OpenAI’s gpt-image-1.5 and Google’s NB2 have been pretty much neck and neck on my comparison site which focuses heavily on prompt adherence, with both hovering around a 70% success rate on the prompts for generative and editing capabilities. With the caveat being that Gemini has always had the edge in terms of visual fidelity.
That being said, gpt-image-1.5 was a big leap in visual quality for OpenAI and eliminated most of the classic issues of its predecessor, including things like the “piss filter.”
I’ll update this comment once I’ve finished running gpt-image-2 through both the generative and editing comparison charts on GenAI Showdown.
Since the advent of NB, I’ve had to ratchet up the difficulty of the prompts especially in the text-to-image section. The best models now score around 70%, successfully completing 11 out of 15 prompts.
For reference, here’s a comparison of ByteDance, Google, and OpenAI on editing performance:
gpt-image-2 has already managed to overcome one of the so‑called “model killers” on the test suite: the nine-pointed star.
Results are in for the generative (text to image) capabilities: Gpt-image-2 scored 12 out of 15 on the text-to-image benchmark, edging out the previous best models by a single point. It still fails on the following prompts:
- A photo of a brightly colored coral snake but with the bands of color red, blue, green, purple, and yellow repeated in that exact order.
- A twenty-sided die (D20) with the first twenty prime numbers (2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61, 67, 71) on the faces.
- A flat earth-like planet which resembles a flat disc is overpopulated with people. The people are densely packed together such that they are spilling over the edges of the planet. Cheap "coastal" real estate property available.
Very useful website. Would you have insight into what models are best at editing existing images?
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.
OpenAI actually has really good adherence, but occasionally tends to introduce its own almost equivalent of "tone mapping", making hyper-localized edits frustrating.
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.
I’ll have to give it another try. Its predecessor, Hunyuan Image 2.0, scored pretty poorly when I tested it last year: 2 out of 15, so it'll be interesting to see how much it has improved.
Here's ZiT, Gpt-Image-2, and Hunyuan Image 2 for reference:
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.
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.
Yeah, I suspect you'd see some solid passing scores if you ran it as many times as some of the others.
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.
I still use Imgur from time to time just because it’s convenient, but I’ve been meaning to build an Imgur-style extension for my site for a while, something that would let me drag and drop media for quick sharing but it being Astro-based (static site generation) makes it tricky.
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?)
So if you zoom in (click the zoom button on the actual gpt-image-2 of the flat Earth), you’ll see that a lot of the people are anatomical impossibilities, which is one of the disallowed criteria on the list. The faces also look like melted candles.
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!
That's lovely. My own personal benchmark has been to ask the various models to generate a functional pair of novelty New Year's Eve glasses on a person, that don't just plonk the year onto the top of regular frames.
Thanks. That's a good one~ Lens type stuff that involves reflections/refraction is a neat challenge for generative models. I did some editing tests that involved replacing an apartment window with a mirror back when Nano-Banana Pro was released and was rather stunned by the results.
That's great, though I wasn't even thinking at the scale of reflection or refraction. My test was if the image generators could come up with a novelty pair of glasses that incorporate the year digits into the shape of the frame itself with some whimsy, rather than just plop the numbers on top of regular boring frames. So something like [this](https://p.kagi.com/proxy/oardefault.jpg?c=-4THVYblKrsgkzFTNE...) rather than [this](https://p.kagi.com/proxy/2026-Glasses-4-Color-New-Year-Glass...). A lot of the initial designs just incorporated the numbers into the frames, with no consideration for relative placement to the eyes, completely obscuring vision. Additional prompting might lead to cutouts for the eyes, but that was unsatisfying. At least as of this past new year's eve, I couldn't get any of the image generators to give me something even passible. Images 2.0 also couldn't give anything acceptable till I gave it some examples.
Oh, I see what you’re saying. I like these types of tests where you incorporate well-known objects from the training data into unusual geometries.
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.
Flounder might even work, since my initial complaints that the generated designs obscured the wearer's eyesight were met with solutions that just moved the offending eye to the side of the person's head :)
Such a fun site, thank you! I was surprised that Seedream4 passed the mermaid test since it's hard to tell whether they are in the water or submerged, and the mermaid has something funny going on with her left hand.
Yeah seedream's attempt does have a bit of an uncanny valley effect: the mermaid/dolphin are only partially submerged, but there’s water above them with sunlight reflecting on the surface, and the mermaid’s hand looks disconnected from the angle of her arm.
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.
Why does Gemini 3.1 get a pass for the same reasons they got image 2 gets a fail on the flat earth one? Gemini has all sorts of random body parts and limbs etc.
That's a mistake~ None of the models successfully passed the Flat Earth composition test. I've updated the passing criteria to be more explicit as well. Thanks for catching that!
So the prompts are tuned and adjusted on a per-model basis. If you look at the number of attempts, each receives a specific prompt variation depending on the model. This honestly isn't as much of an issue these days because SOTA models natural language parsing (particularly the multimodal ones) has eliminated a lot of the byzantine syntax requirements of the SD/SDXL days.
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 goal isn’t the prompt itself. The test is whether a prompt can be expressed in such a way that we still arrive at the author's intent, and of course to do so in a way that isn't unnatural.
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.
That being said, gpt-image-1.5 was a big leap in visual quality for OpenAI and eliminated most of the classic issues of its predecessor, including things like the “piss filter.”
I’ll update this comment once I’ve finished running gpt-image-2 through both the generative and editing comparison charts on GenAI Showdown.
Since the advent of NB, I’ve had to ratchet up the difficulty of the prompts especially in the text-to-image section. The best models now score around 70%, successfully completing 11 out of 15 prompts.
For reference, here’s a comparison of ByteDance, Google, and OpenAI on editing performance:
https://genai-showdown.specr.net/image-editing?models=nbp3,s...
And here’s the same comparison for generative performance:
https://genai-showdown.specr.net/?models=s4,nbp3,g15
UPDATES:
gpt-image-2 has already managed to overcome one of the so‑called “model killers” on the test suite: the nine-pointed star.
Results are in for the generative (text to image) capabilities: Gpt-image-2 scored 12 out of 15 on the text-to-image benchmark, edging out the previous best models by a single point. It still fails on the following prompts:
- A photo of a brightly colored coral snake but with the bands of color red, blue, green, purple, and yellow repeated in that exact order.
- A twenty-sided die (D20) with the first twenty prime numbers (2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61, 67, 71) on the faces.
- A flat earth-like planet which resembles a flat disc is overpopulated with people. The people are densely packed together such that they are spilling over the edges of the planet. Cheap "coastal" real estate property available.
All Models:
https://genai-showdown.specr.net
Just Gpt-Image-1.5, Gpt-Image-2, Nano-Banana 2, and Seedream 4.0
https://genai-showdown.specr.net?models=s4,nbp3,g15,g2