Hacker News

Interesting background! Prompts like this also test the latent space of the image generator - it’s usually the other way round, so if you see a man on top of a horse, you’ve got a less sophisticated embedding feeding the model. In this case, though, that’s quite an image to put out to the interwebs. I looked to see what gender the horse was.

EDIT: After reading the prompt translation, this was more just like a “year of the horse is going to nail white engineers in glorious rendered detail” sort of prompt. I don’t know how SD1.5 would have rendered it, and I think I’ll skip finding out

This is fascinating!

From the article it seems the name is 马启仁, not 马骑人 so the guy's name sounds the same as 'horse riding man', but that's not a literal translation of his name.

There's also the "horse riding astronaut" challenge in image generation: https://garymarcus.substack.com/p/horse-rides-astronaut-redu...

While I don't doubt this was one influence, there was also an infamous problem with Dall-E 2, which was perfectly able to generate an astronaut riding a horse but completely unable to generate a horse riding an astronaut.

This problem is infamous because it persisted (unlike other early problems, like creating the wrong number of fingers) for much more capable models, and the Qwen Image people are certainly very aware of this difficult test. Even Imagen 4 Ultra, which might be the most advanced pure diffusion model without editing loop, fails at it.

And obviously an astronaut is similar to a man, which connects this benchmark to the Chinese meme.

On the topic of modern Chinese culture, Is there the same hostility towards AI generated Imagery in China as there seems to be in America?

For example I think there would be a lot of businesses in the US that would be too afraid of backlash to use AI generated imagery for an itinerary like the one at https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwe...

Very interesting! What's weird though is that the chinese do not even pretend: every single picture has asian-looking people generated.

But on the one picture that honestly looks like a man getting ass-raped by a horse, it's a white man.

I mean even in the west where you can hardly see an ad with a white couple anymore, they don't go that far (at least not yet).

White people are a minority on earth and anti-white racism sure seems to be alive and well (btw my family is of all the colors and we speak three languages at home, so don't even try me).

Why not ask for simply a man or even an Han man given the race of Tsai Kang-yong. Why a white man and why a man wearing medieval clothing. Gives your head a wobble.

Fun fact, the Serbian parliament building has two statues of horses riding men in front of it.

Which is really apt because in Serbian "konj", or horse, is a colloquial word for moron. So, horses riding people is a perfect representation of the reality of the Serbian government.

Another fun fact, the parliament building in HL2's City 17 was modelled from that building.

For the more technical…

Qwen 2512 (December edition of Qwen Image)

* 19B parameters, which was a 40GB file at FP16 and fit on a 3090 at FP8. Anything less than that and you were in GGUF format at Q6 to Q4 quantizations… which were slow, but still good quality.

* used Qwen 2.5 VL. So a large model and a very good vision model.

* And iirc, their own VAE. Which had known and obvious issues of high frequency artifacts. Some people would take the image and pass it through another VAE like WAN Video model’s or upscale-downscale to remove these

Qwen 2 now is

* a 7B param model. Right between Klein 9B (non-commercial) this (license unknown), Z-Image 7B (Apache), and Klein 4B (Apache). Direct competition, will fit on many more GPUs even at FP16.

* upgrades to Qwen 3 VL, I assume this is better than the already great 2.5 VL.

* Unknown on the new VAE. Flux2’s new 128 channel VAE is excellent, but it hasn’t been out long enough for even a frontier Chinese model to pick up.

Overall, you’re right this is on the trend to bring models on to lower end hardware.

Qwen was already excellent and now they rolled Image and Edit together for an “Omni” model.

Z-Image was the model to beat a couple weeks ago… and now it looks like both Klein and Qwen will! Z-Image has been disappointing to see how it just refuses to adhere to multiple new training concepts. Maybe they tried to pack it too tightly.

Open weights for this will be amazing. THREE direct competitors all vying to be “SDXL2” at the same time.

The Qwen convention was confusing! You had Image, 2509, Edit, 2511 (Edit), 2512 (Image) and then the Lora compatibility was unspecified. It’s smart to just 2.0 this mess.

The pace of commoditization in image generation is wild. Every 3-4 months the SOTA shifts, and last quarter's breakthrough becomes a commodity API.

What's interesting is that the bottleneck is no longer the model — it's the person directing it. Knowing what to ask for and recognizing when the output is good enough matters more than which model you use. Same pattern we're seeing in code generation.

Isn’t it still? Antidotally, I work with lots of creators who still prefer it because of its subjective qualities.

What ever happend to midjourney?

Practically anybody actually creating with this class of models (diffusion based mostly) is using ComfyUI. Community takes care of quantization, repackaging into gguf (most popular) and even speed optimizing (lighting loras, layers skip). It's quite extensive

Everything keeps changing so quickly, I basically have my own Python HTTP server with a unified JSON interface, then that can be routed to any of the impls/*.py files for the actual generation, then I have of those per implementation/architecture basically. Mostly using `diffusers` for the inference, which isn't the fastest, but tends to have the new model architectures much sooner than everyone else.

I encourage everyone to at least try ComfyUI. It's come a long way in terms of user-friendliness particularly with all of the built-in Templates you can use.

ComfyUI is the best for stable diffusion

If you're on an AMD platform Lemonade (https://lemonade-server.ai/) added image generation in version 9.2 (https://github.com/lemonade-sdk/lemonade/releases/tag/v9.2.0).

I have my own MIT licensed framework/UI: https://github.com/runvnc/mindroot. With Nano Banana via runvnc/googleimageedit

Ollama is working on adding image generation but its not here yet. We really do need something that can run a variety of models for images.

Stability matrix, it's a manager for models and uis and loras etc, very nice

Koboldcpp has built in support for image models. Model search and download, one executable to run, UI, OpenAI API endpoint, llama.cpp endpoint, highly configurable. If you want to get up and running instantly, just pick a kcppt file and open that and it will download everything you need and load it for you.

Engine:

* https://github.com/LostRuins/koboldcpp/releases/latest/

Kcppt files:

* https://huggingface.co/koboldcpp/kcppt/tree/main

LMStudio is a low barrier to entry for LLMs, for sure. The lowest. Good software!

Other people gave you the right answer, ComfyUI. I’ll give you the more important why and how…

There is a huge effort of people to do everything but Comfy because of its intimidating barrier. It’s not that bad. Learn it once and be done. You won’t have to keep learning UI of the week endlessly.

The how, go to civitai. Find an image you like, drag and drop it into comfy. If it has a workflow attached, it will show you. Install any missing nodes they used. Click the loaders to point to your models instead of their models. Hit run and get the same or a similar image. You don’t need to know what any of the things do yet.

If for some reason that just does not work for you… Swarm UI, is a front end too comfy. You can change things and it will show you on the comfy side what they’re doing. It’s a gateway drug to learning comfy.

EDIT: most important thing no one will tell you out right… DO NOT FOR ANY REASON try and skip the VENV or miniconda virtual environment when using comfy! You must make a new and clean setup. You will never get the right python, torch, diffusers, driver, on your system install.

To be fair it hasn't made LinkedIn any worse than it already was.

Infographics and full presentations are a NanoBananaPro exclusive so far.

Informatics are as bad as the author allows though. There's few people who could make or even describe a good infographic, so that's what we see in the results too.

Correct.

Much like the pointless ASCII diagrams in GitHub readmes (big rectangle with bullet points flows to another...), the diagrams are cognitive slurry.

See Gas Town for non-Qwen examples of how bad it can get:

https://news.ycombinator.com/item?id=46746045

(Not commenting on the other results of this model outside of diagramming.)

some context for the perplexed. https://live2makan.com/2024/08/07/treasures-statue-of-horse-...

To be fair, didn't they release open weights image model only like a ~month ago? Think last one was in December 2025.

"Guy being humped by a horse" wouldn't have been my first choice for demoing the capabilities of the model, but each to their own I guess.

It's the year of the horse in their zodiac. The (translated) prompt is wild:

""" A desolate grassland stretches into the distance, its ground dry and cracked. Fine dust is kicked up by vigorous activity, forming a faint grayish-brown mist in the low sky. Mid-ground, eye-level composition: A muscular, robust adult brown horse stands proudly, its forelegs heavily pressing between the shoulder blades and spine of a reclining man. Its hind legs are taut, its neck held high, its mane flying against the wind, its nostrils flared, and its eyes sharp and focused, exuding a primal sense of power. The subdued man is a white male, 30-40 years old, his face covered in dust and sweat, his short, messy dark brown hair plastered to his forehead, his thick beard slightly damp; he wears a badly worn, grey-green medieval-style robe, the fabric torn and stained with mud in several places, a thick hemp rope tied around his waist, and scratched ankle-high leather boots; his body is in a push-up position—his palms are pressed hard against the cracked, dry earth, his knuckles white, the veins in his arms bulging, his legs stretched straight back and taut, his toes digging into the ground, his entire torso trembling slightly from the weight. The background is a range of undulating grey-blue mountains, their outlines stark, their peaks hidden beneath a low-hanging, leaden-grey, cloudy sky. The thick clouds diffuse a soft, diffused light, which pours down naturally from the left front at a 45-degree angle, casting clear and voluminous shadows on the horse's belly, the back of the man's hands, and the cracked ground. The overall color scheme is strictly controlled within the earth tones: the horsehair is warm brown, the robe is a gradient of gray-green-brown, the soil is a mixture of ochre, dry yellow earth, and charcoal gray, the dust is light brownish-gray, and the sky is a transition from matte lead gray to cool gray with a faint glow at the bottom of the clouds. The image has a realistic, high-definition photographic quality, with extremely fine textures—you can see the sweat on the horse's neck, the wear and tear on the robe's warp and weft threads, the skin pores and stubble, the edges of the cracked soil, and the dust particles. The atmosphere is tense, primitive, and full of suffocating tension from a struggle of biological forces. """

I think they call it "horse riding a human" which could have taken two very different directions, and the direction the model seems to have taken was the least worst of the two.

Wont someone think of the horses.

For reference, below is the prompt translated (with my highlighting of the part that matters). They did very much ask for this version of "horse riding a man", not the "horse sitting upright on a crawling human" version

---

A desolate grassland stretches into the distance, its ground dry and cracked. Fine dust is kicked up by vigorous activity, forming a faint grayish-brown mist in the low sky.

Mid-ground, eye-level composition: A muscular, robust adult brown horse stands proudly, its forelegs heavily pressing between the shoulder blades and spine of a reclining man. Its hind legs are taut, its neck held high, its mane flying against the wind, its nostrils flared, and its eyes sharp and focused, exuding a primal sense of power. The subdued man is a white male, 30-40 years old, his face covered in dust and sweat, his short, messy dark brown hair plastered to his forehead, his thick beard slightly damp; he wears a badly worn, grey-green medieval-style robe, the fabric torn and stained with mud in several places, a thick hemp rope tied around his waist, and scratched ankle-high leather boots; his body is in a push-up position—his palms are pressed hard against the cracked, dry earth, his knuckles white, the veins in his arms bulging, his legs stretched straight back and taut, his toes digging into the ground, his entire torso trembling slightly from the weight.

The background is a range of undulating grey-blue mountains, their outlines stark, their peaks hidden beneath a low-hanging, leaden-grey, cloudy sky. The thick clouds diffuse a soft, diffused light, which pours down naturally from the left front at a 45-degree angle, casting clear and voluminous shadows on the horse's belly, the back of the man's hands, and the cracked ground.

The overall color scheme is strictly controlled within the earth tones: the horsehair is warm brown, the robe is a gradient of gray-green-brown, the soil is a mixture of ochre, dry yellow earth, and charcoal gray, the dust is light brownish-gray, and the sky is a transition from matte lead gray to cool gray with a faint glow at the bottom of the clouds.

The image has a realistic, high-definition photographic quality, with extremely fine textures—you can see the sweat on the horse's neck, the wear and tear on the robe's warp and weft threads, the skin pores and stubble, the edges of the cracked soil, and the dust particles. The atmosphere is tense, primitive, and full of suffocating tension from a struggle of biological forces.

I like how sometimes I get angry at a LLM for not understanding what I meant, but then I realize that I just forgot to mention it in the context. It's fun to see the same thing happen in humans reading websites too, where they don't understand the context yet react with strong feelings anyways.

I agree. They makes me nauseous. The same kind of light nausea as car sickness.

I assume our brains are used to stuff which we dont notice conciously, and reject very mild errors. I've stared at the picture a bit now and the finger holding the baloon is weird. The out of place snowman feels weird. If you follow the background blur around it isnt at the same depth everywehere. Everything that reflects, has reflections that I cant see in the scene.

I dont feel good staring at it now so I had to stop.

The lighting is wrong, that's what's telling to me. They look too crisp. No proper shadows, everything looks crystal clear.

Everything is weightless. When real people stand and gesture there’s natural muscle use, hair and clothing drape, papers lay flat on surfaces.

At least for the real life pictures, there’s no depth of field. Everything is crystal clear like it’s composited.

Qwen always suffered from their subpar rope implementation and qwen 2 seems to suffer from it as well. The uncanny feel is down to the sparsity of text to image token and the higher in resolution you go the worse it gets. It's why you can't take the higher ends of the MP numbers serious no matter the model. At the moment there is no model that can go for 4k without problems you will always get high frequency artifacts.

Agree, looks like the same effect they are applying on YouTube Shorts...

For me the only model that can really generate realistic images is nano banana pro (also known as gemini-3-pro-image). Other models are closing the gap, this one is pretty meh in my opinion in realistic images.

they explicitly called for that in the prompt

Is the problem the position/horse or that Qwen mostly shows asian people?

Do western AI models mostly default to white people?

That's not what they did with Qwen-Image v1 - they announced it and it was available via API, but then they released the weights a few weeks after with an Apache 2.0 license. Let's at least give them the benefit of the doubt here.

Good that we have the arbitrator of what "real open source" is and isn't over here.

Where do you see a press release countdown? Alibaba consistently doesn't release weights for their biggest models, but they also don't pretend that they do.