ChatGPT Images 2.0

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:

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

Price comparison:

GPT Image 2

  Low     : 1024×1024 $0.006 | 1024×1536 $0.005 | 1536×1024 $0.005

  Medium  : 1024×1024 $0.053 | 1024×1536 $0.041 | 1536×1024 $0.041

  High    : 1024×1024 $0.211 | 1024×1536 $0.165 | 1536×1024 $0.165

GPT Image 1

  Low     : 1024×1024 $0.011 | 1024×1536 $0.016 | 1536×1024 $0.016

  Medium  : 1024×1024 $0.042 | 1024×1536 $0.063 | 1536×1024 $0.063

  High    : 1024×1024 $0.167 | 1024×1536 $0.25  | 1536×1024 $0.25

Genuine question: what positive use cases are sufficient to accept the harm from image generators?

One that i can think of:

- replacing photography of people who may be unable to consent or for whom it may be traumatic to revisit photographs and suitable models may not be available, e.g. dementia patients, babies, examples of medical conditions.

Most other vaguely positive use cases boil down to "look what image generators can do", with very little "here's how image generators are necessary for society.

On the flip side, there are hundreds of ways that these tools cause genuine harm, not just to individuals but to entire systems.

I know people like to dunk on ChatGPT and Gemini and say Claude is or used to be better, but you can still use worse models when you're out of usage AND make use of Nano Banana and and ChatGPT Image generation with separate limits for your subscription. I think it could make it a more package as a whole for some people (non-programmers). I do like having the option and am excited for which improvements they've done to ChatGPT Image generation because in the past it had this yellow piss filter and 1.5 it sort of fixed it but made things really generic with Nano Banana beating it (altough Gemini also had a too aggressively tuned racial bias which they fixed), it seems the images ChatGPT generates have gotten better.

The quality of the text is really impressive and I can’t seem to see any artefacts at all. The fake desktop is particularly good: Nano Banana would definitely slip up with at least a few bits of the background.

System card link with safety details https://deploymentsafety.openai.com/chatgpt-images-2-0

direct pdf https://deploymentsafety.openai.com/chatgpt-images-2-0/chatg...

do they have anything similar to SynthID, or are they just pretending that problem doesn't exist?

I know this is probably mega cherry-picked to look more impressive, but some of the images are terrifyingly realistic. They seem to have put a lot of effort into the lighting.

The image of the messy desktop with the ASCII art is so impressive - the text renders, the date is consistent, it actually generated ASCII art in "ChatGPT", etc. I was skeptical that it was cherry-picked but was able to generate something very similar and then edit particular parts on the desktop (i.e. fixing content in the browser window and making the ASCII dog "more dog like"). It's honestly astounding, to me at least.

We were afraid it would be Skynet and instead we got the ultimate meme generator !

Ok, I can hear the sound of entire industries crumbling right now.

Every time a new image gen comes out I keep saying that it won't get better just to be surprised again and again. Some of the examples are incredible (and incredibly scary. I feel like this is truly the point where understanding if something is AI becomes impossible)

Model card for the API endpoint gpt-image-2 (which may or may not reflect the output from ChatGPT Images 2): https://developers.openai.com/api/docs/models/gpt-image-2

API Pricing is mostly unchanged from gpt-image-1.5, the output price is slightly lower: https://developers.openai.com/api/docs/pricing

...buuuuuuuuut the price per image has changed. For a high quality image generation the 1024x1024 price has increased? That doesn't make sense that a 1024x1024 is cheaper than a 1024x1536, so assuming a typo: https://developers.openai.com/api/docs/guides/image-generati...

The submitted page is annoyingly uninformative, but from the livestream it proports the same exact features as Gemini's Nano Banana Pro. I'll run it through my tests once I figure out how to access it.

It's great. Also doesn't seem to have any "slop" standard look, the images it produces are quite diverse.

I would imagine this will hit illustrators / graphics designers / similar people very hard, now that anyone can just generate professional looking graphical content for pennies on the dollar.

The page keeps crashing on my iPhone 17 Pro.

fal has the endpoint under openai/gpt-image-2

Fake layouts, fake handwritten kid story, fake drunk photos? All from training on real things people did.

As with anything AI, we are not ready for the scale of impact. And for what? Like, why are you proud of this?