▲ 192 points
back
55 comments
If I had a nickel for every model card, repo, or spec posted here that had zero samples of the actual output, I could probably start financing a rack of MI455Xs.
How is the pro version different from this?
OpenAI does surprisingly well in this regard with their blog posts, and I'm more of an Anthropic fanboy. I wish most tech releases for AI were as well done as OpenAIs they actually showcase what they've built, honestly, if I had to compare it to another company, I wouldn't be surprised if former Apple employees are doing OpenAI's press briefings. Whoever does press briefings at Apple, they are worth their weight in gold.
The last chief communications officer of OpenAI used to work at Apple for 7 years according to her linkedin.
As I was writing this it dawned on me, and I had a feeling someone would confirm. :)
I am not a Sam Altman fan, but I do give him credit where its due, their press briefings are chef's kiss.
I prefer no samples over cherry picked samples, though.
Really? You prefer not to see the top end of the distribution? Why?
Because I don't like baby sitting an AI. Just tell me what it is capable of; not what it is capable of when I invest a lot of work into it myself. The whole point of an AI is that it does the work for me.
If you register you can try it for free.
I was playing with this a bunch when it dropped a week ago. It's a big step up over their prior model, but it's nowhere near gpt-image-2 at least for high density UI design. Photos are a solved domain in my eyes, so the real question is if it can do infographics/web design.
Here are some samples, each with the same prompt:
Cannabis site
gpt-image-2: https://image.non.io/9fbf3396-889a-445e-b51b-ab4468ded269.we...
qwen-3: https://image.non.io/b5061e30-fafa-496f-9dad-2071e9473998.we...
Bookstore site
gpt-image-2: https://image.non.io/7157afee-914e-4433-9d5e-e3d0c8d8b3a3.we...
qwen-3: https://image.non.io/9556857b-e6f3-4754-8958-74b2265d946f.we...
While it does text fairly well, the overall layout/aesthetics are simply behind.
The bookstore design is really nice. Did you use diffui.ai?
Also, mind if I take it for my open-source bookstore SaaS?
Yea, the prompt was the expanded json the diffui harness uses. And re grabbing the designs, go for it.
I expaneded out the designs here - feel free to copy them for your agent to build: https://diffui.ai/app/canvas/95934269-5dc8-4145-a33a-d3a4dc2...
I'd recommend adding something like this to the copied prompt:
"Focus heavily on the angular cuts. Use image-to-svg to generate the hero graphics and also generate images of the hero graphics - let me choose between the two of them."
That should help steer the agent a bit and will give you some optionality for the hero graphics.
Thanks for expanding the designs for it, they look great! I need nice storefronts for my open-source SaaS but 5.6 Sol has been making them look really generic. I tried Google Stitch but it was much of the same. I really like diffui, but is there a way for me to use in Codex directly?
Looks like this model is meaningfully less good than gpt-image-2. Arena.ai score is 1263 vs 1380.
Everything is less good than gpt-2-image and I suspect that will be the case for awhile, until potentially Nano Banana Pro 2.
However, cost is significantly lower in this case. A Pro image here is $0.04, a gpt-2-image high is $0.21 and lower resolution.
gpt-2-image has such a yellow tinge though. It stands out badly on nearly any screen. The image quality is great otherwise, but I generally prefer nano banana for its color balance.
True but it’s at least addressable with some basic tone-mapping changes. It’s easier to correct an issue like this than to deal with an image that simply doesn’t follow your prompt.
Here's a quick trend of the "piss filter" in the gpt image series:
Gpt 2 image is unlimited on a chatgpt pro plan
Which makes it a clear price win if you already subscribe to Pro for some other reason, otherwise the price crossover where ChatGPT Pro unlimited beats the per-image pricing cited here is 80+ images/day.
As I’ve said before, I wouldn’t put a lot of stock in Arena’s scoring system. They have MAI Image 2.5 ranked above Gemini Nano Banana Pro, and maybe that’s true on paper but good luck using it. Microsoft’s censorship makes Google feel like the wild west by comparison.
They also have Meta’s Muse Image ranked above NB Pro, which is just patently absurd. In my own GenAI benchmark it only managed a lackluster 7 passes out of 15. Even the open‑weight Ideogram 4 scored higher than that.
For reference, here’s GPT‑Image‑2, NB Pro, and Muse compared:
The photo rankings on this page are so absurd that the only reasonable explanation is that they were judged by an AI.
The rankings are for prompt adherence, not subjective quality.
Less than 10% is meaningful?
If these were processors, I wouldn’t spend another hundred on the faster one…
The scores are pairwise ELO rankings, not simple metric scores.
Also available on OpenRouter (whose image output support is getting better): https://openrouter.ai/qwen/qwen-image-3-pro
is this going to get released on huggingface or is it cloud/hosted only?
They promised weights for Qwen Image 2 around 6 months ago and there's still nothing, so wouldn't get your hopes up...
The page lists it as "No" for open source, so I wouldn't hold my breath.
The page does not mention open source.
>The page does not mention open source.
on the right-hand side, it says:
"Version Tag: MAJOR
Open Source: No
Updated: Jul 20, 2026"
Given that Qwen Image 2 didn’t even get an open-weight release, I really doubt we’re going to see any kind of open release for it. They seem to have completely shifted over to proprietary models sadly.
When I tried to sign up for Qwen last month (after getting an ad on youtube), after entering my credit card for a "free" trial, Alibaba wanted me to upload a picture of my driver's license to be able to use it! I almost sold by BABA.
Will be interesting to see if they release the weights for local deployments. I love running Flux2 locally via ComfyUI.
I doubt any labs will be releasing weights for frontier image models.
Why would they want to deal with the amount of bad press they'll get for unsavory porn? Look at what happened to Grok.
Minimax just released the weights for a video model (https://huggingface.co/MiniMaxAI/MiniMax-H3) that apparently only has limited safeguards.
Pretty annoying user experience trying to test out the model on the official Qwen site. You get an immediate popup with "Security Notice: For security purposes, please add a payment method" that blocks you from clicking anywhere.
Honestly if you're launching a new model, you should probably just make it free for the 24 hours after launch so people can test it out. Otherwise just wait till it's on openrouter.
Not a great time to announce a pile of closed-weight excuses and locked gates when MiniMax-H3 just dropped.
Qwen lost some key people after 3.6 and it looks like we're seeing the consequences of that, or perhaps the cause.
And now you got 90% of the people saying the model sucks as it timed out while serving them
What's the difference between Qwen Cloud and Alibaba Cloud? Same company?
Similar to aistudio vs vertex.
Or rather, Alibaba cloud is equivalent to google cloud or aws.
Old news, no?
The model was added to OpenRouter yesterday.
Seems like the title order is incorrect at the time of this comment. It says Qwen 3.0 Image, but it's Qwen Image 3.0.