Yeah, I've been mystified for years why they can't just train the "AI look" out of images. Is it, like, something undetectable by machine? But then how come we can often detect it so easily?
I think this is fundamental. People will probably dislike whatever it is that AI images look like, because whatever they look like is going to be associated with cheap. They could have a totally different aesthetic and I think most people's conception of cheap would adapt to whatever that is.
It's the opposite: the average person likes the AI look of images (the more "aesthetic" look) and RLHF keep converging toward that.
It is entirely possible to prompt engineer away from that style, but that requires effort and knowledge that said average person does not have, so it's stuck in a feedback loop.
Using the grid trick for animations to generate the individual frames has been around for awhile and not specific to ChatGPT: I'm surprised it took this long for one of the official providers to make it explicit. However, adherence between frames is generally tricky, and even in the examples there is a lot of noise between frames although that could be due to compression.
SVG support is interesting but I suspect shenanigans are afoot since that's not how image generation models work, and it would likely strip any SynthID watermarks. It's possible OpenAI has pipelined a raster-to-vector tool that does the conversion automatically. (I developed a similar pipeline and it works extremely well; I'll open source it at some point)
EDIT: Per documentation, there is an explicit `output_format: "svg"`, and I am indeed correct there is raster-to-vector in play:
> The API creates a reference image and uses it to produce the SVG. The quality setting controls that reference image. Decode data[0].b64_json as base64 and save the resulting UTF-8 XML bytes with a .svg extension.
OpenAI claims 50% faster generation with Images 2.6 but curiously price is the same as Image 2.5.
I would caution against using the webapp to test the quality of this new model because OpenAI does not indicate which image model and effort is being used to generate the image, to my continued frustration.
ChatGPT Images (and some others like Nano Banana) are autoregressive models, and the text model generates the tokens in the same way which are then passed in the correct form for an image model to decode. That is how it can reason the prime numbers + Pokemon + art style conditional.
It is different from earlier image models which just use a simple text encoder and struggle with conditional logic if not explicitly trained for it.
That prompt seems almost deliberately obtuse. The natural phrasing would be “greying blonde hair”. If you described someone as having half-X, half-Y hair, I’d expect to see something like that image, too.
Decided to try out one of the new features in GPT Image 2.6: the ability to natively generate animated GIF files.
I didn't have high expectations, and I still think you can probably get better results by doing a First-Last-Frame using something like Minimax H3 but I was still rather surprised that it could do it at all.
Here are some of the results. The last one shows its ability to take an existing image and convert it into an animated GIF. It's a picture of my husky inside a snow globe.
That's a bit hyperbolic. The windmill one is honestly not terrible (slight flicker on the door).
Honestly, you have to think about the use case. GIFs are going to be little throwaway things that you're pasting into Discord or meme-ing.
Nobody's reaching for a GIF if they're looking for professional results.
It's a shame that the SVG output seems to be limited to static vector graphics only though - or rather I couldn't get it to produce anything with the <animate> attribute.
Everyone is highly sensitive to slop, and it still looks like slop.
If you actually want something that looks decent, it feels like tool use would at least help cover the obvious blunders. Building layers to composite would go a long way to removing some slop here.
Did a considerably worse job than Image Gen 2.0 on my standard test prompt.
"A macro close-up photograph of an old watchmaker's hands carefully replacing a tiny gear inside a vintage pocket watch. The watch mechanism is partially submerged in a shallow dish of clear water, causing visible refraction and light caustics across the brass gears. A single drop of water is falling from a pair of steel tweezers, captured mid-splash on the water's surface. Reflect the watchmaker's face, slightly distorted, in the curved glass of the watch face. Sharp focus throughout, natural window lighting from the left, shot on 100mm macro lens."
On a phone, if you share a gif, it creates a link to ChatGPT.com with a static image. You can also save the image to photos and share directly from there which feels like an unnecessary step.
It feels like a sleazy pattern and defeats the purpose of why I’d want to generate a gif in the first place, which to make clear is likely going to be to share with friends in a chat. I’m not interested in driving traffic to ChatGPT.com to impress my friends.
That stained glass whale thing makes me especially grumpy. When it scrolled onto the page, my first thought was "whoah, what a neat idea" and then it was _immediately_ followed by "what a terribly execution."
Like, I'm not even an artist, so I'm sure there's a million things I'm missing, but it's just so upsetting to look at. There's weird random reflections on things that shouldn't have reflections, the coloured light absolutely doesn't actually track through the stained glass onto any surface that makes any sense, light is coming from every which way, it's just... gross to look at for anything more than a quarter of a second.
And this is what they're using to hype people up for it. /shrug
The input and output pricing per token remains the same between 2.5 and 2.6, although using their image pricing calculator shows a quadrupling of token usage at low and medium quality settings for 2.6. For e.g. at 1024x1024:
2.5 low, 196 tokens, $0.006
2.6 low, 823 tokens, $0.025
2.5 med, 439 tokens, $0.013
2.6 med, 1449 tokens, $0.043
Interestingly, higher quality settings show no change in token usage between the models.
Similar results for my workflow. And the Round art comes out significantly less round without prompt changes (more input tokens)
Setting (1024x1024) Tokens Cost/image Time
2.5 medium 439 $0.014 12.0 s
2.6 low 824 $0.025 6.7 s
2.6 medium 1,449 $0.044 7.9 s
2.6 high 1,758 $0.053 8.9 s
I've been playing with it for the last hour or so. I immediately updated my platform to use both since 2.5 is my primary model (i run a prompt->design site).
After playing with it for an hour, I'm rolling back 2.6-flare to 2.5-flare. I've found 2.6-flare specifically to either be too simplistic in its approaches, or create outputs that are inherently broken in other cases.
I really don't like this generative AI. I may look at it like an artist but completed generated software, writing, music and imagery is a robbery of human worth. Absolutely.
No, it's created by data scanned from the internet, books and fed to the models. How is that a good thing for anyone? I'd love your take. People spend years giving themselves value to others to live a life of worth and value.
The data just scanned itself and organized itself into an neural network? Would be an awesome display of bootstrapped intelligence if that happened! Sadly, it was the result of decades of human research.
What's the state of the art for image gen? I'm far removed and pretty happy with the free Gemini tier but its notable to me that I rarely see boasting about being "the best image generator yet" and I wonder if that's because they're all just good enough?
It has to be Google, since OpenAI always compares to other companies when it feels its winning at some benchmarks. They didn't even compare on cost, speed, or any benchmarks at all.
I've been running a comparative image benchmark site for years now (see my profile for the link), ever since SDXL back in 2023, and the answer really comes down to it depends. My benchmark focuses primarily on prompt adherence, but that's just one metric you can use. (fidelity, censorship, explorative, speed, features, etc are others)
In terms of overall state of the art, the two major contenders are probably Nano Banana Pro (Google) and the GPT Image 2 series (OpenAI).
For pure prompt adherency, OpenAI has consistently lead the pack since the original gpt-image-1 released back in March of 2025.... but its OpenAI's fidelity has always left something to be desired. Version 2 of the series has definitely closed the gap a bit.
I wonder if there's some weird accounting going on, or if this is real. It would mean around 24M images an hour and that's in the area of "more images than anyone will see". It means there's lots of wasted automated effort in this process somewhere, even if we assume some creators will request 8x variants of their ideas.
I don't use AI image generation anymore, but when I was fiddling with it I would generate dozens of images per idea, limited only by my patience. "The same thing but a little different" is the easiest feedback you can give anything.
Same, when I messed around with it to learn how the models worked I would generate batches of hundreds or thousands with slightly tweaked prompts and negative prompts to get a sense of how the inputs changed the outputs.
If you define pixel art as just a vague overall look, then yes, it can do pixel art. If you define pixel art as being limited to a fixed color palette, like 16 or 256 colors and aligned to a square grid, then no, not without some manual post-processing.
It's not helped by the fact that a lot of people who are not artistically inclined will generate pixel art on these image generators at 1MP resolutions (far higher than traditional pixel art), such that even the casual observer can typically see the bad smearing and fringing in the final result.
That depends on your definition of pixel art. (people are exceptionally pedantic about this)
With the right prompt engineering, it (and similar quality models) can indeed do relevant pixel art, but due to the existence of blotchiness/SynthID watermarks you'll need to do some sort of postprocessing regardless.
There are SDXL pixel art models out there that do actual pixel art. They generate images that are 8x too large, you need to downscale them and reduce colors externally.
yeah, the original Retro Diffusion models were an early example of that - IIRC it had models separately fine-tuned for 32x32, 64x64, etc. so you'd nearest neighbor squash'em down to that size after the initial gen to kill the high frequency noise.
It looks like the images in there are 'based' on the "hot springs symbol" - a Japanese map symbol encoded into Unicode that seems to have become synonymous with "piping hot bread" in the West!
The images in the GPs link show a common pictogram for artisanal bread. The image itself is actually a close representation of the real thing. An image search for sourdough will return many images showing bread with a similar appearance.
Anecdotally, I’ve never seen that hot springs symbol used for bread, and I’ve lived in the west my whole life.
Not without irony, but connecting the bread symbol with the hot springs symbol feels like the sort of mistake an llm would make.
It is entirely possible to prompt engineer away from that style, but that requires effort and knowledge that said average person does not have, so it's stuck in a feedback loop.
This kind of content is here to stay and we’re going to see more and more of it. Unfortunately…
SVG support is interesting but I suspect shenanigans are afoot since that's not how image generation models work, and it would likely strip any SynthID watermarks. It's possible OpenAI has pipelined a raster-to-vector tool that does the conversion automatically. (I developed a similar pipeline and it works extremely well; I'll open source it at some point)
EDIT: Per documentation, there is an explicit `output_format: "svg"`, and I am indeed correct there is raster-to-vector in play:
> The API creates a reference image and uses it to produce the SVG. The quality setting controls that reference image. Decode data[0].b64_json as base64 and save the resulting UTF-8 XML bytes with a .svg extension.
OpenAI claims 50% faster generation with Images 2.6 but curiously price is the same as Image 2.5.
I guess I should have been careful about “half”?
These models have a hard time imagining men with naturally blond hair getting grey.
That said, I gave it one of my more difficult image prompts and the model now surprisingly passes (without plagiaristic cheating): https://chatgpt.com/share/6aebcbcd-6b8c-83e8-b52b-dbedf20ce4...
When I did a similar prompt with ChatGPT 2.0 it was worse, so interesting to see improvement.
It is different from earlier image models which just use a simple text encoder and struggle with conditional logic if not explicitly trained for it.
I didn't have high expectations, and I still think you can probably get better results by doing a First-Last-Frame using something like Minimax H3 but I was still rather surprised that it could do it at all.
Here are some of the results. The last one shows its ability to take an existing image and convert it into an animated GIF. It's a picture of my husky inside a snow globe.
https://imgpb.com/OzSMKJ
Honestly, you have to think about the use case. GIFs are going to be little throwaway things that you're pasting into Discord or meme-ing.
Nobody's reaching for a GIF if they're looking for professional results.
It's a shame that the SVG output seems to be limited to static vector graphics only though - or rather I couldn't get it to produce anything with the <animate> attribute.
If you actually want something that looks decent, it feels like tool use would at least help cover the obvious blunders. Building layers to composite would go a long way to removing some slop here.
Ain’t wrong buuuuut…
But trust, I'm a have a field day with that.
"A macro close-up photograph of an old watchmaker's hands carefully replacing a tiny gear inside a vintage pocket watch. The watch mechanism is partially submerged in a shallow dish of clear water, causing visible refraction and light caustics across the brass gears. A single drop of water is falling from a pair of steel tweezers, captured mid-splash on the water's surface. Reflect the watchmaker's face, slightly distorted, in the curved glass of the watch face. Sharp focus throughout, natural window lighting from the left, shot on 100mm macro lens."
2.6: https://gov.f9.is/JRDpHJGu
2.0: https://gov.f9.is/cqyCwtq9
It feels like a sleazy pattern and defeats the purpose of why I’d want to generate a gif in the first place, which to make clear is likely going to be to share with friends in a chat. I’m not interested in driving traffic to ChatGPT.com to impress my friends.
Like, I'm not even an artist, so I'm sure there's a million things I'm missing, but it's just so upsetting to look at. There's weird random reflections on things that shouldn't have reflections, the coloured light absolutely doesn't actually track through the stained glass onto any surface that makes any sense, light is coming from every which way, it's just... gross to look at for anything more than a quarter of a second.
And this is what they're using to hype people up for it. /shrug
It's an image depicting a whale made of stained glass; realism is off the table there.
Just an absolute race to post a contrarian take to hit the top of the comments.
Never change bros
2.5 low, 196 tokens, $0.006
2.6 low, 823 tokens, $0.025
2.5 med, 439 tokens, $0.013
2.6 med, 1449 tokens, $0.043
Interestingly, higher quality settings show no change in token usage between the models.
I've been playing with it for the last hour or so. I immediately updated my platform to use both since 2.5 is my primary model (i run a prompt->design site).
After playing with it for an hour, I'm rolling back 2.6-flare to 2.5-flare. I've found 2.6-flare specifically to either be too simplistic in its approaches, or create outputs that are inherently broken in other cases.
Comparison examples: https://diffui.ai/app/canvas/178db5d5-ff87-4b43-aa98-2897773...
2.6-sunburst seems solid however.
Update: seems to be confirmed on arena.ai - 2.5 scores higher than 2.6 https://arena.ai/leaderboard/image/text-to-image
This may be why text elements in particular appear more "zoomed in", and why it's providing less detail for UI design.
In two of the examples the images from 2.6 are inherently broken, ie: https://image.non.io/161d8f02-fc2a-40e0-91b2-9d27ceb8fef1.we...
Objectively there are tests here where 2.5 are better. Subjectively prefering the layout of 2.6 is fine, but there are blocking flaws.
> Images 2.6 is available globally today in ChatGPT and Codex across all subscription tiers on desktop, mobile, and web.
But I just updated my codex-cli to the latest version (still 0.166.0) and it is still unable to generate gifs. Very disappointing
In terms of overall state of the art, the two major contenders are probably Nano Banana Pro (Google) and the GPT Image 2 series (OpenAI).
For pure prompt adherency, OpenAI has consistently lead the pack since the original gpt-image-1 released back in March of 2025.... but its OpenAI's fidelity has always left something to be desired. Version 2 of the series has definitely closed the gap a bit.
It's not helped by the fact that a lot of people who are not artistically inclined will generate pixel art on these image generators at 1MP resolutions (far higher than traditional pixel art), such that even the casual observer can typically see the bad smearing and fringing in the final result.
With the right prompt engineering, it (and similar quality models) can indeed do relevant pixel art, but due to the existence of blotchiness/SynthID watermarks you'll need to do some sort of postprocessing regardless.
https://emojipedia.org/hot-springs
SVG is cool though, I wonder how good it is.
> Create a GIF where xkcd 2347 gets hit by an airliner like the twin towers and the pilot is AI
and got a refusal.
> We’re so sorry, but the image we created may violate our guardrails around violence.
Follow by
“Make it”
Seems to resolve the refusals
https://solyin.com/dependency-ai.gif