Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

ChatGPT Images 2.0

openai.com

Post reply on HN
Print view View original

81–90 of 1001 posts

Re: ChatGPT Images 2.0

#81

Post by throw310822 »

lehmacdj wrote: post #49
thevinter wrote: post #19

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)

So do you think there will be a better image model in a year?

I'll bite: no I don't think so. If the examples are not cherry-picked and by "image model" we mean just the ability to generate pictures, this looks like parity with human excellence, there isn't much space for further improvement. The images don't just look real, they look tasteful- the model is not just generating a credible image, it's generating one that shows the talent of a good photographer/ designer/ artist.

Re: ChatGPT Images 2.0

#82

Post by thevinter »

Vachyas wrote: post #77
lehmacdj wrote: post #49

Earlier quoted context omitted.

So do you think there will be a better image model in a year?

I'm honestly unsure what could be improved at this point. Consistency? So it fails less often? Based on the released images, (especially the one "screenshot" of the Mac desktop) I feel like the best images from this model are so visually flawless that the only way to tell they're fake is by reasoning about the content of the image itself (ex. "Apple never made a red iPhone 15, so this image is probably fake" or "Cost…

There is definitely room for improvement: https://gist.github.com/simonw/88eecc65698a725d8a9c1c918478a...

Especially when it comes to detailed outputs or non-standard prompts.

I do believe it will get even better - not sure it will happen within a year but I wouldn't be incredibly surprised if it did.

Re: ChatGPT Images 2.0

#83

Post by vunderba »

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

Re: ChatGPT Images 2.0

#84

Post by cwillu »

makira wrote: post #61
simonw wrote: post #37

Earlier quoted context omitted.

I just got a much better version using this command instead, which uses the maximum image size according to https://github.com/openai/openai-cookbook/blob/main/examples... OPENAI_API_KEY="$(llm keys get openai)" \ uv run 'https://raw.githubusercontent.com/simonw/tools/refs/heads/main/python/openai_image.py' \ -m gpt-image-2 \ "Do a where's Waldo style image but it's where is the raccoon holding a ham radio" \ --quali…

Fed into a clear Claude Code max effort session with : "Inspect waldo2.png, and give me the pixel location of a raccoon holding a ham radio.". It sliced the image into small sections and gave: "Found the raccoon holding a ham radio in waldo2.png (3840×2160). - Raccoon center: roughly (460, 1680) - Ham radio (walkie-talkie) center: roughly (505, 1650) — antenna tip around (510, 1585) - Bounding box (raccoon + radio):…

I had one problem: finding the raccoon. Now I have two: finding the red-and-white striped souvenir umbrella, and finding the raccoon.

Re: ChatGPT Images 2.0

#85

Post by makira »

cwillu wrote: post #84
makira wrote: post #61

Earlier quoted context omitted.

Fed into a clear Claude Code max effort session with : "Inspect waldo2.png, and give me the pixel location of a raccoon holding a ham radio.". It sliced the image into small sections and gave: "Found the raccoon holding a ham radio in waldo2.png (3840×2160). - Raccoon center: roughly (460, 1680) - Ham radio (walkie-talkie) center: roughly (505, 1650) — antenna tip around (510, 1585) - Bounding box (raccoon + radio):…

I had one problem: finding the raccoon. Now I have two: finding the red-and-white striped souvenir umbrella, and finding the raccoon.

simonw posted 2 different images: make sure to look at the second one.

Re: ChatGPT Images 2.0

#86

Post by cwillu »

makira wrote: post #85
cwillu wrote: post #84

Earlier quoted context omitted.

I had one problem: finding the raccoon. Now I have two: finding the red-and-white striped souvenir umbrella, and finding the raccoon.

simonw posted 2 different images: make sure to look at the second one.

Yeah, I noticed that just now, but too late to delete the comment :p

Re: ChatGPT Images 2.0

#87

Post by vunderba »

thevinter wrote: post #82
Vachyas wrote: post #77

Earlier quoted context omitted.

I'm honestly unsure what could be improved at this point. Consistency? So it fails less often? Based on the released images, (especially the one "screenshot" of the Mac desktop) I feel like the best images from this model are so visually flawless that the only way to tell they're fake is by reasoning about the content of the image itself (ex. "Apple never made a red iPhone 15, so this image is probably fake" or "Cost…

There is definitely room for improvement: https://gist.github.com/simonw/88eecc65698a725d8a9c1c918478a... Especially when it comes to detailed outputs or non-standard prompts. I do believe it will get even better - not sure it will happen within a year but I wouldn't be incredibly surprised if it did.

Yep. “Where’s Waldo” has been a classic challenge for generative models for a while because it requires understanding the entire concept (there’s only one Waldo), while also holding up to scrutiny when you examine any individual, ordinary figure.

I experimented with the concept of procedural generation of Waldo-style scavenger images with Flux models with rather disappointing results. (unsurprisingly).

Re: ChatGPT Images 2.0

#88

Post by Flere-Imsaho »

toraway wrote: post #69
Flere-Imsaho wrote: post #42

Earlier quoted context omitted.

I wake up everyday, read the tech news, and usually see some step change in AI or whatever. It's wild to think I'm living through such a massive transformation in my lifetime. The future of tech is going to be so different from when I was born (1980), I guess this is how people born in 1900 felt when they got to see man land on the moon? > Wow, the difference between AI and non-AI images collapses. I hate the future…

"Solved" seems a tad overstated if you scroll up to Simonw's Where's Waldo test with deformed faces plus a confabulated target when prompted for an edit to highlight the hidden character with an arrow.

It's "solved" in that we have a way forward to reduce the errors down to 0.00001% (a number I just made up). Throwing more compute/time/money at these problems seems to reduce that error number.

Re: ChatGPT Images 2.0

#89

Post by throw310822 »

thevinter wrote: post #82
Vachyas wrote: post #77

Earlier quoted context omitted.

I'm honestly unsure what could be improved at this point. Consistency? So it fails less often? Based on the released images, (especially the one "screenshot" of the Mac desktop) I feel like the best images from this model are so visually flawless that the only way to tell they're fake is by reasoning about the content of the image itself (ex. "Apple never made a red iPhone 15, so this image is probably fake" or "Cost…

There is definitely room for improvement: https://gist.github.com/simonw/88eecc65698a725d8a9c1c918478a... Especially when it comes to detailed outputs or non-standard prompts. I do believe it will get even better - not sure it will happen within a year but I wouldn't be incredibly surprised if it did.

I wonder if at this point you could just ask the agent to iteratively refine the image in smaller portions.

Re: ChatGPT Images 2.0

#90

Post by vunderba »

makira wrote: post #18
simonw wrote: post #15

I've been trying out the new model like this: OPENAI_API_KEY="$(llm keys get openai)" \ uv run https://tools.simonwillison.net/python/openai_image.py \ -m gpt-image-2 \ "Do a where's Waldo style image but it's where is the raccoon holding a ham radio" Code here: https://github.com/simonw/tools/blob/main/python/openai_imag... Here's what I got from that prompt. I do not think it included a raccoon holding a ham radio…

> though the problem with Where's Waldo tests is that I don't have the patience to solve them for sure I see an opportunity for a new AI test!

There have already been several attempts to procedurally generate Where’s Waldo? style images since the early Stable Diffusion days, including experiments that used a YOLO filter on each face and then processed them with ADetailer.

It's a difficult test for genai to pass. As I mentioned in a different thread, it requires a holistic understanding (in that there can only be one Waldo Highlander style), while also holding up to scrutiny when you examine any individual, ordinary figure.

Post reply on HN