Originally posted at https://openai.com/index/introducing-chatgpt-images-2-0/
Images are a language, not decoration. A good image does what a good sentence does—it selects, arranges, and reveals. It can explain a mechanism, stage a mood, test an idea, or make an argument.
A year ago, we released ChatGPT Images, showing that images created by AI can be both beautiful and useful. ChatGPT Images 2.0 is the next step: a state-of-the-art model that can take on complex visual tasks and produce precise, immediately usable visuals.
This model is a step change in detailed instruction following, placing and relating objects accurately, and rendering dense text, with the ability to generate across aspect ratios. Its sense of composition and visual taste means results feel less AI-generated and more intentionally designed. It’s accurate across languages and uses its expanded visual and world knowledge to fill in the gaps for you, so you get smarter images with less prompting.
To extend the model’s capabilities for the most complex tasks, Images 2.0 is our first image model with thinking capabilities. When a thinking or pro model is selected in ChatGPT, Images 2.0 can search the web for real-time information, create multiple distinct images from one prompt, and double-check its own outputs. With thinking, the model can take on even more of the heavy lifting between idea and image, especially when accuracy, up-to-date information, consistency, and visual cohesion matter most.
With both the intelligence of OpenAI’s reasoning models and a vast understanding of the visual world, this model moves image generation from rendering to strategic design, from a tool to a visual system, helping people turn ideas into outputs they can understand, share, teach with, and build from. It’s available starting today to all users in ChatGPT, Codex, and the API.
Greater precision and control
Images 2.0 brings an unprecedented level of specificity and fidelity to image creation. It can not only conceptualize more sophisticated images, it actually brings that vision to life effectively, able to follow instructions, preserve requested details, and render the fine-grained elements that often break image models: small text, iconography, UI elements, dense compositions, and subtle stylistic constraints, and at up to 2K resolution in the API. Instead of getting something vaguely in the neighborhood of what you meant, you get something you can actually use.
Stronger across languages
To date, our image generation models have been more consistent in English and other Latin-script languages, but less precise beyond that, especially when text was complex or dense.
Images 2.0 moves beyond that barrier with stronger multilingual understanding and significant gains in non-Latin text rendering, particularly in Japanese, Korean, Chinese, Hindi, and Bengali. It can produce images with non-English text that’s not only rendered correctly but with language that flows coherently.
That includes not just translating a label or two, but generating visually coherent outputs where language is part of the design itself, from posters and explainers to diagrams and comics. This makes the model more globally useful and helps people create visuals that work in the languages they actually use.
Stylistic sophistication and realism
Images 2.0 also shows significantly improved fidelity across a wide range of visual styles. It is better able to capture the defining characteristics of photos—including the tiny flaws that add realism—as well as cinematic stills, pixel art, manga, and other distinctive visual languages, with greater consistency in texture, lighting, composition, and fine detail.
As a result, the model can produce outputs that more faithfully reflect the style requested, rather than approximating it. This is especially useful for game prototyping, storyboarding, marketing creative, and creating assets in a particular medium or genre.
Flexible aspect ratios
The new model also gives you more flexibility in how those images are delivered. With support for aspect ratios as wide as 3:1 and as tall as 1:3, Images 2.0 can generate outputs that are ready to fit the formats you need, from wide banners and presentation slides to posters, mobile screens, bookmarks, and social graphics. Ask for the aspect ratio you want in the prompt, or select from preset options to regenerate any image in new dimensions.
Real-world intelligence
Images 2.0 brings a more up-to-date understanding of the world into image creation, with a knowledge cutoff of December 2025, for more relevant and contextually accurate outputs. This is especially important for artifacts like explainers, educational graphics, and visual summaries, where correctness and clarity matter just as much as aesthetics.
Its intelligence allows it to expertly handle tasks end-to-end: synthesizing information, writing the story, and laying it out with clean structure, intentional whitespace, and strong visual flow.
A visual thought partner
When a thinking model is selected in ChatGPT, the model takes more time and does more agentically behind the scenes to thoroughly understand and execute the task. It can use the web to find relevant information, transform uploaded materials into clear visual explainers, and reason through the structure of the image before generating. In this mode, Images 2.0 acts more like a visual thought partner, helping carry a project from rough concept to finished asset with significantly less work on your part.
With thinking, it can also produce multiple distinct images at once, a first for image generation in ChatGPT. That opens up workflows that were previously cumbersome: a sequence of manga pages, a set of redesign directions for every room in a house, a family of poster concepts, or a collection of social graphics in different aspect ratios and languages.
Instead of prompting one image at a time and stitching the project together yourself, you can ask for a coherent set of up to eight outputs in one go with character and object continuity, that sequentially build on one another.
Using image generation in Codex
Images in Codex brings visual creation into one workspace for creating, iterating, and shipping apps, slide decks, and other work, making Codex more useful for broader tasks across design, marketing, product, sales, and learning & development.
For example, you can generate multiple UI directions, concepts, and prototypes, compare options quickly, and then turn the strongest ideas into live products or website experiences without leaving the Codex app. You can create images in Codex with your ChatGPT subscription without creating a separate API key.
Description
Comments (68)
there should be more option for aspect ratio yea?
Yes the model can do up to 4K (Experimental) but is very flexible
From their site:
Maximum edge length must be less than 3840px
Both edges must be a multiple of 16
Ratio between the long edge and short edge must not be greater than 3:1
Total pixels must not exceed 8,294,400
Total pixels must not be less than 655,360
https://developers.openai.com/cookbook/examples/multimodal/image-gen-models-prompting-guide
I can't find ChatGPT Images 2.0 when I search for it, i only discovered it from a image, i need to link the images i made from chatgpt to this model, can you fix this please?
Finally understands gay horse marriage instead of just making straight horse marriage and labeling it gay horse marriage. Maybe with GPT Image 3.0, OpenAI will understand 16:9 as an aspect ratio 🤞
in chatgpt, there are more option of aspect ratio including 16:9, civit need to include it, its available
@Fujido good to know they finally added that option, thanks!
It should have these AR, even on replicate, the new API endpoint is not fully set up. But looking at the specs, it has more than just AR, it has higher resolution too.
generated 40+ images. Good with almost everything. Text is exceptional. Surreal is low. the Article https://civitai.red/articles/28986/chatgpt-images-20-test sometimes GPT 1.5 is better.
Some tips I have learned in the last few days:
Using shorter prompts and not using SDXL tags can improve the image quality and reduce the noise from this model.
Using Medium quality on Civitai is a lot faster, cheaper and doesn't degrade the quality that much.
If you are getting yellow images/art the prompt "give it a neutral hue and white balance" helps a lot.
If you want a less detailed , less photo real and noisy or more anime style image the prompt "make it smooth" helps.
Didn't found this model...because it is no model. Name is gpt-image-2
Maybe rename this model page name?
good
Time to update your api calls. GPT image 2 supports up to 3840 x 2160.
"supports variable aspect ratios and higher resolutions, moving beyond the fixed sizes of previous models to offer native 2K and 4K outputs, supporting resolutions up to 3840×2160 pixels"
>generation only again
I WANNA DOWNLOAD IT FFS - (RTX 2060 - don't be jelly)
the GGUF version is comming next week ! They said it will work even on Android !
Overall, this is a good model. I was wrong; the main problem isn't the brown filter in the image.
It's the light grey, dark grey, and gray filters.
I have found adding "give it a neutral hue and white balance" to the prompt can really help with this issue.
I tried your advice, and it sometimes helps. The new ChatGPT Images model is much better at understanding concepts and scene construction, but I still don't like its color correction.
grey filters? in my experience, the is a side effect, when you mix different images, (reference images). the only way I can get rid of it, is by postprocessing the images which are worth the effort.
2.0 feels utterly fantastic, especially compared to 1.5.
Whereas I found 1.5 to be a side grade, and honestly slight downgrade in comparison to late 1.0 in terms of style adherence and really had to employ lots of trickery to make it work for me (though it still very much did) 2 just feels like a real step up— if you're using it in-app and have a very fine tuned GPT with your own styles, characters, etc. I think it's just just a clear upgrade in every way.
It's a little buggy, as most of the latest OAI projects are (likely due to rushed crunch)— but as they've showed in the last decade they're EXTREMELY receptive to feedback so until they prove otherwise I fully expect to see those issues ironed out.
Cult of Gremmy approved.
Big improvement from 1.5. However it inconsistently refuse to generate image based on its own filters. The randomness is confusing. This result in long wait time, multiple attempts and you still don't know what to change to get results.
It's hilarious to me that this thing made in the 'land of the free' is more censored (by a considerable stretch) than the model from China (Qwen) lmao.
& it's overzealous to the point of idiotic! You can't ask for an image of a woman in a bikini on a beach (like you couldn't just go out & see this sort of 'obscenity' at any random public swimming pool lol.
Literally the other day my prompt/image was rejected because it had the word "Sensual" in it..... I remove it, & it works again. The context wasn't even sexual (neither was any of the rest of the prompt), but the whole thing got canned just because of one innocuous word.
@ShadowCell Yeah I definitely agree. The more I use it the more I feel the censorship. I never knew they block the gen because of just one word. I thought Openai's moderation system to be more nuanced and sophisticated but I guess not. The funny part is sometimes I just retry the same prompt and it goes through.
The blocking is done by CivitAI, not OpenAI, if you generate the images here. I already have written my own app, which uses the Replicate site to render all images, which are "important".
@Silmas There are two warning message 'This request did not pass the external content moderation policy.' and 'Potentially ToS violating content detected'. There might be more but I am not aware. So both are CivitAI filter?
@Silverglasses67 No, I get a red sign underneath, the prompt, and it states the bad words I use... even if they are not in the prompt...
@Silmas @Silverglasses67 Agreed, the censorship comes from Civitai, not from Open AI ( on .com). sensual, seductive and so on all pass with chat GPT but are blocked here for .com ( not for .red)
People in here acting like this is revolutionary 😭am I the only one who thinks this is just a reskinned SD1.5 💀 be honest… this looks mid compared to older models ! Also why does nobody talk about how inconsistent this is ?
Me who actually used Stable Diffusion 1.5, I can tell you this model is nowhere near Stable Diffusion 1.5 quality at all. It's the same level as Nano Banana with way infact better prompt adherence. I'm curious to know why you think this model looks like a reskinned SD 1.5
Maybe get better at prompting? OFC if your goal is to generate generic waifus then SD1.5 is the way to go.
You are not the only one to have complaints about the model, and it does have some issues, but I think the blind benchmark scores do not lie, it is the best model available right now and can make some amazing stuff that you just cannot do as well with other models.
I have found best quality is achieved with minimal prompts (as state in the prompting guide) : https://developers.openai.com/cookbook/examples/multimodal/image-gen-models-prompting-guide
I mean, if you're having quality issues that might be on your end mate. Just a quick look at some of the best images output with it shows the potential it has when you're skilled with the model; it definitely has a few bugs here and there, especially in long-context edits, but come on now.
@Valentinedesigns “that’s kinda my point though—if it needs very specific ‘skill’ or conditions to shine, then consistency is still an issue no? like potential ≠ reliability”
@J1B “benchmarks are cool but they don’t really reflect real usage though… like yeah it scores high but then you try something slightly off-template and it just starts freestyling”
@WhyNaNWhenNyan You cant get better at prompting when the model dosent't understand prompting ... Also it so censored the waifu always have clothes on ...
@elevendr “it’s not 1:1 obviously, it’s more like the same failure patterns but polished… like it looks better at first glance but the moment you push composition or details it starts doing that familiar SD1.5 chaos. hard to explain but once you see it you can’t unsee it”
it’s like it optimized for aesthetic coherence over structural consistency… looks impressive until you analyze it properly ..
@L10n_H34r7 No. I get consistently fantastic outputs just as I have with the last year of models— just because a model is better in the hands of someone very good when that type of model or a skilled user or someone who takes the time to train their GPT doesn't mean it's bad it means it just takes more time and practice to master— that's not necessarily a bad thing or a good thing, I'd argue it's also the single most versatile toolset that also works amazingly when paired inline with other tools both traditional (photoshop, procreate, etc.) and other generative tools as part of a workflow.
why even have this on civitai if the checkpoint is not downloadable???
Because its a Trillion dollar proprietary model.... and 'I think' it wouldn’t fit on your local PC...🤣
@AIBOZO i know, i'm just saying this shouldn't be the place for it
@KoinnAI oh no how terrible you can use a proprieatary model for that you would normally have to subcribe to gen more then a couple of images, completely for free here with just blue buzz which they give you just for using the site and which has no other use. Adding APIs to proprietary models, does not even in the slightest tanget all the free and open source models here. How can one be so stubborn?
“I tested it again and yeah it’s good… but only when you don’t touch anything. The moment you try to control it, it spiritually reverts back to SD1.5 energy” Also I managed to downloaded this Model and now my GPU is asking for emotional support, pls advise”
I wish I could generate some friends as easily now
yeah why does this generate masterpieces but I still can’t generate a stable relationship ???
@L10n_H34r7 because AI is better than us. 😭
This is a something-with-something model!
I really want to get a portable version, otherwise it hits my BUZZ reserves hard (I've already spent 3,000 in 2 days)...
medium quality really save your buzz and the quality is still ok
@Fujido Thanks for the advice. What an insane difference in cost (according to my calculations, 3.8 times, while the quality drops visually by ~15-30%, which is very acceptable!)
@_MUKU_ also check the creator tip, its a closed source checkpoint so creator tip only a waste, low quality can get as low as 14 buzz, medium 53, and high 209 buzz
@Fujido It's kind of slimy how civitai keep turning the creator tip back on every time you log in again though. Besides, at these extortionate prices per image.....a further tip just seems insulting lol.
it's realy good
Grok Imagine = 26 buzz
ChatGPT Images 2.0 = 209 buzz
I'll preface this by saying this may be the best model in terms of content breadth, whilst also having good anatomy, etc, & the image quality output is really good (though I don't personally love the style). But it is also the most censored model!
Here's the real rub though; I just don't think this is even remotely worth the cost. You can use Grok Imagine for 26 buzz an image, it's way less censored (& whilst quality is kind of subjective in a lot of ways; I actually prefer Grok Imagine overall, but ChatGPT might be like 15-20% metric based objectively better @ "High Quality")..... I don't really see anything to warrant the insane price discrepancy on the user's end though. It just feels like a rip-off by comparison.
PS - I know OpenAI have to pass the training/compute costs, etc to consumers.....but it just feels hard to justify it in terms of bang for buck (which is really poor here).
I know there's "Medium" & "Low" quality options, also..... but "High" quality is barely better than Grok to begin with, & those options just further exacerbate the fake smudge like oil painting style ChatGPT 2.0 already has to begin with. "Medium" is still double the cost of Grok....
"Low" makes somewhat sense as competition to Z-Image, etc.....but it's still heavily censored.....& at that point; I'm just thinking "Why don't I just generate the image myself in 10 seconds, free from any lame censor restrictions lol"? (Though I concede that's not an option for everyone).
ChatGPT Images 2.0 = 53 buzz using medium which is the best option
@L10n_H34r7 Sure, but that's still twice the price of Grok Imagine. In your mind, what justifies double the price per image (in addition to the massively increased censorship ChatGTP has)?
It is slightly better in some ways over most of the competition (I'm not suggesting otherwise)......but not x2 the price better (@ medium quality). Not even close.
Having said all that, I have changed my mind somewhat. "Low quality" is competitive for price; just for spamming out low effort trash quick & easy. Quality is notably poorer, but it still maintains tight natural language understanding & prompt adherence. Otherwise, the value proposition over Grok is still poor.
Honestly, a large part of it, is that the sharp egdes are so rounded off, that I just don't find ChatGPT all that fun to use.
@ShadowCell So don't use it? And comparing the OpenAI API call which costs CivitAI money, with your free account, is a little unfair, isn't it?
Thank goodness I have a ChatGPT subscription.....it would be terrible to generate only three in the ChatGPT app then having to spend more blue buzz for them here lol
@elevendr Or you could just spend the money on a subscription for a better service like Grok? I used to primarily use ChatGPT, but it became hard to justify with the noticeable dip in quality/bias with the actual chat model (which I use to expedite scientific work). After it became useless in that regard (too many glaringly incorrect ‘answers’, that it would also refuse to concede were wrong; on their premium tier chat model btw). Answers so bad, no professional in the same field would even remotely accept them. Even when I linked it actual peer review consensus sources on the matter (just to see what would happen)…. still refused to budge. Wouldn’t give me any scientific sources for its insane quack stance either, but still insisted it was right. I re-tested just to see if it was chat session degrade. Nope. It gave the same dumb answers again.
Then at that point, it's kind of hard to justify paying just to generate meh images. Before I left, the chat model became so poor with scientific impartiality, I wouldn't even consider using their most expensive subscription model for free (the time wasted literally loses me money).
But ultimately, these companies aren't your friends.....so it boggles my mind when you get comments like @Silmas fanboying over them. You owe them nothing. It’s just a product. Not a cult. If a service is better elsewhere, why wouldn't you switch?
Everyone has different preferences, however. If you’re happy with it, fair enough. However, I still expect more mature answers from here than just “So don't use it” from the other poster @Silmas. It's just childish & cringe, so I'm not even going to waste any time engaging with it.
https://grok.com/imagine Check the image feed here. Compare with ChatGPT. I think If most people actually used both, then they wouldn’t waste any further time on ChatGPT (primarily due to how terrible even their premium LLM has become).
@Silmas it’s a bot account
@TheRedDread550 I'd definitely trust this guy. With his 0 followers and content.....ok, wait a moment...nevermind lol
What a clown.
I read your opinion with interest. I have little opportunity for testing. I managed to achieve something with this model that I hadn't been able to before. I even wrote a small article on this topic. https://civitai.red/articles/30579/a-small-town-on-a-flower-petal Here is another image. https://civitai.red/images/130738437 Previously, I hadn't been able to get hair intertwined in the form of jewelry like beads or a necklace. This model understood exactly what I wanted. That's why I marked this model for myself. I generated little in Grok Imagine. The overall impression is above average, but I have doubts that this model will work with my prompts and produce an image of the required quality. So overall, your opinion is correct, but when it comes to something complex, I would like to be wrong, but I am not sure that Grok Imagine can handle such tasks.
Chat GPT 2.0 is unmatched when it comes to historically accurate and cinematic images, or complex scenes with 5, 6 or even more protagonists. Grok can't, even remotely , compete whith it. Nano Banana is also quite good but not as good as GPT 2.0. I have to agree though that the cost is somewhat expensive and that's why I took a green membership . I had totally dropped GPT with their huge set back with GPT 1.5. I started using it again with GPT 2.0 for very complex scenes ( multiples people) or historical figures. There are still a few issues ( faces tend to be always the same clones, censorship is often ridiculously exagerated) but nonetheless it is my number 1 checkpoint for high end image generation.
"The provided image URL is not accessible or has expired. Please ensure the URL is valid and publicly accessible." guhh
It’s a good model, but I’m puzzled by the fact that it refuses to generate—for instance, due to rejecting reference images—yet the Buzz aren't refunded. I’m losing thousands of Buzz just testing out a few images...