Models can even go a step further, offloading entropy to the input. An efficient model exploits predictable structure; the remaining details can be expensive to compress further. The input can supply whatever additional information is necessary to produce the desired output within some margin of error. The information required depends on the model’s efficiency over a given domain, the structure of that domain, and the required precision of the output.
Consider generating an image of a pebble driveway. Most models today will generate a plausible arrangement of pebbles using their learned distributions. Reproducing a particular arrangement requires the input to convey the details that distinguish it from other plausible arrangements. The more closely you want to reproduce it, the more informative the input needs to be—potentially including a reference image of the arrangement itself.
This certainly raises a lot of open questions that are fun to pursue, such as the smallest possible model that can recover a complete description of our universe from minimal input.
It kind of blows me away how “unreasonably effective” these small image models seem for their size.
Meanwhile a decent LLM that can do tool use and code and reason tends to weigh in at a minimum of 30B parameters, and you need over 100B to get good performance.
Do LLMs just contain more information or are diffusion models more efficient at compression?
I saw some work on language diffusion models a while back but I think they tend to have problems with reasoning versus transformer type models.
These models are, in a sense, distillations of language models. They wouldn't be as good without a pre-trained language model, so it's unfair to compare them.
Image models are essentially a mapping of text embedding space -> image. So the clearer and cleaner the text embedding space is the better model you get as you need less weights trying to interpret a text prompt, recent image models even use multimodal versions of LLMs (Flux uses Qwen3 VL) so all the visual separation was done beforehand making the model even more efficient.
edit: Maybe a better analogy is that the image models are only half the model in a encoder-decoder sense, where the LLM is a pretrained frozen encoder and the image model is the decoder.
To make a rather speculative argument - since an LLM should be able to describer an image losslessly, images are a subset of what an LLM needs to understand. It is possible that the statistical model being learned for images is just simpler.
The get a proper LLM they have to model all of human knowledge regarding patterns, physics, biology, social interaction, skills, history, etc, etc. For images they have to recall what a dog looks like in a still photo. The image models aren't even necessarily good at a wide range of novel situations; usually they get used for stock scenes.
Natural evolution has also had a much easier time coming up with vision models (everything with eyeballs) than complex language models (as far as we know only a smallish family of monkeys).
this is the right answer. ultimately text and general llm usage incl coding is far more internally complex than just memorizing the looks and composition of concepts placed in an image. dont forget that most of our reasoning benchmarks basically require the llm to have a search engine and calculator in its own weights.
These diffusion models are not just memorizing the looks and composition of concepts. These models are increasingly capable of representing advanced world knowledge. Things like shadows require a comprehensive, if indirect, understanding of physical light modeling. Things like weight/size estimation, etc., which are all important to a cohesive generated image of natural phenomenon, require an adequate understanding of physics.
Some of this understanding is in the input embeddings themselves, embedding models do a lot of heavy lifting behind image generation models.
I posit that you can theoretically achieve the same capability in a model trained on the right sequence of "images" as you can purely text; both are turned into embeddings and can ultimately convey the same exact relationships. I don't think it's very useful to try and create delineations between text and image inference, especially as more models move towards native multimodality.
> since an LLM should be able to describer [sic] an image losslessly
Any lossless text description of a given image would just be a structured text output of the exact pixel values, or an encoding that losslessly exploits regularities. Anything less is categorically lossy. An ordinary semantic description of a reasonably complex image will generally be lossy. A short, lossless text description would need to exploit regularities in the image, such as a solid-color image, or a basic pattern or sequence which is easily reproducible, such as a precisely specified checkerboard image.
And being able to represent images losslessly in text simply establishes text’s capacity to encode information. It doesn't establish an LLM’s holistic understanding of images. There is heavy lifting in the phrase “images are a subset.” An image can be almost any conceivable perceptual arrangement of data. The bounds are not clear. Maybe what you meant to suggest was that images of our natural world contain exploitable regularities.
> Natural evolution has also had a much easier time coming up with vision models (everything with eyeballs) than complex language models (as far as we know only a smallish family of monkeys).
It isn't appropriate to explain complex biological systems in terms of today's technologies. This is something people have been doing for a long time. The mind is like a a book, a loom, a camera, an engine, a computer, an LLM/vision model. Etc. The cognitive abilities required for an organism to meaningfully interpret its vision can possibly largely be the same cognitive abilities underlying speech, and all that is missing for a brain to exploit this might be sufficiently advanced vocal cords or manual dexterity; nothing to do with the brain itself, but its auxiliary organs.
At first, I considered that diffusion models struggle with reasoning because they underweight 'prior beliefs', ie the propositions by which an argument advances.
Then I considered whether the universe of imagery is simply smaller than the universe of ideas—which must, itself, include the universe of images, too.
But finally, I'm thinking maybe images are just harder to benchmark.
Or maybe they're _not_ actually worth 1,000 words.
When it comes to reasoning specifically, it's a far simpler explanation IMO. Modern RLVR techniques like GRPO rely on the model understanding the likelihood of the generated phrase. This is really easy to compute for LLMs that are trained with next-word prediction. Diffusion models have a harder time accessing this likelihood value (i.e., you can approximate it but the approximation isn't good). Hence, reasoning doesn't work (yet) for diffusion models. Though, alternative approaches are currently being tried out in academia; so someone might crack it!
When you say "they", what are you referring to? Image generation models? LLMs? All of them or older/more recent ones?
When you say "look the same", what do you mean?
I understand this was just a casual one-off comment, but as presented I'm having a hard time understanding what you're trying to convey and if you have appropriate domain knowledge: Your "looks the same" and my "looks the same" might be vastly different, especially depending on context.
Very disappointing compared to Flux 2 Klein. I was excited to see what the next frontier of small local image gen might be like, but this model is absolutely atrocious. I can't believe they bothered releasing it.
There is a story that Stability.AI went under with their trash release of Stable Diffusion 3, and that their better people went on to form and work at Black Forest Labs… apparently “3” is viral infected for these people.
People on reddit claim it's bad at generating images of people, apparently because they removed anything close to nsfw from the training set. Anyone already tried it?
I've been running it through tests on my local machine. The Text-to-Image component is VERY poor. It's not a good sign when the stock ComfyUI Flux3 Klein template with built-in config/prompts produces SD 1.5 levels of half-melted faces and various other body horror.
Flux 2 Klein 9b also had anatomical issues as well, but this one feels practically lobotomized.
https://imgpb.com/wDGGs
That looks worse than Flux 2.
What the heck are they doing?
Is this fast/cheap at the pareto frontier? Doesn't really work if humans look deformed.
I honestly don’t know what the hell BFL was thinking with this release. There’s idle conjecture that they may have been too heavy-handed when removing even the slightest risqué images from the training data, and as a result the final product has such a grossly poor understanding of basic human anatomy that no amount of LoRAs is going to be able to patch it up.
It’s bafflingly worse than other comparably sized ones: Qwen-Image 2.1 7B, Z-Image Turbo 6B which was released practically a year ago, even its own predecessor Flux 2 Klein 9b.
Even if we concede that the Klein variants are intended more as editor models, the results I've seen on Reddit aren't that impressive either.
I feel like they just handed their collective heads on a platter over to the team over at Krea especially since Krea 3 addresses the major limitation of Krea 2 which was strictly text-to-image only.
They don’t just filter out images in the dataset, but also have said they do targeted training and RL to neuter the model’s capability’s NSFW capability.
Safetymaxxing at the expense of benign output quality is sad to see.
>What the heck are they doing?
European edit model cannot make it easy to nudify people… they lobotomize it, it’s trash, Mission Accomplished!
Another reason that China is pulling ahead daily. They can leave in some anatomy and not knee cap their own product, leave in IP, and give away the model so as to undercut API/credit sales of US and EU models.
Finally… the Neo Puritan death march is having a direct financial toll.
Some examples of the Reddit posts/memes:
https://reddit.com/r/StableDiffusion/comments/1x79gth/why/
https://reddit.com/r/StableDiffusion/comments/1x7ktbu/enshit...
I gave Flux 3 Klein my typical human test that has never failed on newer image models and it did fail; in addition to lower quality, the fingers were definitely not the correct proportions.
Whether the cause is due to more safeguards against NSFW content is hard to confirm technically; it was more an intuition back in the Stable Diffusion XL days but I suspect modern training paradigms are better at correlative inference.
i think its instructive how the ancient sculptors and artists used to have to draw nudes in order to understand the human body. perhaps our image models need this
It isn't just the ancient sculptors and artists (unless 20 years ago is ancient :-P), when i went to an art school there were nude drawing lessons there too.
did you try with https://huggingface.co/black-forest-labs/flux-3-image-klein-... apparently the model is not going to work well unless this type of "prompt upsampler" is added
I did it in OpenRouter, which hits the direct BFL API.
Yes, a prime example of when the performance vs "bias" tradeoff is handled anti-business
I've tried it some, only locally. I rarely use APIs, so can't say what the official inference API results look like. It correctly followed some prompts that almost every model I've ever tried performed poorly at, but failed in strange ways on prompts that other models handle with ease.
At high resolutions, structural integrity on people can definitely fall apart like older models, where you get one person turning into a long spaghetti spanning across the frame. It might be that you need to do grounding by generating some kind of scene reference at its stable resolutions first then having it use that for the larger resolution like older models too.
They say that their prompt "upscaler" model or system prompt are essentially necessary for the model, but my results have been hit or miss with that claim. Some basic prompts are better, some upscaled prompts are better. I think that if you gain an intuition for how the model understands things, you might get better results by default by hand crafting prompts specifically for it.
Some of the stylized art I'm getting out of it at higher resolutions are really nice and I prefer them over other models.
All of these models seem to have their own character in terms of system requirements, performance, prompt adherence, artistic outputs and so on. None of the models I have tried are a complete homerun, so maybe in this case the model feels a little inconsistent in some ways, but it's a model I will probably keep around.
Lame. Back to Chroma it is then.
There's a base version for LoRA training so a porn adapter is likely already being built.
Comments on Krea 3 releasing yesterday 2026.10.15:
https://news.ycombinator.com/item?id=50105648
It's editing capabilities have been really solid, but I've been disappointed with its ability to do finer text details (a niche use case for my own purposes). There's also a lot of reports of body horror outputs, but I haven't encountered any myself.
Still though, it's a lot like the porsche mission S vs the tesla roadster situation with the weights release. Moss Image dropped which looks far better at the moment, leading to a "why Flux?" question in the community.
I still think the edit capabilites are absolutely excellent, but for t2i I'd try out moss: https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base/tree/ma...
I've been VERY impressed with Moss' design capabilites out of the box. Some samples:
https://image.non.io/9a32f732-668e-4fb3-bfef-13f0f40a0972.we...
https://image.non.io/8b5ee315-1ffd-4985-8ff0-d62e19d33f29.we...
https://image.non.io/5130501a-5c07-4635-82f5-fa64d60e8871.we...
IMO the taste is lacking, but it has some of the best text reproductions I've seen with their new VAE. Here's a type ramp example:
https://image.non.io/57246870-a245-4875-983f-993656e09e3d.we...
Looks like the people over /r/SD seem to like MOSS (i'm not up to date on the image stuff) but what i notice is that it is Apache2 licensed when every other openweight model seems to have some weird license (from a quick check Qwen-Image even downright forbids any commercial use at all).
Yeah, unfortunately, a lot of them have been moving to their own bespoke, non-commercial licenses (Ideogram, Krea, Flux, Qwen 2, etc).
Annoyingly, the original Qwen-Image was actually Apache 2 before they switched over unfortunately.
In addition to Moss, Boogu Image 0.1 is a unified model which came out in June and is kind of being slept on right now, is actually pretty decent and falls under the Apache 2.0 license. It scored surprisingly high on my GenAI benchmark (comparable to Krea 2 and Ideogram 4).
https://github.com/boogu-project/Boogu-Image
Looks like MOSS was made by some Chinese university as part of an AI education/research program so i guess they don't have to care about funding as much as the other labs :-P
I find this stuff really insane to think about, that it has "knowledge" about how most of the world looks like in less than 10GB.
The smaller we go with good models, the more I wonder about tiny animals.
By memorizing the fact that the sun rises in the East and sets in the West, you no longer have to make complex predictions and draw arcs every day. By memorizing the fact that the Earth rotates, and that generally most bodies in solar systems do, you can extrapolate this fact to any planet or solar system.
This example is a rough comparison to what actually happens, but it does attempt to demonstrate what compression and knowledge are, and how patterns, symmetries, and extrapolation/interpolation can make describing/modeling a complex system much more efficient.
Models can even go a step further, offloading entropy to the input. An efficient model exploits predictable structure; the remaining details can be expensive to compress further. The input can supply whatever additional information is necessary to produce the desired output within some margin of error. The information required depends on the model’s efficiency over a given domain, the structure of that domain, and the required precision of the output.
Consider generating an image of a pebble driveway. Most models today will generate a plausible arrangement of pebbles using their learned distributions. Reproducing a particular arrangement requires the input to convey the details that distinguish it from other plausible arrangements. The more closely you want to reproduce it, the more informative the input needs to be—potentially including a reference image of the arrangement itself.
This certainly raises a lot of open questions that are fun to pursue, such as the smallest possible model that can recover a complete description of our universe from minimal input.
> Pricing starts at $0.009 per 1K image.
This seems like a very competitive price for image models if it can compete on quality, no?
For example, Krea 3 released yesterday starting at $0.06/image.
That's true. What's also true is that it cannot compete performance-wise.
I’m can get 6000 trash images for the cost of one good one though!! (What are these companies thinking?)
backrooms energy
archived, no trackers: https://nonograph.com/flux-3-image-klein-10-16-2026-1
It kind of blows me away how “unreasonably effective” these small image models seem for their size.
Meanwhile a decent LLM that can do tool use and code and reason tends to weigh in at a minimum of 30B parameters, and you need over 100B to get good performance.
Do LLMs just contain more information or are diffusion models more efficient at compression?
I saw some work on language diffusion models a while back but I think they tend to have problems with reasoning versus transformer type models.
These models are, in a sense, distillations of language models. They wouldn't be as good without a pre-trained language model, so it's unfair to compare them.
Image models are essentially a mapping of text embedding space -> image. So the clearer and cleaner the text embedding space is the better model you get as you need less weights trying to interpret a text prompt, recent image models even use multimodal versions of LLMs (Flux uses Qwen3 VL) so all the visual separation was done beforehand making the model even more efficient.
edit: Maybe a better analogy is that the image models are only half the model in a encoder-decoder sense, where the LLM is a pretrained frozen encoder and the image model is the decoder.
To make a rather speculative argument - since an LLM should be able to describer an image losslessly, images are a subset of what an LLM needs to understand. It is possible that the statistical model being learned for images is just simpler.
The get a proper LLM they have to model all of human knowledge regarding patterns, physics, biology, social interaction, skills, history, etc, etc. For images they have to recall what a dog looks like in a still photo. The image models aren't even necessarily good at a wide range of novel situations; usually they get used for stock scenes.
Natural evolution has also had a much easier time coming up with vision models (everything with eyeballs) than complex language models (as far as we know only a smallish family of monkeys).
this is the right answer. ultimately text and general llm usage incl coding is far more internally complex than just memorizing the looks and composition of concepts placed in an image. dont forget that most of our reasoning benchmarks basically require the llm to have a search engine and calculator in its own weights.
These diffusion models are not just memorizing the looks and composition of concepts. These models are increasingly capable of representing advanced world knowledge. Things like shadows require a comprehensive, if indirect, understanding of physical light modeling. Things like weight/size estimation, etc., which are all important to a cohesive generated image of natural phenomenon, require an adequate understanding of physics.
Some of this understanding is in the input embeddings themselves, embedding models do a lot of heavy lifting behind image generation models.
I posit that you can theoretically achieve the same capability in a model trained on the right sequence of "images" as you can purely text; both are turned into embeddings and can ultimately convey the same exact relationships. I don't think it's very useful to try and create delineations between text and image inference, especially as more models move towards native multimodality.
> since an LLM should be able to describer [sic] an image losslessly
Any lossless text description of a given image would just be a structured text output of the exact pixel values, or an encoding that losslessly exploits regularities. Anything less is categorically lossy. An ordinary semantic description of a reasonably complex image will generally be lossy. A short, lossless text description would need to exploit regularities in the image, such as a solid-color image, or a basic pattern or sequence which is easily reproducible, such as a precisely specified checkerboard image.
And being able to represent images losslessly in text simply establishes text’s capacity to encode information. It doesn't establish an LLM’s holistic understanding of images. There is heavy lifting in the phrase “images are a subset.” An image can be almost any conceivable perceptual arrangement of data. The bounds are not clear. Maybe what you meant to suggest was that images of our natural world contain exploitable regularities.
> Natural evolution has also had a much easier time coming up with vision models (everything with eyeballs) than complex language models (as far as we know only a smallish family of monkeys).
It isn't appropriate to explain complex biological systems in terms of today's technologies. This is something people have been doing for a long time. The mind is like a a book, a loom, a camera, an engine, a computer, an LLM/vision model. Etc. The cognitive abilities required for an organism to meaningfully interpret its vision can possibly largely be the same cognitive abilities underlying speech, and all that is missing for a brain to exploit this might be sufficiently advanced vocal cords or manual dexterity; nothing to do with the brain itself, but its auxiliary organs.
At first, I considered that diffusion models struggle with reasoning because they underweight 'prior beliefs', ie the propositions by which an argument advances.
Then I considered whether the universe of imagery is simply smaller than the universe of ideas—which must, itself, include the universe of images, too.
But finally, I'm thinking maybe images are just harder to benchmark.
Or maybe they're _not_ actually worth 1,000 words.
When it comes to reasoning specifically, it's a far simpler explanation IMO. Modern RLVR techniques like GRPO rely on the model understanding the likelihood of the generated phrase. This is really easy to compute for LLMs that are trained with next-word prediction. Diffusion models have a harder time accessing this likelihood value (i.e., you can approximate it but the approximation isn't good). Hence, reasoning doesn't work (yet) for diffusion models. Though, alternative approaches are currently being tried out in academia; so someone might crack it!
They can make good images but they all look the same. Guess it's not hard to replicate a small section of what exists.
When you say "they", what are you referring to? Image generation models? LLMs? All of them or older/more recent ones?
When you say "look the same", what do you mean?
I understand this was just a casual one-off comment, but as presented I'm having a hard time understanding what you're trying to convey and if you have appropriate domain knowledge: Your "looks the same" and my "looks the same" might be vastly different, especially depending on context.
Very disappointing compared to Flux 2 Klein. I was excited to see what the next frontier of small local image gen might be like, but this model is absolutely atrocious. I can't believe they bothered releasing it.
There is a story that Stability.AI went under with their trash release of Stable Diffusion 3, and that their better people went on to form and work at Black Forest Labs… apparently “3” is viral infected for these people.