Qwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledge
qwen.ai381 points by ilreb 7 hours ago
381 points by ilreb 7 hours ago
To me it feels so weird that people are trying to push these model for online shopping like "here is how this dress/shirt/pants would look on you". But these models will always make the clothes fit your body and show you in flattering light and so on. How the actual garment fits is still as elusive as before these tools
The short-term goal of a tool like this is to sell products. The more ambitious long-term goal is to shift cultural norms, blurring the lines between advertising and reality until the question you're asking is no longer consciously asked. At least, not by average people, and not at the point of purchase.
I find it easy to envision a world, maybe 50 years from now, in which the very concept of "truth in advertising" is viewed as a lost, idyllic fantasy. Something people are nostalgic for, but feel powerless to regain.
> I find it easy to envision a world, maybe 50 years from now, in which the very concept of "truth in advertising" is viewed as a lost, idyllic fantasy.
From what I hear that's already the case right now with much more consequential transactions, like renting real estate in NYC. Square footages that are blatant lies, etc.
> I find it easy to envision a world, maybe 50 years from now, in which the very concept of "truth in advertising" is viewed as a lost, idyllic fantasy.
I find it easy now!
I've noticed something similar in Facebook marketplace ads for used furniture. Most of the images are AI generated to look like a Pottery Barn catalog, then the last image will be the actual item, full of scratches and other damage, sitting in a messy garage.
I've started to see this on Etsy and Wayfair too, where there will be a listing that is clearly just MDF flatpack being resold from China, but the AI-generated images wildly exaggerate the proportions of it.
Here's a recent example: https://www.etsy.com/ca/listing/4509158065/corner-wall-shelf...
Ironically, ChatGPT is decently good at ferreting these out. Like I sent it a screenshot of that listing and it not only helped me find where the original item was for sale, but also pointed out how the dimensioned diagram shows it as being just 49" tall, whereas the "in real life" image looks like it's at least six feet, based on it coming up over the top of the picture frame.
I ended up engaging a local woodworker to make me a piece like it instead. Obviously way, way more expensive, but it will be real solid walnut finished to match my dining table.
Estate agents are also doing it; making interiors of houses look very different to how they really are.
Yes and even if all they do is ask it to stage an empty room picture with furniture and decorations it will make improvements like adding a doorway to a room that doesn’t exist.
The worst I've seen changed the view out a window from an alley to a private garden. Most also add natural light that isn't possible, and in some cases I'm convinced the generate image is depicting the space as larger than it actually is, with furniture and spaces that wouldn't fit in reality
I once heard about a company that specialized in making furniture about 10% smaller than normal for use in show rooms to make the spaces look larger. Not surprised they do it with AI if they can.
This is a thing in show homes on new-build estates in the UK. British houses are often tiny.
Wide angle lenses have been a thing in real estate photography since forever, but you would always have the reference of the furniture to ground your perception. Having furniture be slightly downsized is diabolical.
And knocking out the messy garage background to replace it with a showroom is the easiest prompt.
Is it gonna be less distorted than just seeing the shirt on a model photographed by a professional in the perfect light?
Many product/model shots actually use clips to make the clothing look like it's perfectly fitted to the body. [1] This image is actually over 15 years old at this point. I think there should be laws that prevent this because it veers into false advertising though others believe it's alright because you can theoretically tailor the clothes to fit like this.
Regardless, I'd rather see real clothing on a real person when it comes to my purchasing decisions. I buy a lot of vintage clothes online and I've noticed a dramatic uptick in AI images of models wearing the clothes. I've never once bought from those sellers because it feels disingenuous. Sometimes they have fake runways which is actual false advertising because it makes the item appear more expensive than it really is. I've also noticed that the AI models' body types are always thin even if the item is a L or XL. Needless to say, the AI isn't showing me what an XL looks like on a small model; it's showing what a small model would look like if the item fit perfectly.
[1] https://www.primermagazine.com/wp-content/uploads/2011/02/St...
That image isn't from a clothes catalog, they'd generally use the unaltered garment there.
Yes, because at least that won't be flattering you (not to mention it would be showing the actual garment).
Sounds exactly like something the marketing department would use to up the convertion rate
The more honest ones are “how it looks on you” and don’t promise fit that depends on so many measurements that aren’t even visible in pics.
Sure, it’s idealized, but some people benefit from seeing color / neckline / etc on themselves as a visual reference.
Me, I’m a text-learner so I don’t get it at all. But I know people who get value.
If it doesn’t fit, then you must have gained weight while the item was in transit
I use nanobanana 2 to test changes in paint and flooring/tiling to great success. And the real pro move is taking those images to a designer to tweak the remaining 20% or so.
You just described how people are using AI in general. And I think that last step is actually a huge gap. People will brainstorm or improve their understanding of the task with AI and create something. Ideally they would hand it off to a professional for finalization like you did.
I wish a service existed where I could make something in AI (text, images, whatever), then pass that AI output to an actual human who would use it as a guide to produce an actual product.
There's definitely a predatory "everything will look good", but you can also leverage exactly that part for your own purposes - I've done pre-shopping a couple of times by asking for something like "a grid of 9 versions of my photo, wearing different types of X that look good". Definitely helped with choosing a good style.
I bet if you had a way to easily train a LORA on you trying on various clothing types the models would do pretty well but I agree without that I can't think of a way to get it to work. Reminds me of the flattering mirrors scam https://finance.yahoo.com/news/company-swears-controversial-...
Good or bad, I’m pretty sure showing things in the best light has always been the point of “marketing”. There is a reason ads aren’t filled with ugly people with misshaped bodies, and it isn’t because the intention is to reflect reality, so what you allude to as problematic is what a lot of businesses call a feature.
I'm pretty sure I saw somewhere people can sell re-sell clothes. They can include a picture of themselves wearing the item and it replaces literally everything but the clothing with someone/someplace "prettier".
In the end the one thing that is completely honest is the portion of the picture that is the item you are re-selling. But somehow to me the entire thing feels disingenuous.
The cynic in me says this is just an advancement and logical continuation from the known problem of sycophantic behavior in text-to-text chatbot format LLM, to image generation models.
To me the tools are newer, but fit people in a controlled context to sell clothing is as old as photography itself.
i'd like it to notice things i would miss, like "this is ring-spun shirt, so it will sit like this on your torso" or "these pleats will require you to iron them" etc
A prompt went viral recently where people were sharing their pictures and asking chatgpt to visualise what their looksmatched partner would look like
The result was always someone extremely good looking
There’s going to be an entirely new class of mental disorders that will emerge from people being deluded by AI
> But these models will always make the clothes fit your body and show you in flattering light and so on
This is something that can be fixed over time. And if this forces clothing manufacturers to stick more to their advertised "specs" (width/length), then it's a win for us.
Nobody that's pushing it has any incentive for fixing it - and they have all the opposite incentives.
Maybe this will lead to more business for tailors doing alterations, assuming the clothes people end up buying are expensive enough to justify it.
I understand they are already doing decent business since the advent of the magic weight loss pen
I've been struggling with this question myself. But isn't this a model training/use problem (a.k.a skill issue )? Isn't there a way to make these models be faithful to how people will actually look?
To be fair in many cases the actual clothes aren't much better. I've had two pair of the same pants, same brand, same size fit noticeably differently.
I have a similar use case at work for previewing construction material and such in our catalogue applied to user uploaded images.
Results are mixed, expensive, but it really feels you're few months off the next improvement to really nail it. It's already good enough.
Wonder what Qwen image will provide over nano banana.
The meta keywords in the HTML is very interesting. 100+ references to NSFW topics such as hentai, nudes, etc.
That's hilarious. These keywords are applied globally, even on pages like https://qwen.ai/usagepolicy, where they very kindly ask you not to use their products for sexual content.
Run this in console to see all the tags:
document.querySelector('meta[name="keywords"]').contentWTF lol
who made qwen stefani's dress
spiderman into the spider verse did qwen meet Peter
qwen stacey porn
do blake shelton and qwen stafani have children togetherSome of it is a reference to https://en.wikipedia.org/wiki/Spider-Woman_(Gwen_Stacy)
(i.e. the porn references)
Why would they added URLs for porn websites to their keywords? Like literally the names of websites with ".com" and such at the end.
I assume they use a SEO tool that automatically adds these meta keywords to optimize for some search engines.
It apparently adds common search terms that contain words like "qwen". This evidently includes possibly mistyped searches for "gwen" or "ben" in a NSFW context.
Maybe someone knows more about how such SEO tools work, and where they pull the data from.
What value does the 'keywords' meta tag even have these days? 77+ KB of crap added to the page weight. Web development is full of idiots.
Bonus points for also covering misspellings like "pregbant".
how is babby formed? how girl get pragnent?
If a women has starch masks on her body does that mean she has been pargnet before.?
show bobs and vegana!
That's racist!
(See kids, it is possible to fight memes/racism with memes! And well... yeah, this really is racist.)
This reddit slop is unbecoming of HN.
Please don’t violate community guidelines like this. All HN threads look like this. Don’t make me report you to a submod.
> Please don't post comments saying that HN is turning into Reddit. It's a semi-noob illusion, as old as the hills.
> Don’t make me report you to a submod.
This is pathologically outside scope. I don't think I have ever seen somebody threaten somebody else on hn before. The 'don't make me do it to you' abuser trope is next level.
Wow you really weren't kidding. Fully automated AI slop tentacles for everyone!
It also seems to have an obsession with this celebrity, based on how many times ctrl-f for "stefani" turns up a result.
Wow, crazy things - "shemale porn", "spider fucks venom", and various porn URLs like tushy.com. What the hell did they do to their AI? Is that the training corpus or intended use case?
> spider qwen thicc, spider qwen tied up, spider qwen toddler costume, spider qwen trans, spider qwen wallpaper, spider qwen xxx hentai
> ben 10 hentai game where qwen gets drunk
erm... what?
Surely it was meant to be "Gwen" (the protagonist's cousin and the other main character in the show). I guess it was somehow (incorrectly) assumed to be a misspelling of Qwen and included in the tags. Or perhaps many people were misspelling Gwen as "Qwen" and it was all hoovered up.
> qwen ten hentai, qwen ten xxx, qwen tennason footjob, qwen tennison hentai, qwen tennyson, qwen tennyson ass expansion, qwen tennyson giantess pussy, qwen tennyson hentai, qwen tennyson porn, qwen tft, qwen the milk maid, qwen the milkmaid, qwen the parts over, qwen the wolf witcher, qwen tire, qwen tneyson porn, qwen to go feelies youtube, qwen tokenizer java,
Whelp today’s a unique day on hacker news, wow! Didn’t expect to read that at my desk lol
That is hilarious as porn is illegal in China. But I guess pron SEO is allright, if it is against the west.
Western search engines simply ignore the meta keywords tag for over a decade now
So what is the point then?
There is no point. It's from people that have no clue which is most people that do search engine optimization
That said it's possible search engines in China or other countries might use it, but it's very easy to game so it doesn't really make sense
uh, are you implying that porn is bad somehow? I can show you a hundred articles that say only nazi incel chuds believe so.
https://www.nytimes.com/2019/06/07/us/hate-groups-porn-consp...
porn is as American as apple pie.
> porn is as American as apple pie
So not at all then as Apple Pie is a traditional English desert. :P
Incels are probably the biggest porn consumers.
Apart from that, I made no judgement about porn in general, just about porn tags in Chinese backed AI websites.
No pricing table? No benchmarks? This is just marketingslop.
Image token pricing has been fairly steady while text token prices fall, yet image model release discussion seems to be more focused on how beautiful the women the model generates are versus any sort of substantive discussion.
They must have trained on GPT Image 1 outputs. The yellow tint is unmistakable.
https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen-Image/i...
https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen-Image/i...
https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen-Image/i...
https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen-Image/i...
GPT Image 1 ended up with a yellow tint without training on another image-generation model's output. It's just that humans like pictures with a soft sunset glow, and this is a very easy global signal for a preference model to pick up on, and for a image-generation model to imitate. So optimizing for aesthetic appeal makes everything slightly tinted by default, unless you make sure to countersteer.
yellow/red tint is an extremely common problem not matter the photograph source you train on
source: work at a photograph start, even training on raw images things get tinted, it is an uphill battle
AI-generated images are part of the web now, if you're doing ordinary web scraping, you can't avoid training on generated images.
Very cool but the Arabic text in the title image is obviously and hopelessly broken, which is oddly not the case when actually using the model. Could it be that the hero image was not generated by Qwen Image 3.0?
that's a very interesting point to test new models. I speak an Indian language called "Tamil" and have always tested new models with Tamil but also have tried little bit of Arabic (quranic verses) with previous GPT image 2 and Nanobanana pro and they have nailed it. Don't know if it was because of extensive training data.
What I can not wrap my head around: How are these models trained?
What training mechanism or model architecture provides the glue to go from human text to images?
Don't you need to have millions of really descriptively labelled images?
That's exactly how they do it.
There are ML models that do the reverse and output image to text, which assist quite a lot.
The better the text represents the unique thing in the photo, the better the model understands what that text means.
Short answer yes.
Slightly longer answer for older text to image models you teach them how to encode images and text into the same latent space. Then you simply do a conversion, take a text input, put it into latent space and then extract the image that latent space represents.
I found this 3b1b guest video on diffusion helpful: https://www.youtube.com/watch?v=iv-5mZ_9CPY
Random question, but has there been any improvement in OCR/document understanding in these newer models? Last time I checked (1mo ago) SOTA was still sadly Gemini, unless you wanted to pay $$$ for e.g. Sol
> to precisely describe the full 3×3 grid takes a full 3.7k tokens
It's a shame they didn't share that prompt - it would make that demo more convincing.
Impressive, but their woman face generation always use very similar, too perfect, same prettiness faces, it's very obvious.
I’d argue that talking about “authentic” AI-generated images is oxymoronic.
Not a single word about when/if they'll actually release the weights for this, or am I missing it somewhere?
Considering Qwen-Image-2.0 weights have not been released either, it unfortunately looks unlikely.
Why release it just so Cursor/Azure/Amazon can profit off of it? Unless OpenAI actually opens it up fat chance.
I am surprised by the rather bad output. It doesn't achieve qwen image 1 quality in composition or anatomical correctness. Tested on chat.qwen.ai
I am seeing third legs and glowing eyes. It's a Microsoft Lens level of quality and that one was pulled.
I truly wish these models were available when I was in University. As a visual learner it would have been much easier for me to understand certain topics with illustrative diagrams rather than reading a wall of text.
…but the illustrative diagrams are a simulacrum; if you ask Qwen, or any image-generator, for an “accurate” poster-design featuring a representation of a model of an atom and explaining its constituent parts I expect you’ll get an imitation-airbrush rendering of red, blue, and grey table-tennis balls orbiting in perfect circles; you might get an electron-shell diagram if you’re lucky. What you won’t get is anything remotely related to probability-clouds.
Edit: just to test myself I asked Nano Banana 2 to generate “an undergraduate infographic poster about how atoms work” - and the result was something right out of a middle-school science textbook and very Bohr…
This is the image I got for the same prompt: https://jumpshare.com/s/mqqBdl7U59FWPiEXjwoM. It's more like high-school level, but not bad. I can imagine a collage professor can improve the prompt the create a more accurate and detailed diagram
This seems like something that could be solved by asking an LLM to write the prompt for the image model. You can also feed in the output of an image model into an LLM and ask it to check it/make improvements.
My wife has taken all her recipes and fed them through ChatGPT image gen to make zine pages and they’re really cool! She’s building a recipe book for the kids so they’ll know all the recipes from their childhood.
The "piss filter" is still everywhere.
It failed to create a simple overlay on a map - something ChatGPT had no issues with.
> Supports up to 4.5k token input, effortlessly generating complex layouts such as newspapers, storyboards, and exam papers.
Impressive.
Btw, what is currently the best model to run locally on a 16GB Vram? Is it Z-Image Turbo?
Krea-2-Turbo. I've even got it working locally on my M5 iPad Pro.
> We implemented safety measures across the full model development lifecycle.
Any suggestions for the best open, non-opinionated model?
It is a reasonably non-opinionated model. My usual test is to ask these models to generate comic book and cartoon characters that hosted image generation models refuse to generate. I think that text is just CYA legalese.
The examples posted on their launch blog page are quite impressive, especially for fine details, multi-panel/multi-page and text rendering.
But: not open-source/open-weights, and no indication that weights/source will be released either.
Zooming in on mobile on that website causes a large white area to obstruct the page. Might wanna look into that.
As for the image model, wow...
How long until we get rid of the AI "plasticness" in portrait kind of generated images?
The right models and LORAs can get rid of it right now. People apparently really like everyone to look like over exposed over filtered mannequins, so the big companies target that look.
There is an IRL phenomena, glass skin skincare and makeup, where the person's skin is evenly flat and toned and glossy. Did you think they are more plasticky than IRL models?
That's supposed to make skin appear lively. Human skins in AI images tend to look clouded, opaque, and overall un-alive, so to speak.
Seems that this will not be open weights?
I assume Van Gogh didn't paint enough hands to train from!
> In the three examples below, the model accurately renders Japanese, Korean, and Spanish respectively.
And yet the Korean text is not accurate... [1]
[1] E.g. "드레스 컬렉션 dress collection" has vowels ㅔ mixed with ㅐ, "초웜한" should be "초월한 exceeding", "신키한" should be "실키한 silky", "디자언되다" should be "디자인되다 have been designed", "로얼" should be "로열 royal", and so on.
the natural endpoint of this technology is product photos that look better than the actual product, which is going to make unboxing videos the last remaining source of truth on the internet
I wonder if they got permission to generate that (admittedly impressive) Berserk image.
Again not released on huggingface immediately?
I am curious whether the model requires a font to be installed. Does it also generate the glyphs for the text?
Yes, it generates the text without using a font. Same is true of other image models like ChatGPT Images and Gemini Nano Banana and Midjourney.
The generated latex pdf!
Appears to be closed-weights entirely. No word at all on any weights release.
It will be interesting to compare this to Flux 2.
The red-dress woman's vestigial pinkie toes …
The blog write-up style is so casual haha.
Wow, it displays Korean properly without breaking. But there are still a lot of typos. Haha, it's good that Korean displays properly, but there are a lot of incorrect sentences
I find it mildly interesting how your comment repeats itself but with different phrasing
It's because of the structure of Korean. I think in Korean first, so I end up translating it directly.
When I want to emphasize something, I tend to repeat it
Midjourney already knew that image generation was going to zero. Again yet another reason why the model was never a moat in the first place.
interesting
This looks impressive！
The real performance is nowhere close to what is presented in the marketing materials, which is pretty annoying. Especially text rendering and accuracy.
Try asking it for a plot of Polish GDP growth over the past 20 years. It's slop.
This model isn't supposed to contain all the numerical data. It will give you a (usually) matching graph transformed from one you provide or from a table of information you provide. Or you can pipeline from an LLM doing research on that data first. But expecting an image gen model to get you GDP info has got to be one of the worst possible approaches.
> Especially text rendering
That's true though. I still got some completely fried letters in headings.
I am not expecting it, it's the Qwen team is claimingthey can do much harder tasks than this, like rendering a consistent page of a maths paper, or creating true to fact explainers.
They can't.
Giving it numerical data in an LLM-generated prompt doesn't seem to help much: https://imgur.com/a/KFhczOd
It included the table verbatim and even managed to hallucinate a reasonable heading for it, but then the graph doesn't even manage to align the data points with the time axis, leading to an unfortunate collision in the middle.
I guess you should use a traditional graphing library for your presentation slides for now.
... but if you want an accurate graph, why not ask the LLM model to put the data points into a graphing library?
[flagged]
[flagged]
[flagged]
[flagged]
THIS IS EXACTLY WHAT I WAS LOOKING FOR
why I cant submitte new posts here ? My account since 2016
