I think it shows problems with your tests tbh. The bigger models are way more capable than you make them out to be. They are also better in training and understanding of CGI render outputs as reference like normal maps or id-masks. Your testing suite is the perfect example that structured data implies false confidence. Pure t2i is not a good benchmark anymore.

Can you fix the information bubble on mobile please? When pressing one, it vanishes instantly...

There are probably some more subtle tipping points that small models hit too. One of the challenges of a 100GB model is that there is non-trivial difficulty in downloading and running the thing that a 4GB model doesn't face. At 4GB I think it might be reasonable to assume that most devs can just try it and see what it does.

Is there a theoritical minimum for params for a given output? I saw news about GPT 3.5, then Deepseek training models at a fraction of that cost, then laptops running a model that beats 3.5. When does it stop?

Quality is increasing, but these small models have very little knowledge compared to their big brothers (Qwen Image/Full size Flux 2). As in characters, artists, specific items, etc.

When given an image of an empty wine glass, it can't fill it to the brim with wine. The pogo stick drawers and wine glass fillers can enjoy their job security for months to come!

It's a tough test for local models - (gpt-image and NB had zero problems) - the only one that came reasonably close was Qwen-Image

Z-Image / Flux 2 / Hidream / Omnigen2 / Qwen Samples:

https://imgur.com/a/tB6YUSu

This is where smaller models are just going to be more constrained and will require additional prompting to coax out the physical description of a "pogo stick". I had similar issues when generating Alexander the Great leading a charge on a hippity-hop / space hopper.

You are right, just tried even with reference images it can't do it for me. Maybe with some good prompting.

Because in theory I would say that knowledge is something that does not have to be baked in the model but could be added using reference images if the model is capable enough to reason about them.

Those are both good benchmark prompts. Z-Image Turbo doesn't like them either:

Tiger on pogo stick: https://i.imgur.com/lnGfbjy.jpeg

Dunno what this is, but it's not a pogo stick: https://i.imgur.com/OmMiLzQ.jpeg

Nano Banana Pro FTW: https://i.imgur.com/6B7VBR9.jpeg

Local models. I'm not gonna wait 10 min for one image on my computer like I did back in the Stable Diffusion days. And image editing in particular.

Maybe fast image editing, since it supports that.

> Generally, text compresses extremely well. Images and video do not.

Is that actually true? I'm not sure it's fair to compare lossless compression ratios of text (abstract, noiseless) to images and video that innately have random sampling noise. If you look at humanly indistinguishable compression, I'd expect that you'd see far better compression ratios for lossy image and video compression than lossless text.

I find it likely that we are still missing a few major efficiency tricks with LLMs. But I would also not underestimate the amount of implicit knowledge and skill an LLM is expected to carry on a meta level.

Images and video compress vastly better than text. You're lucky to get 4:1 to 6:1 compression of text (1), while the best perceptual codecs for static images are typically visually lossless at 10:1 and still look great at 20:1 or higher. Video compression is much better still due to temporal coherence.

1: Although it looks like the current Hutter competition leader is closer to 9:1, which I didn't realize. Pretty awesome by historical standards.

The team behind Z-Image Turbo has told us multiple times in their paper that the output quality of the Turbo model is superior to the larger base model.

I think that information still did not get through to most users.

"Notably, the resulting distilled model not only matches the original multi-step teacher but even surpasses it in terms of photorealism and visual impact."

"It achieves 8-step inference that is not only indistinguishable from the 100-step teacher but frequently surpasses it in perceived quality and aesthetic appeal"

https://arxiv.org/abs/2511.22699

I’m a bit confused, both you and another commenter mention something called Z-Image, presumably another Flux model?

Your frame of it is speculative, i.e. it is forthcoming. Theirs is present tense. Could I trouble you to give us plebes some more context? :)

ex. Parsed as is, and avoiding the general confusion if you’re unfamiliar, it is unclear how one can observe “the way it is looking”, especially if turbo was released months ago and there is some other model that is unreleased. Chose to bother you because the others comment was less focused on lab on lab strategy.