• It's a heck of a lot smaller than Qwen-Image 1 (20b parameters) at only 7b, making it one of the smaller open-weight models available (Z-Image Turbo is one of the few that is smaller at 6b) when compared to Ideogram, Krea2, Flux2, etc.
• It supports native transparency (Qwen's team, as far as I know, is the only one attempting to tackle this). Even though it's relatively trivial to set up background removal postprocessors, it's also neat to see it natively supported.
• It's fast using QwenImage2.1 convrot, a 1MP image took around ~5 seconds on an RTX4090.
Negatives
• The license (assuming you respect it) is far more restrictive. The original Qwen Image 1 was released under the standard Apache license; this one explicitly forbids commercial usage without obtaining a separate license. On the other hand, a lot of us didn't expect the Qwen team to ever release "weights-available" ever again.
Qwen-Image 1.0, released about a year ago, only scored 4/15 on my GenAI Showdown Benchmarks. Since that time, they've been upstaged by Krea 2 (6/15) and Ideogram4 (8/15). I'll post the new results once I have some more time to run them.
That's because the users don't care about the licence. They did try to take down a porn Lora repo on hugging face. The trick is, don't say you're using minimax, and you can make porn with it.
Can do! GPT-Image-2 already scored unsurprisingly very high: 12 out of 15 on text-to-image, and 10 out of 12 on image-to-image.
The three benchmarks it failed on (D20, Flat Earth, and Banded Snake) are pretty difficult, so I'd be surprised if 2.5 manages to pass them, but I’ll add it for completeness’ sake later this week.
That benchmark might have some issues. You prompted the models to generate an image of striking a ring against a crucible. Then you (presumably, manually?) scored the images that depicted an anvil higher than the ones striking something resembling a crucible.
That’s a good catch. Yes, all scoring is done through manual review since relying on a VL model for these kinds of meta-metrics is a sort of loose equivalent of gödel's second incompleteness theorem.
I’ll have to think about this one. When I crafted the prompt, I wasn’t really thinking about the differences between a crucible and an anvil. It was more the visual of an archangel smelting halos for newly arrived heavenly beings.
I'm not sure why one would even strike metal against a crucible! It's a container for liquid metal. One of the outputs shows it being smashed by the manoeuvre, which is probably the most realistic outcome of all of them.
Sorry, I'm not trying to nitpick. I'm just joining in because I'm interested in how the models dealt with the request.
Well this is HN - original home of the "ummm actually..." - so I appreciate when people pick all the nits. :)
Even though I prompted for a crucible in the prompt, I think the fact that the prompt also contained terms like “blacksmith” and “hammer,” caused it to lean towards anvils over crucibles in some of the pictures (which as you brought up makes more sense anyway).
I promise I searched on my phone before posting that, but it looks like I misinterpreted an hf search result that was actually Qwen-Image-2512 but the url was truncated to Qwen-Image-2 due to my screen width (ugh). For some reason I was thinking Qwen Image 2.1 was the one that had been available for a while but not released, but you're right that was 2.0.
I run a prompt-to-ui design site that uses image models for the design process[1]. The text rendering especially makes this model deeply interesting to me, despite the license. Here are some tests using my harness comparing the outputs of gpt-image-2 and qwen 2.1:
The text rendering definitely is much, much better than anything else on the open weights market right now. Small text fidelity is quite good. It seems like the text encoder however gets a little bit overloaded with larger prompts - note the presence of hex codes in the design output, those were inputs from the expanded prompt.
I'll be trying a post-training run on this for web design, it has some serious potential.
> The text rendering definitely is much, much better than anything else on the open weights market right now
Really? Because basically everything in those screenshots is completely garbled. I didn't follow it super closely but I thought Ideogram or whatever was really good for this particular use, with actual clear text.
This is my experience as well. Ideogram4 (assuming you are willing to put in the work to use the proper structured JSON input) is very accurate when it comes to text rendering in an image.
Totally normal for modern models due to training on the same datasets supplied by third parties, dataset contamination, and mode collapse, especially for simple prompts that don't have enough semantic capacity. -isms are often very similar even without distillation, and tend to come and go in waves along with model generations.
My harness expands the prompt into a json representation that specifies layout much more rigorously, which is why you see such that amount of alignment between the two.
That internal json backing helps significantly when you want to maintain consistent design system components/patterns across multiple pages. The aligned layout is it working as intended.
I love the non-commercial clauses because of how many people are using these for deceptive ads and “virtual staging” and fake social media accounts. Anything that makes those guys lives harder while still letting me make silly pictures for my kids and tapestries for my D&D campaign feel fine by me.
Why would you even do that? Just... use it? There hasn't been any legal precedent on if models can even be copyright restricted. Labs just keep publishing license documents as if they matter.
Well, it is an indication that it matters to the lab, so if you don't want legal fees to be the first one to set precedent, then it does matter a great deal.
You could say the same thing about "license-washing" the model. It seems like you're just going through a guaranteed expensive process to have roughly the same risk as just using the model and potentially getting hit with legal fees.
It's not going to matter unless you plan to commercially deploy the model, as far as I see.
If you were to generate outputs for commercial use, I think it would still violate this research license, but it's not like they are going to know, are they?
That said, I am disappointed that the model is not actually open-weights as I expected based on the headline.
Usually the way the licenses are interpreted, if I run it myself and use it to generate $X million in revenue in my own business, I'm subject to the license. In this case, the license seems to cover any commercial usage at all, which would certainly include self-hosting for internal business purposes.
Why even bother demanding such terms? Release the weights or don't release the weights, but this is silly. No one is going to pay them to run or host a 7B static image model.
You are not wrong, but will a judge and jury be competent enough to understand the difference after you've spent several hundred thousand in litigation?
Boogu-Image has the Apache 2.0 License [1] (good coherence, but outputs can look synthetic).
And Krea 2 has a community license [2] that is fairly permissive - I think commercial usage is allowed under $1 million.
Boogu-Image scored 6/15 and Krea 2 scored 7/15 on my GenAI Showdown benchmark [3] - only Ideogram4 eclipses them in terms of local models, but its got a far more restrictive license and the JSON structured inputs can be a pain to work with.
Calling open-weights as open-source in marketing materials is the usual misrepresentation. But now with the restriction on commercial use (which is against opensource definition) it is not even open-weights, technically it would be more accurate to call it weights-available.
The capabilities of local LLM text-to-image is honestly pretty damn impressive. IMO, I think local image generation is currently ahead of local code generation. I can get an image in seconds locally with the quality being way higher than what I'd expect from a local model. However with coding it's much slower and much less impressive. I'm sure there's a reason for this and I'm not an AI expert so I'll let the smarter folks tell me why, but that's just been my observation thus far.
Text is in a sense way harder to do than images because of radical nonlocality. A word at the start of one paragraph can directly influence the meaning of a word five paragraphs away. Whereas images typically represent the real world, or at least a spatial domain, which gives you a lot of structure 'for free'. If you are drawing a human, you can make a reasonable guess where their hands go in relation to their face. If someone hands you the first half of an essay, finishing it is not trivial.
I've played with diffusion models on and off since the first release of Stable Diffusion - just for amusement, without a particular goal.
Recently, I've been helping a friend's wife with some basic vector images for her sewing hobby (she has what is essentially a CNC sewing machine) and have been super-impressed with FLUX.1-Kontext, which I've been running on my Macbook Pro with mflux. Its ability to (for example) take a photo of a human or an animal and return a line drawing which is recognisably them (rather than just a generic similarish image as I've experienced with other models) is excellent.
It's an older model now, but (AIUI) has the text-handling features baked in, and in my various testing is very reliable at giving me the outputs that I want, without the randomness I've experienced previously. It's big and relatively slow (~3 mins per 512x512 image edit on my M1 Max Mac) but excellent to work with. It's also very straightforward to set up, without the harness complexity of e.g. comfyui.
Off the shelf - it’s a Brother. It prints via a proprietary file format (.PES) but there’s an extension for Inkscape that supports creation and export.
Remember that quality output is a necessary but insufficient property of a generative model.
Prompt-adherence is really hit-or-miss—especially if one lacks the visual vocabulary. Likewise with coding, I find junior devs don't think to prompt re: respecting this-or-that interface, or refactoring to point-free style, etc.
Probably ComfyUI is one of the easiest way to get started with local image/video models. Or perhaps vLLM, if they have support for it already, would be something like `vllm serve <model> --omni --port 9080`
Seems I'm missing something. Does this model support other inputs?
Image outputs are supported, videos I'm not sure but I don't think that's an output, just a preview of the equirectangular example, so, same question here, what does this model outputs that isn't supported?
Not all architectures are supported by llama.cpp . The GGUF format encodes the NN in a standardized way, but then you need code that can use that NN structure.
I understand that llama.cpp could only output text, last time I checked (I do not know how to find a good source for that though).
I haven't used llama.cpp for image generation either but I recall an issue about it. Unfortunately I can't pinpoint it now and there is the older closed issue https://github.com/ggml-org/llama.cpp/issues/4408 so unless mtmd supports also multimodal outputs out of the box safe to assume output is still limited to text.
> Diffusers, ComfyUI, vLLM-Omni, SGLang, and LightX2V
I think that's all Python (not a direct executable).
You could just do (see the "Quick Start") four `pip install` and have a dozen lines script to generate the image. But `llama.cpp` and similar do not require e.g. installing Torch (or PyTorch) - you can use `llama.cpp` on a non-specialized machine.
There is difussion.cpp which is intended for those types of models. I set up krea-2-turbo with the help of ChatGPT 2 months ago, if you have a capable computer that's what I would suggest once it becomes supported.
Additional question is what kind of local hardware would be required for this? 7B parameters sounds very light weight, but I'm not sure. (Edit: The download is 33 GB).
Edit x2: As usual I'm in a twisty maze of pip packages that don't work together, with obscure errors about missing modules, even though I followed the instructions on the page to the letter. I really wish people didn't use Python for this stuff. A simple C/C++ program would be so much better.
It's about 16 GiB at Q8 quants (combining both the image and language parts). (Meaning, community quantized models from HuggingFace).
I think it will technically run on anything that has enough memory. I just tried it on a standard laptop (dual-channel DDR5), and it took about 3 minutes for a 512x512. If you'd want to run it at interactive speeds, you would want a GPU (one which fits this in VRAM).
> "I really wish people didn't use Python for this stuff. A simple C/C++ program would be so much better."
I've set it up on my local machine just now, as my first local image diffuser. I can confirm it's very easy.
I tried stable-diffusion.cpp, following its compile guide here[0], and its Qwen Image-2.1 specific instructions here[1]. It works out of the box. I made a test pelican[2]. It took 3 minutes on a CPU.
I use opencode + <a decent saas llm> to set up all this new ai generation stuff. GLM-5.3 is my current gun. Safely inside podman containers too because I dont trust this fast moving python eco system at all. Never do I want this running on my main OS.
I have FLUX.2 klein and dev, Ideogram, LaDA-Image and SenseNova locally. Works great. Ive never touched a file.
The days of making container yamls myself is over. I read them but I dont edit anymore.
latents go from 16ch @ 8x compression to 64ch @ 16x, so roughly the same total latent budget but much more channel heavy. It’s also deeper/wider, and the old 2x2 transformer patching is gone.
On some images it still produces artifacts but can't say if it's the transformer or the VAE yet.
Just think about how recently we got that feature in the official ChatGPT image gen. And now we have that running locally — assuming that is, I can figure out how to get this running on my Mac — blows my mind.
I want to say that they released a version since then before 2.5, but I'm not entirely sure. I should also note that I do not know a single Chinese character, so it's possible they are broken in such a way that I wouldn't necessarily notice. I do remember, however, that English text used to be broken; now it's completely readable.
Its happy to see a new open image model from qwen. But the license is a let down. And it dosent even beat their closed qwen3 image wich is already a bit old.
My first impression is that it's not so good at following prompt directions. I asked it to place a 3D text made of glass in a particular city. It instead gave me a broken 3D text on a white background. Maybe with different seeds it gets better, but it's more of a trial and error process than reliable results.
I am really grateful to the Chinese Labs for open sourcing their best models. If it was left to the Americans, we would be forced to pay obscene API fees to use them.
> You shall not use the Materials for any commercial purpose without obtaining a separate commercial license from us.
It probably will be much cheaper to use than other image models, but it seems that will be up to the whims of Qwen/Alibaba rather than just being the cost of putting it in a cloud provider.
Qwen and Alibaba are the biggest competitor for basically every model out there. They're beating the benchmarks like top-frontier models, focused on open-source and much cheaper than the competitors.
Not sure there's a better avatar for the absurdity of AI slop imagery than the "cowboy on horseback". That's a pony with a child's saddle on it, and they've composited a grown man on top of it.
So thoughts
Positives
• It's a heck of a lot smaller than Qwen-Image 1 (20b parameters) at only 7b, making it one of the smaller open-weight models available (Z-Image Turbo is one of the few that is smaller at 6b) when compared to Ideogram, Krea2, Flux2, etc.
• It supports native transparency (Qwen's team, as far as I know, is the only one attempting to tackle this). Even though it's relatively trivial to set up background removal postprocessors, it's also neat to see it natively supported.
• It's fast using QwenImage2.1 convrot, a 1MP image took around ~5 seconds on an RTX4090.
Negatives
• The license (assuming you respect it) is far more restrictive. The original Qwen Image 1 was released under the standard Apache license; this one explicitly forbids commercial usage without obtaining a separate license. On the other hand, a lot of us didn't expect the Qwen team to ever release "weights-available" ever again.
Qwen-Image 1.0, released about a year ago, only scored 4/15 on my GenAI Showdown Benchmarks. Since that time, they've been upstaged by Krea 2 (6/15) and Ideogram4 (8/15). I'll post the new results once I have some more time to run them.
https://genai-showdown.specr.net
They're trying to cash in but this is just sad
Within a few days this seems a total pivot from Xiaomi’s op RL dashboard and the praise of Chinese open model? What is the sentiment now?
Qwen can't train anymore with openai reasoning tokens? I kid I kid.
Everything is combined and uneven, including the opinions of hackernews commenters?
There is no single opinion, and clearly no single Chinese approach.
Also Chinese labs are in particular very careful about anything which be used to create pornographic content, which is highly illegal in the PRC.
> Also Chinese labs are in particular very careful about anything which be used to create pornographic content, which is highly illegal in the PRC.
MiniMax H3 is dominating AI porn right now, and they are Chinese.
That's because the users don't care about the licence. They did try to take down a porn Lora repo on hugging face. The trick is, don't say you're using minimax, and you can make porn with it.
Could you add new OAI 2.5 image models?
Can do! GPT-Image-2 already scored unsurprisingly very high: 12 out of 15 on text-to-image, and 10 out of 12 on image-to-image.
The three benchmarks it failed on (D20, Flat Earth, and Banded Snake) are pretty difficult, so I'd be surprised if 2.5 manages to pass them, but I’ll add it for completeness’ sake later this week.
That benchmark might have some issues. You prompted the models to generate an image of striking a ring against a crucible. Then you (presumably, manually?) scored the images that depicted an anvil higher than the ones striking something resembling a crucible.
That’s a good catch. Yes, all scoring is done through manual review since relying on a VL model for these kinds of meta-metrics is a sort of loose equivalent of gödel's second incompleteness theorem.
I’ll have to think about this one. When I crafted the prompt, I wasn’t really thinking about the differences between a crucible and an anvil. It was more the visual of an archangel smelting halos for newly arrived heavenly beings.
I'm not sure why one would even strike metal against a crucible! It's a container for liquid metal. One of the outputs shows it being smashed by the manoeuvre, which is probably the most realistic outcome of all of them.
Sorry, I'm not trying to nitpick. I'm just joining in because I'm interested in how the models dealt with the request.
Well this is HN - original home of the "ummm actually..." - so I appreciate when people pick all the nits. :)
Even though I prompted for a crucible in the prompt, I think the fact that the prompt also contained terms like “blacksmith” and “hammer,” caused it to lean towards anvils over crucibles in some of the pictures (which as you brought up makes more sense anyway).
It would make more sense to compare with Qwen Image 2, since that was the last open weights Qwen model.
Edit: This is wrong.
Wait... is that true? I don't think the original Queen Image 2.0 was ever released beyond an API. At least, I don't remember a public weights release.
I promise I searched on my phone before posting that, but it looks like I misinterpreted an hf search result that was actually Qwen-Image-2512 but the url was truncated to Qwen-Image-2 due to my screen width (ugh). For some reason I was thinking Qwen Image 2.1 was the one that had been available for a while but not released, but you're right that was 2.0.
I run a prompt-to-ui design site that uses image models for the design process[1]. The text rendering especially makes this model deeply interesting to me, despite the license. Here are some tests using my harness comparing the outputs of gpt-image-2 and qwen 2.1:
https://html.non.io/qwen-comparison/
The text rendering definitely is much, much better than anything else on the open weights market right now. Small text fidelity is quite good. It seems like the text encoder however gets a little bit overloaded with larger prompts - note the presence of hex codes in the design output, those were inputs from the expanded prompt.
I'll be trying a post-training run on this for web design, it has some serious potential.
[1] diffui.ai
> The text rendering definitely is much, much better than anything else on the open weights market right now
Really? Because basically everything in those screenshots is completely garbled. I didn't follow it super closely but I thought Ideogram or whatever was really good for this particular use, with actual clear text.
This is my experience as well. Ideogram4 (assuming you are willing to put in the work to use the proper structured JSON input) is very accurate when it comes to text rendering in an image.
Those simple prompts produce nearly the exact same layout in the 2 different models?
Equally confused with this. They must be using a lot more guidance than just the provided prompt.
Qwen is trained off of gpt’s outputs. This is both a positive and negative
Totally normal for modern models due to training on the same datasets supplied by third parties, dataset contamination, and mode collapse, especially for simple prompts that don't have enough semantic capacity. -isms are often very similar even without distillation, and tend to come and go in waves along with model generations.
Qwen's latest image models have a ton of distillation from gpt-image, same with Grok Imagine.
Even the artifacts are getting picked up.
My harness expands the prompt into a json representation that specifies layout much more rigorously, which is why you see such that amount of alignment between the two.
That internal json backing helps significantly when you want to maintain consistent design system components/patterns across multiple pages. The aligned layout is it working as intended.
A lot of the previous Qwen models seem to have used Apache licenses, among others:
https://en.wikipedia.org/wiki/Qwen#List_of_models
Unfortunately, it looks like this model is using a much more restrictive license:
https://github.com/QwenLM/Qwen-Image-2.1/blob/main/LICENSE
Was going to post about this: the last image models with Apache 2.0 license seem to be from 2025, recent Qwen models are "non-commercial use".
I love the non-commercial clauses because of how many people are using these for deceptive ads and “virtual staging” and fake social media accounts. Anything that makes those guys lives harder while still letting me make silly pictures for my kids and tapestries for my D&D campaign feel fine by me.
You think they care about the probably unenforceable license terms?
This achieves absolutely nothing to that end.
People can continue to use closed SOTA models to generate outputs for commercial or malicious purposes.
What this research license achieves is that we cannot use this model in applications we publish.
Companies can use llm to license-wash open source code regardless of license.
How difficult would it be to use this model to create a second model without licensing issues?
Why would you even do that? Just... use it? There hasn't been any legal precedent on if models can even be copyright restricted. Labs just keep publishing license documents as if they matter.
Well, it is an indication that it matters to the lab, so if you don't want legal fees to be the first one to set precedent, then it does matter a great deal.
You could say the same thing about "license-washing" the model. It seems like you're just going through a guaranteed expensive process to have roughly the same risk as just using the model and potentially getting hit with legal fees.
What.
I'm willing to bet a nonzero amount of its training material is GPL, so I'll treat it as GPL licensed instead and use it however the fuck I want.
If AI labs get to ignore licenses, so do we.
Agreed.
It's not going to matter unless you plan to commercially deploy the model, as far as I see.
If you were to generate outputs for commercial use, I think it would still violate this research license, but it's not like they are going to know, are they?
That said, I am disappointed that the model is not actually open-weights as I expected based on the headline.
That seems quite important. Definitely requires any grown up company to have to do a bunch of legal paperwork to use it.
It could have an attempt at steganographic watermarking trained into the model.
It's not going to matter unless you plan to commercially deploy the model, as far as I see.
It's not going to matter then, either. What are they going to do, sue me for copyright infringement?
Yes? If you deploy the model and provide it on Openrouter (or elsewhere) you better have a license.
No need to deploy it publicly. It's 7B, it'll run on a 3090.
You talked about commercially deploying it. Of course nobody cares what you are doing on your own hardware for fun.
Usually the way the licenses are interpreted, if I run it myself and use it to generate $X million in revenue in my own business, I'm subject to the license. In this case, the license seems to cover any commercial usage at all, which would certainly include self-hosting for internal business purposes.
Why even bother demanding such terms? Release the weights or don't release the weights, but this is silly. No one is going to pay them to run or host a 7B static image model.
Why not? If it's cheap enough, and presuming I had any reason to use an image model, I'd consider paying for it on OpenRouter over hosting it myself.
You are not wrong, but will a judge and jury be competent enough to understand the difference after you've spent several hundred thousand in litigation?
You really think Alibaba is going to go around and sue in US courts for something like this?
It’s more of something to scare companies with legal teams. If you’re an individual or hobbyist doing a side project the risk is essentially zero.
What are the top image models that still use a less restrictive license today?
Boogu-Image has the Apache 2.0 License [1] (good coherence, but outputs can look synthetic).
And Krea 2 has a community license [2] that is fairly permissive - I think commercial usage is allowed under $1 million.
Boogu-Image scored 6/15 and Krea 2 scored 7/15 on my GenAI Showdown benchmark [3] - only Ideogram4 eclipses them in terms of local models, but its got a far more restrictive license and the JSON structured inputs can be a pain to work with.
[1] - https://github.com/Boogu-Project/Boogu-Image
[2] - https://www.krea.ai/krea-2-licensing
[3] - https://genai-showdown.specr.net/?models=fd,hd,kd,qi,f2d,zt,...
Calling open-weights as open-source in marketing materials is the usual misrepresentation. But now with the restriction on commercial use (which is against opensource definition) it is not even open-weights, technically it would be more accurate to call it weights-available.
And the only reason source available had any significance is that you could look at something and understand it. Weights are much much more opaque.
It's just freeware.
Tune the weights a bit and call them derivative work.
The capabilities of local LLM text-to-image is honestly pretty damn impressive. IMO, I think local image generation is currently ahead of local code generation. I can get an image in seconds locally with the quality being way higher than what I'd expect from a local model. However with coding it's much slower and much less impressive. I'm sure there's a reason for this and I'm not an AI expert so I'll let the smarter folks tell me why, but that's just been my observation thus far.
I mean I’m sure it’s the reverse for an artist. They would be less impressed with the image and more impressed with the code quality
To generalize, LLMs are great at what you are not skilled at.
That's how they're sold, isn't it?
Well, not exactly. Many, many companies and managers think that they can replace skilled people or have unskilled monitoring them.
That's a fair statement, I agree. I'm quite an abysmal artist so I could be a victim of my own bias here
The point I think is interesting is that this is just 7B. The current SOTA 7B LLMs are barely usable for quite simple coding.
Text is in a sense way harder to do than images because of radical nonlocality. A word at the start of one paragraph can directly influence the meaning of a word five paragraphs away. Whereas images typically represent the real world, or at least a spatial domain, which gives you a lot of structure 'for free'. If you are drawing a human, you can make a reasonable guess where their hands go in relation to their face. If someone hands you the first half of an essay, finishing it is not trivial.
I've played with diffusion models on and off since the first release of Stable Diffusion - just for amusement, without a particular goal.
Recently, I've been helping a friend's wife with some basic vector images for her sewing hobby (she has what is essentially a CNC sewing machine) and have been super-impressed with FLUX.1-Kontext, which I've been running on my Macbook Pro with mflux. Its ability to (for example) take a photo of a human or an animal and return a line drawing which is recognisably them (rather than just a generic similarish image as I've experienced with other models) is excellent.
It's an older model now, but (AIUI) has the text-handling features baked in, and in my various testing is very reliable at giving me the outputs that I want, without the randomness I've experienced previously. It's big and relatively slow (~3 mins per 512x512 image edit on my M1 Max Mac) but excellent to work with. It's also very straightforward to set up, without the harness complexity of e.g. comfyui.
is the cnc sewing machine an off the shelf model or something DIY? I'd love to hear more
Off the shelf - it’s a Brother. It prints via a proprietary file format (.PES) but there’s an extension for Inkscape that supports creation and export.
How are you converting the bitmaps into vector images?
Remember that quality output is a necessary but insufficient property of a generative model.
Prompt-adherence is really hit-or-miss—especially if one lacks the visual vocabulary. Likewise with coding, I find junior devs don't think to prompt re: respecting this-or-that interface, or refactoring to point-free style, etc.
So, as others have said, the artist knows better.
How do you use this model locally, similarly to using `llama-server -m <model>`?
(I mean: outside direct or substantial use of Python, and running the Neural Network in the most efficient way.)
Probably ComfyUI is one of the easiest way to get started with local image/video models. Or perhaps vLLM, if they have support for it already, would be something like `vllm serve <model> --omni --port 9080`
why not just as you suggested i.e. https://qwen.readthedocs.io/en/latest/run_locally/llama.cpp.... then get the result either via a UI or wget/curl it back?
I am not sure that llama.cpp also supports image generation models.
it's multimodal, see https://github.com/ggml-org/llama.cpp/blob/master/docs/multi...
Multimodal doesn't guarantee input and output.
> Currently, we support image, audio and video input.
Seems I'm missing something. Does this model support other inputs?
Image outputs are supported, videos I'm not sure but I don't think that's an output, just a preview of the equirectangular example, so, same question here, what does this model outputs that isn't supported?
It's a diffusion model, completely different from autoregressive attention models.
like https://github.com/ggml-org/llama.cpp/tree/master/examples/d... ?
At the risk of stating the obvious llama.cpp isn't just about LLaMa as https://github.com/ggml-org/llama.cpp/blob/master/src/llama-... someone else pointed out.
Aha I was wrong. Thanks for sharing that!
no worries, I was wrong too, it is multimodal but only for inputs apparently, so for now there seem to only be text as output but no image as output
Not all architectures are supported by llama.cpp . The GGUF format encodes the NN in a standardized way, but then you need code that can use that NN structure.
I understand that llama.cpp could only output text, last time I checked (I do not know how to find a good source for that though).
See https://github.com/ggml-org/llama.cpp/blob/master/src/llama-... , the
...
I haven't used llama.cpp for image generation either but I recall an issue about it. Unfortunately I can't pinpoint it now and there is the older closed issue https://github.com/ggml-org/llama.cpp/issues/4408 so unless mtmd supports also multimodal outputs out of the box safe to assume output is still limited to text.
on the linked GitHub page they list support Diffusers, ComfyUI, vLLM-Omni, SGLang, and LightX2V with links to each
> Diffusers, ComfyUI, vLLM-Omni, SGLang, and LightX2V
I think that's all Python (not a direct executable).
You could just do (see the "Quick Start") four `pip install` and have a dozen lines script to generate the image. But `llama.cpp` and similar do not require e.g. installing Torch (or PyTorch) - you can use `llama.cpp` on a non-specialized machine.
"just"
I don't think I have ever once run "pip install transformers" and had it work without three rounds of fiddling
There is difussion.cpp which is intended for those types of models. I set up krea-2-turbo with the help of ChatGPT 2 months ago, if you have a capable computer that's what I would suggest once it becomes supported.
Additional question is what kind of local hardware would be required for this? 7B parameters sounds very light weight, but I'm not sure. (Edit: The download is 33 GB).
Edit x2: As usual I'm in a twisty maze of pip packages that don't work together, with obscure errors about missing modules, even though I followed the instructions on the page to the letter. I really wish people didn't use Python for this stuff. A simple C/C++ program would be so much better.
It's about 16 GiB at Q8 quants (combining both the image and language parts). (Meaning, community quantized models from HuggingFace).
I think it will technically run on anything that has enough memory. I just tried it on a standard laptop (dual-channel DDR5), and it took about 3 minutes for a 512x512. If you'd want to run it at interactive speeds, you would want a GPU (one which fits this in VRAM).
> "I really wish people didn't use Python for this stuff. A simple C/C++ program would be so much better."
You mean besides stable-diffusion.cpp ?
> You mean besides stable-diffusion.cpp ?
Yes, thanks, I didn't know about that. Will try it.
I am using sd.cpp, which is the cousin of llama.cpp: https://github.com/leejet/stable-diffusion.cpp
it already has day-0 qwen image 2.1 support!
I've set it up on my local machine just now, as my first local image diffuser. I can confirm it's very easy.
I tried stable-diffusion.cpp, following its compile guide here[0], and its Qwen Image-2.1 specific instructions here[1]. It works out of the box. I made a test pelican[2]. It took 3 minutes on a CPU.
[0] https://github.com/leejet/stable-diffusion.cpp/blob/master/d...
[1] https://github.com/leejet/stable-diffusion.cpp/blob/master/d...
[2] https://i.ibb.co/yMknC2K/output.png
Unsloth Desktop is the easiest way imo. There are already gguf quants of this model, or simply wait until the official one comes out.
I use opencode + <a decent saas llm> to set up all this new ai generation stuff. GLM-5.3 is my current gun. Safely inside podman containers too because I dont trust this fast moving python eco system at all. Never do I want this running on my main OS.
I have FLUX.2 klein and dev, Ideogram, LaDA-Image and SenseNova locally. Works great. Ive never touched a file.
The days of making container yamls myself is over. I read them but I dont edit anymore.
They finally fixed their VAE. It really held back their models over the last 2 years.
EDIT: It still produces artifacts it's better but unusable for production work. In midvalues you will see a slight dot pattern.
> finally fixed their VAE
Can you share the sources?
It's right there in the hugging face link?
latents go from 16ch @ 8x compression to 64ch @ 16x, so roughly the same total latent budget but much more channel heavy. It’s also deeper/wider, and the old 2x2 transformer patching is gone.
On some images it still produces artifacts but can't say if it's the transformer or the VAE yet.
> In midvalues you will see a slight dot pattern.
Is this not simply some sort of watermark instead of an artifact?
No, it's probably their rope implementation. They had a similar problem with the old qwen image but to be sure it needs some digging.
A 7B diffusion model can now render CJK text better than Microsoft Windows.
Just think about how recently we got that feature in the official ChatGPT image gen. And now we have that running locally — assuming that is, I can figure out how to get this running on my Mac — blows my mind.
Pretty sure comfyui has a mac executable
Is ChatGPT really that good?
Back in Apr, ChatGPT Images 2.0 has some broken Chinese texts in its featured examples, and they later removed that from blog post. Is 2.5 better now?
I want to say that they released a version since then before 2.5, but I'm not entirely sure. I should also note that I do not know a single Chinese character, so it's possible they are broken in such a way that I wouldn't necessarily notice. I do remember, however, that English text used to be broken; now it's completely readable.
Ideogram 4 has been around for a while haha
Its happy to see a new open image model from qwen. But the license is a let down. And it dosent even beat their closed qwen3 image wich is already a bit old.
Qwen Image 3 was released two months ago: https://qwen.ai/blog?id=qwen-image-3.0 I think you have it confused with another model.
My first impression is that it's not so good at following prompt directions. I asked it to place a 3D text made of glass in a particular city. It instead gave me a broken 3D text on a white background. Maybe with different seeds it gets better, but it's more of a trial and error process than reliable results.
Try translating your prompt to Chinese first, it seems a lot better at understanding and following Chinese prompts even with the translation hop.
God I love the Qwen team. Easily the most diverse set of models from all the Chinese labs. Only Gemini/DeepMind comes close.
I am really grateful to the Chinese Labs for open sourcing their best models. If it was left to the Americans, we would be forced to pay obscene API fees to use them.
Note that the license on this has this in it:
> You shall not use the Materials for any commercial purpose without obtaining a separate commercial license from us.
It probably will be much cheaper to use than other image models, but it seems that will be up to the whims of Qwen/Alibaba rather than just being the cost of putting it in a cloud provider.
https://github.com/QwenLM/Qwen-Image-2.1/blob/main/LICENSE
Good luck to them enforcing that license.
Interesting in the example of assembling the Cheers team how the otherwise great result genericizes Shelley Long.
The result seems a pretty good representation given the source image wasn't that great. I think that Woody Harrelson comes across much worse.
Qwen and Alibaba are the biggest competitor for basically every model out there. They're beating the benchmarks like top-frontier models, focused on open-source and much cheaper than the competitors.
Excited to see what the future holds for them!
Very impressive, and kind of worrying a 7B model can have such capabilities. The implications are huge. And Qwen does no watermarking (yet) yeah?
They always had a fourier space mark in their models even without the VAEs are usually pretty easy to detect.
Ah I wasn't aware, thank you
I don't want only cherry picked examples. Show me failure modes too.
failure modes are so load bearing.
Boy do I love waking up to find a new awesome toy from the Qwen team waiting for me to play with! Pulling it now
I would call this a license trap: Qwen RESEARCH LICENSE AGREEMENT
Code on github, models on huggingface, nice intro text: "We are excited to open-source Qwen-Image-2.1 [...]".
meh...
Not sure there's a better avatar for the absurdity of AI slop imagery than the "cowboy on horseback". That's a pony with a child's saddle on it, and they've composited a grown man on top of it.
While the license of this model is a shame it is still unenforceable.
I know a few friends of mine who are running models and are ignoring the licence.
Whether it is AGPL 3.0, or a completely restrictive license, it is going to get broken anyway and be used for commercial purposes.
I don't know anyone who looks at the licenses of the OSS software they are using.
In today’s world OSS is synonymous with "Free" and the AI model providers are proof of that with their training of code, datasets, etc.
So it begs the question, why should we abide by their licenses of their models?
And they shouldn’t be enforceable considering how the training data was slurped up without concern for licensing.
Woody Harrelson?
Still fails to generate smoothly animated sprites, although the native RGBA transparency is nice. Anyone found one that can?
Is it just me or are Alibaba / qwen’s websites often appear broken / very slow?
While I'm impressed with the Bluey example, the lack of Muffin disappoints me.
Image gen you eyeball one frame and stop, code needs hundreds of tokens all correct in sequence, one bad line and the whole thing fails.