It can be (slowly) run at home, but needs 96GB RTX 6000-level hardware so it is not very popular.

Here's ZiT, Gpt-Image-2, and Hunyuan Image 2 for reference:

https://genai-showdown.specr.net/?models=hy2,g2,zt

Note: It won't show up in some of the newer image comparisons (Angelic Forge, Flat Earth, etc) because it's been deprecated for a while but in the tests where it was used (Yarrctic Circle, Not the Bees, etc.) it's pretty rough.

Ring toss: https://i.imgur.com/Zs6UNKj.png (arguably a pass)

9-pointed star: https://i.imgur.com/SpcSsSv.png (star is well-formed but only has 6 points)

Mermaid: https://i.imgur.com/R6MbMPX.png (fail, and I can't get Imgur to host it for some reason even though it's SFW)

Octopus: https://i.imgur.com/JTVH7xy.png (good try, almost a pass, but socks don't cover the ends of all the tentacles)

Above are one-shot attempts with seed 42.

You're killing me Smalls. This one is a 404. I'm really curious what it actually showed.

That ring toss is definitely leagues better than its predecessor. I’m not going to fault it too much for the star though, that one is an absolute slate wiper. The only locally hostable model that ever managed it for me was the original Flux, and I’m still not entirely convinced it wasn’t a fluke. Despite getting twice as many attempts, Flux 2, a much larger model, couldn’t even pull it off.

For the mermaid, https://i.imgur.com/R6MbMPX.png sometimes seems to work but not consistently. It is probably triggering a porn filter of some kind. I need to find another free image host, as imgur has definitely jumped the shark.

The image shows a mermaid of evident Asian extraction lying on a beach, face down. There is a dolphin lying on top of her, positioned at a 90-degree angle. It doesn't show any interaction at all, so a definite fail.

The template prompt seen in each comparison gets adjusted through a guided LLM which has fine-tuned system prompts to rewrite prompts. The goal is to foster greater diversity while preserving intent, so the image model has a better chance of getting the image right.

Getting to your suggestion for posting all the raw prompts, that's actually a great idea. Too bad I didn't think about it until you suggested it. And if you multiply it out - there's 15 distinct test cases against 22 models at this point, each with an average of about 8 attempts so we’re talking about thousands of prompts many of which are scattered across my hard drive. I might try to do this as a future follow-up.

The prompts despite their variation are still expressed in natural language.

The idea is that if you can rephrase the prompt and still get the desired outcome, then the model demonstrates a kind of understanding; however more variation attempts also get correspondingly penalized: this is treated more as a failure of steering, not of raw capability.

An example might help - take the Alexander the Great on a Hippity-Hop test case.

The starter prompt is this: "A historical oil painting of Alexander the Great riding a hippity-hop toy into battle."

If a model fails this a couple of times (multiple seeds), we might use a synonym for a hippity-hop, it was also known as a space hopper.

Still failing? We might try to describe the basic physical appearance of a hippity-hop.

Thus, something like GPT-Image-2 scored much higher on the compliance component of the test, requiring only a single attempt, compared with Z-Image Turbo, which required 14 attempts.

I often have to make very specific edits while keeping the rest of the image intact and haven't yet found a good model. These are typically abstract images for experiments.

I asked gpt-image-2 to recolor specific scales of your Seedream 4 snake and change the shape of others. It did very poorly.

I don’t know how much work it is for you, but one thing a lot of people do, myself included, is take the original image, make a change to it using something like NB, then paste that as the topmost layer in something like Krita/Pixelmator. After that, we’ll mask and feather in only the parts we actually want to change. It doesn’t always work if it changes the overall color balance or filters out certain hues, it can be a real pain but it does the job in some cases.

The Flux models (like Kontext) are actually surprisingly good at making very minimal changes to the rest of the image, but unfortunately their understanding of complex prompts is much weaker than the closed, proprietary models.

I will say that I’ve found Gemini 3.0 (NB Pro) does a relatively decent job of avoiding unnecessary changes - sometimes exceeding the more recent NB2, and it scored quite well on comparative image-editing benchmarks.

https://genai-showdown.specr.net/image-editing

1 - Gpt-image-2 seems to pass the Flat Earth test? (if not, I'm sure the paid thinking 2k version passes it).

2 - Since NB2 was earlier, many gold medals are assigned to it, even though now GI2 passes them too, example the Octopus test NB2 14 attempts but GI2 just 2 (BTW number of attempts should affect the score I guess?)

This is one of those areas where even state-of-the-art models still struggle. You’re asking for a high level of detail at a per-person level, which means you end up with lots and lots of very small objects that all need to be rendered with convincing detail.

I should probably explain the scoring rubric better - it's in the (i) info icon. If you click the pass/fail button towards the top, it switches from a simple pass/fail view to a weighted score. That weighted score is based on three things: level of adherence to the prompt, visual fidelity, and the number of attempts.

I've tried to keep my criteria as objective as possible, but there's just a certain level of unavoidable subjectivity to it.

For example, with the octopus image: Even though the minimum criteria might be five tentacles covered, having all eight is much closer to the ideal of “an octopus,” so it usually gets bumped up to a higher rating (bronze, silver, gold).

Honestly, I think I agree that the gpt-image-2 probably should be upgraded to a gold medal. Thanks for pointing that out!

https://mordenstar.com/blog/edits-with-nanobanana/#through-t...

Kind of makes me want to take advantage of the multi-image editing capability, since you can use gpt-image-2 with multiple images.

Take a photo of an existing pair of glasses frames (maybe even snapped at an optometrist’s office) then take a picture of an animal, like a spider with an unusual number of eyes, or something like a flounder, where the eyes eventually migrate to the top of its body.

Then you could see if the system can realistically adapt the design and show how those glasses might look if they were redesigned for these unusual optical situations.

That’s why I gave it a bronze. To me, it falls into that “barely passing” category, similar to Gemini 2.5 Flash Image on that test. Seedream also took a major hit to its weighted score because of how many attempts it took to get something even remotely passable out of it.

Thanks for the feedback!