4o Image Generation | Hacker News Reader

4o Image Generation

openai.com

1,072 pointsmeetpateltech599 comments

What's important about this new type of image generation that's happening with tokens rather than with diffusion, is that this is effectively reasoning in pixel space.

Example: Ask it to draw a notepad with an empty tic-tac-toe, then tell it to make the first move, then you make a move, and so on.

You can also do very impressive information-conserving translations, such as changing the drawing style, but also stuff like "change day to night", or "put a hat on him", and so forth.

I get the feeling these models are quite restricted in resolution, and that more work in this space will let us do really wild things such as ask a model to create an app step by step first completely in images, essentially designing the whole app with text and all, then writing the code to reproduce it. And it also means that a model can take over from a really good diffusion model, so even if the original generations are not good, it can continue "reasoning" on an external image.

Finally, once these models become faster, you can imagine a truly generative UI, where the model produces the next frame of the app you are using based on events sent to the LLM (which can do all the normal things like using tools, thinking, etc). However, I also believe that diffusion models can do some of this, in a much faster way.

> What's important about this new type of image generation that's happening with tokens rather than with diffusion, is that this is effectively reasoning in pixel space.

I do not think that this is correct. Prior to this release, 4o would generate images by calling out to a fully external model (DALL-E). After this release, 4o generates images by calling out to a multi-modal model that was trained alongside it.

You can ask 4o about this yourself. Here's what it said to me:

"So while I’m deeply multimodal in cognition (understanding and coordinating text + image), image generation is handled by a linked latent diffusion model, not an end-to-end token-unified architecture."

>You can ask 4o about this yourself. Here's what it said to me:

>"So while I’m deeply multimodal in cognition (understanding and coordinating text + image), image generation is handled by a linked latent diffusion model, not an end-to-end token-unified architecture."

Models don't know anything about themselves. I have no idea why people keep doing this and expecting it to know anything more than a random con artist on the street.

You're incorrect. 4o was not trained on knowledge of itself so literally can't tell you that. What 4o is doing isn't even new either, Gemini 2.0 has the same capability.

Models are famously good at understanding themselves.

I think this is actually correct even if the evidence is not right.

See this chat for example:

https://chatgpt.com/share/67e355df-9f60-8000-8f36-874f8c9a08...

A lot of convoluted explanations about something we don't even know if it really works all the time. I feel like in the third year of LLM-Hype and after reminde-me-how-many billions of dollars burned, we should by now not have to imagine what 'might happen' down to road, it should have been happening already. The use-case you are describing, sure sounds very interesting, until I remember asking asked Copilot for a simple scaffolding structure in React, and it spat out something which lacked half of imports and proper visual alignments. A few years ago I was excited about the possibility of removing all the scaffolding and templating work so we can write the cool parts, but they cannot even do that right. It's actually a step back compared to automatic code generators of the past, because those at least produced reproducible results every single time you used them. But hey sure, the next generation of "AI" (it's not really AI) will probably solve it.

That scene is changing so quickly that you will want to try again right now if you can.

While LLM code generation is very much still a mixed bag, it has been a significant accelerator in my own productivity, and for the most part all I am using is o1 (via the openAI website), deepseek, and jetbrains' AI service (Copilot clone). I'm eager to play with some of the other tooling available to VS Code users (such as cline)

I don't know why everyone is so eager to "get to the fun stuff". Dev is supposed to be boring. If you don't like it maybe you should be doing something else.

Consider using a better AI IDE platform than Copilot ... cursor, windsurf, cline, all great options that do much better than what you're describing. The underlying LLM capabilities also have advanced quite a bit in the past year.

> truly generative UI, where the model produces the next frame of the app

Please sir step away from the keyboard now!

That is an absurd proposition and I hope I never get to use an app that dreams of the next frame. Apps are buggy as they are, I don't need every single action to be interpreted by LLM.

An existing example of this is that AI Minecraft demo and it's a literal nightmare.

This argument could be made for every level of abstraction we've added to software so far... yet here we are commenting about it from our buggy apps!

First it will dream up the interaction frame by frame. Next, to improve efficiency, it will cache those interaction representations. What better way to do that than through a code representation.

While I think current AI can’t come close to anything remotely usable, this is a plausible direction for the future. Like you, I shudder.

I guess you have not heard that NVidia is generating frames with AI on their GPUs now: https://www.nvidia.com/en-us/geforce/news/dlss4-multi-frame-...

> “DLSS Multi Frame Generation generates up to three additional frames per traditionally rendered frame, working in unison with the complete suite of DLSS technologies to multiply frame rates by up to 8X over traditional brute-force rendering. This massive performance improvement on GeForce RTX 5090 graphics cards unlocks stunning 4K 240 FPS fully ray-traced gaming.”

It still can't generate a full glass of wine. Even in follow up questions it failed to manipulate the image correctly.

https://i.imgur.com/xsFKqsI.png

"Draw a picture of a full glass of wine, ie a wine glass which is full to the brim with red wine and almost at the point of spilling over... Zoom out to show the full wine glass, and add a caption to the top which says "HELL YEAH". Keep the wine level of the glass exactly the same."

Are you sure you are using the new 4o image generation?

https://imgur.com/a/wGkBa0v

I don't buy the meme or w/e that they can't produce an image with the full glass of wine. Just takes a little prompt engineering.

Using Dall-e / old model without too much effort (I'd call this "full".)

https://imgur.com/a/J2bCwYh

They're glass-half-full type models.

Yeah, it seems like somewhere in the semantic space (which then gets turned into a high resolution image using a specialized model probably) there is not enough space to hold all this kind of information. It becomes really obvious when you try to meaningfully modify a photo of yourself, it will lose your identity.

For Gemini it seems to me there's some kind of "retain old pixels" support in these models since simple image edits just look like a passthrough, in which case they do maintain your identity.

Also still seems to have a hard time consistently drawing pentagons. But at least it does some of the time, which is an improvement since last time I tried, when it would only ever draw hexagons.

I think it is not the AI but you who is wrong here. A full glass of wine is filled only up to the point of max radius so that the surface to air is maxed an the wine can breathe. This is what we taught the AI to consider „a full glass of wine“ and it perfectly gets it right.

Nope, it can.

Got it in two requests, https://chatgpt.com/share/67e41576-8840-8006-836b-f7358af494... for the prompts.

The question remains: why would you generate a full glass of wine? Is that something really that common?

> What's important about this new type of image generation that's happening with tokens rather than with diffusion

That sounds really interesting. Are there any write-ups how exactly this works?

Would be interested to know as well. As far as I know there is no public information about how this works exactly. This is all I could find:

> The system uses an autoregressive approach — generating images sequentially from left to right and top to bottom, similar to how text is written — rather than the diffusion model technique used by most image generators (like DALL-E) that create the entire image at once. Goh speculates that this technical difference could be what gives Images in ChatGPT better text rendering and binding capabilities.

https://www.theverge.com/openai/635118/chatgpt-sora-ai-image...

There are a few different approaches. Meta documents at least one approach quite well in one of their llama papers.

The general gist is that you have some kind of adapter layers/model that can take an image and encode it into tokens. You then train the model on a dataset that has interleaved text and images. Could be webpages, where images occur in-between blocks of text, chat logs where people send text messages and images back and forth, etc.

The LLM gets trained more-or-less like normal, predicting next token probabilities with minor adjustments for the image tokens depending on the exact architecture. Some approaches have the image generation be a separate "path" through the LLM, where a lot of weights are shared but some image token specific weights are activated. Some approaches do just next token prediction, others have the LLM predict the entire image at once.

As for encoding-decoding, some research has used things as simple as Stable Diffusion's VAE to encode the image, split up the output, and do a simple projection into token space. Others have used raw pixels. But I think the more common approach is to have a dedicated model trained at the same time that learns to encode and decode images to and from token space.

For the latter approach, this can be a simple model, or it can be a diffusion model. For encoding you do something like a ViT. For decoding you train a diffusion model conditioned on the tokens, throughout the training of the LLM.

For the diffusion approach, you'd usually do post-training on the diffusion decoder to shrink down the number of diffusion steps needed.

The real crutch of these models is the dataset. Pretraining on the internet is not bad, since there's often good correlation between the text and the images. But there's not really good instruction datasets for this. Like, "here's an image, draw it like a comic book" type stuff. Given OpenAI's approach in the past, they may have just bruteforced the dataset using lots of human workers. That seems to be the most likely approach anyway, since no public vision models are quite good enough to do extensive RL against.

And as for OpenAI's architecture here, we can only speculate. The "loading from top to be from a blurry image" is either a direct result of their architecture or a gimmick to slow down requests. If the former, it means they are able to get a low resolution version of the image quickly, and then slowly generate the higher resolution "in order." Since it's top-to-bottom that implies token-by-token decoding. My _guess_ is that the LLM's image token predictions are only "good enough." So they have a small, quick decoder take those and generate a very low resolution base image. Then they run a stronger decoding model, likely a token-by-token diffusion model. It takes as condition the image tokens and the low resolution image, and diffuses the first patch of the image. Then it takes as condition the same plus the decoded patch, and diffuses the next patch. And so forth.

A mixture of approaches like that allows the LLM to be truly multi-modal without the image tokens being too expensive, and the token-by-token diffusion approach helps offset memory cost of diffusing the whole image.

I don't recall if I've seen token-by-token diffusion in a published paper, but it's feasible and is the best guess I have given the information we can see.

EDIT: I should note, I've been "fooled" in the past by OpenAI's API. When o* models first came out, they all behaved as if the output were generated "all at once." There was no streaming, and in the chat client the response would just show up once reasoning was done. This led me to believe they were doing an approach where the reasoning model would generate a response and refine it as it reasoned. But that's clearly not the case, since they enabled streaming :P So take my guesses with a huge grain of salt.

> truly generative UI, where the model produces the next frame of the app

I built this exact thing last month, demo: https://universal.oroborus.org (not viable on phone for this demo, fine on tablet or computer)

Also see discussion and code at: http://github.com/snickell/universal

I wasn't really planning to share/release it today, but, heck, why not.

I started with bitmap-style generative image models, but because they are still pretty bad at text (even this, although it’s dramatically better), for early-2025 it’s generating vector graphics instead. Each frame is an LLM response, either as an svg or static html/css. But all computation and transformation is done by the LLM. No code/js as an intermediary. You click, it tells the LLM where you clicked, the LLM hallucinates the next frame as another svg/static-html.

If it ran 50x faster it’d be an absolutely jaw dropping demo. Unlike "LLMs write code", this has depth. Like all programming, the "LLMs write code" model requires the programmer or LLM to anticipate every condition in advance. This makes LLM written "vibe coded" apps either gigantic (and the llm falls apart) or shallow.

In contrast, as you use universal, you can add or invent features ranging from small to big, and it will fill in the blanks on demand, fairly intelligently. If you don't like what it did, you can critique it, and the next frame improves.

Its agonizingly slow in 2025, but much smarter and in weird ways less error prone than using the LLM to generate code that you then run: just run computation via the LLM itself.

You can build pretty unbelievable things (with hallucinated state, granted) with a few descriptive sentences, far exceeding the capabilities you can “vibe code” with the description. And it never gets lost in its rats nest of self generated garbage code because… there is no code to in.

Code is medium with a surprisingly strong grain. This demo is slow, but SO much more flexible and personally adaptable than anything I’ve used where the logic is implemented cia a programming language.

I don’t love this as a programmer, but my own use of the demo makes me confident that programming languages as a category will have a shelf life if LLM hardware gets fast, cheap and energy efficient.

I suspect LLMs will generate not programming language code, but direct wasm or just machine code on the fly for things that need faster traction than they can draw a frame, but core logic will move out of programming languages (not even llm written code). Maybe similar to the way we bind to low level fast languages but a huge percentage of “business” logic is written in relatively slower languages.

FYI, I may not be able to afford the credits if too many people visit, I put a a $1000 of credits on this, we'll see if that lasts. This is claude 3.7, I tried everything else, a claude had the visual intelligence today. IMO this is a much more compelling glance at the future than coding models. Unfortunately, generating an SVG per click is pricey, each click/frame costs me about $0.05. I’ll fund this as far as I can so folks can play with it.

Anthropic? You there? Wanna throw some credits at an open source project doing something that literally only works on claude today? Not just better, but “only Claude 3.7 can show this future today?”. I’d love for lots more people to see the demo, but I really could use an in-kind credit donation to make this viable. If anyone at anthropic is inspired and wants to hook me up: snickell@alumni.stanford.edu. Very happy to rep Claude 3.7 even more than I already do.

I think it’s great advertising for Claude. I believe the reason Claude seems to do SO much better at this task is, one it shows far greater spatial intelligence, and two, I distract they are the only state of the art model intentionally training on SVG.

I’m a bit late here - but I’m the COO of OpenRouter and would love to help out with some additional credits and share the project. It’s very cool and more people could be able to check it out. Send me a note. My email is cc at OpenRouter.ai

This is super cool! I think new kinds of experiences can be built with infinite generative UIs. Obviously there will need to be good memory capabilities, maybe through tool use.

If you end up taking this further and self hosting a model you might actually achieve a way faster “frame rate” with speculative decoding since I imagine many frames will reuse content from the last. Or maybe a DSL that allows big operations with little text. E.g. if it generates HTML/SVG today then use HAML/Slim/Pug: https://chatgpt.com/share/67e3a633-e834-8003-b301-7776f76e09...

Wonderful, good job! Reminds me of https://arstechnica.com/information-technology/2022/12/opena...

Do you have any demo videos?

It’s like a lucid dream version of using and modifying the software at the same time.

This is so impressive. You are building the future!

It also would mean that the model can correctly split the image into layers, or segments, matching the entities described. The low-res layers can then be fed to other image-processing models, which would enhance them and fill in missing small details. The result could be a good-quality animation, for instance, and the "character" layers can even potentially be reusable.

Pretty sure the modern Gemini image models can already do token based image generation/editing and are significantly better and faster.

Yeah Gemini has had this for a few weeks, but much lower resolution. Not saying 4o is perfect, but my first few images with it are much more impressive than my first few images with Gemini.

It's faster but it's definitely not better than what's being showcased here. The quality of Flash 2 Image gens are generally pretty meh.

That's very interesting. I would have assumed that 4o is internally using a single seed for the entire conversation, or something analogous to that, to control randomness across image generation requests. Can you share the technical name for this reasoning process so I could look up research about it?

multimodal chain of thought / generation of thought

Nobody has really decided on a name.

Also chain of thought is somewhat different from chain of thought reasoning so mb throw in multimodal chain of thought reasoning

>You can also do very impressive information-conserving translations, such as changing the drawing style, but also stuff like "change day to night", or "put a hat on him", and so forth.

You can do that with diffusion, too. Just lock the parameters in ComfyUi.

> Finally, once these models become faster, you can imagine a truly generative UI, where the model produces the next frame of the app you are using based on events sent to the LLM

With current GPU technology, this system would need its own Dyson sphere.

Is it able to break the usual failure modes of these models, that all clocks are at 10 min past two, or they can't produce images of people drawing with the left hand?

> writing the code to reproduce it

I'm super excited for all the free money and data our new AI written apps will be giving away.

Hmmm, I wanted to do that tic tac toe example, and it failed to create a 3x3 grid, instead creating a 5x5 (?) grid with two first moves marked.

https://chatgpt.com/share/67e32d47-eac0-8011-9118-51b81756ec...

Wrt reasoning I’ll believe it when I see it. I just tried several variants of “Generate an image of a chess board in which white has played three great moves and black has played two bad moves.” Results are totally nonsensical as always.

Ran through some of my relatively complex prompts combined with using pure text prompts as the de-facto means of making adjustments to the images (in contrast to using something like img2img / inpainting / etc.)

https://mordenstar.com/blog/chatgpt-4o-images

It's definitely impressive though once again fell flat on the ability to render a 9-pointed star.

Didn't work for me on the first prompting (got a 10-pointed one), but after sending [this is 10 points, make it 9] it did render a 9-pointed one too

Have u had any luck with engineering/schematics/wireframe diagrams such as [1] ??

[1] https://techcrunch.com/wp-content/uploads/2024/03/pasted-ima...

Great question. I haven't tested the creation of such an image from scratch, but I did add an adjustment test against that specific text-heavy diagram and I'd say it passed with "flying colors". (pun intended).

Here's what it created based on a text description of your schematic

https://i.imgur.com/sGfdtWo.png

Funny how it changed the spelling in the Magic coloring book. Westher, wntility, substittutesfor, yyears. What's worse is that it removed the CO2 tank and changed some vital numbers, eg 8' to 3', completely altering the meaning of the schematic. Not what i would call passing with flying colors. Still pretty cool as a party trick, like most other AI outputs, useful only with careful review.

Armless Venus with bread is true art

Fantastic prompts!

I’ve just tried it and oh wow it’s really good. I managed to create a birthday invitation card for my daughter in basically 1-shot, it nailed exactly the elements and style I wanted. Then I asked to retain everything but tweak the text to add more details about the date, venue etc. And it did. I’m in shock. Previous models would not be even halfway there.

share prompt minus identifying details?

> Draw a birthday invitation for a 4 year old girl [name here]. It should be whimsical, look like its hand-drawn with little drawings on the sides of stuff like dinosaurs, flowers, hearts, cats. The background should be light and the foreground elements should be red, pink, orange and blue.

Then I asked for some changes:

> That's almost perfect! Retain this style and the elements, but adjust the text to read:

> [refined text]

> And then below it should add the location and date details:

> [location details]

> Introducing 4o Image Generation: [...] our most advanced image generator yet

Then google:

> Gemini 2.5: Our most intelligent AI model

> Introducing Gemini 2.0 | Our most capable AI model yet

I could go on forever. I hope this trend dies and apple starts using something effective so all the other companies can start copying a new lexicon.

We're in the middle of a massive and unprecedented boom in AI capabilities. It is hard to be upset about this phrasing - it is literally true and extremely accurate.

If that's so then there's no need to be hyperbolic about it. Why would they publish a model that is not their most advanced model?

This is my latest and most advanced comment yet.

Has post-Jobs Apple ever come up with anything that would warrant this hope?

Every iPhone is their best iPhone yet

Apple isn't really the best software company and though they were early to digital assistants with Siri, it seems like they've let it languish. It's almost comical how bad Siri is given the capabilities of modern AI. That being said, Android doesn't really have a great builtin solution for this either.

Apple is more of a hardware company. Still, Cook does have a few big wins under his belt: M-series ARM chips on Macs, Airpods, Apple watch, Apple pay.

Apple silicon chips

No, but I think they stopped with "our most" (since all other brainless corps adopted it) and just connect adjectives with dots.

Hotwheels: Fast. Furious. Spectacular.

When you keep improving, it's always going to be the best or most: https://www.youtube.com/watch?v=bPkso_6n0vs

Every step of gradient descent is the best model yet!

Not if you do gradient descent with momentum.

Maybe it’s not useless. 1) it’s only comparing it to their own products and 2) it’s useful to know that the product is the current best in their offering as opposed to a new product that might offer new functionality but isn’t actually their most advanced.

Which is especially relevant when it's not obvious which product is the latest and best just looking at the names. Lots of tech naming fails this test from Xbox (Series X vs S) to OpenAI model names (4o vs o1-pro).

Here they claim 4o is their most capable image generator which is useful info. Especially when multiple models in their dropdown list will generate images for you.

What's the problem?

It's a nitpick about the repetitive phrasing for announcements

<Product name>: Our most <superlative> <thing> yet|ever.

This actually makes sense because the versioning is so confusing they could be releasing a lesser/lightweight model for all we know.

https://www.youtube.com/watch?v=CUPDRnUWeBA

Maybe they used AI to come up with the tag line.

OpenAI's livestream of GPT-4o Image Generation shows that it is slowwwwwwwwww (maybe 30 seconds per image, which Sam Altman had to spin "it's slow but the generated images are worth it"). Instead of using a diffusion approach, it appears to be generating the image tokens and decoding them akin to the original DALL-E (https://openai.com/index/dall-e/), which allows for streaming partial generations from top to bottom. In contrast, Google's Gemini can generate images and make edits in seconds.

No API yet, and given the slowness I imagine it will cost much more than the $0.03+/image of competitors.

As a user, images feel slightly slower but comparable to the previous generation. Given the significant quality improvement, it's a fair trade-off. Overall, it feels snappy, and the value justifies a higher price.

Maybe this is the dialup of the era.

LLMs are autoregressive, so they can't be (multi-modality) integrated with diffusion image models, only with autoregressive image models (which generate an image via image tokens). Historically those had lower image fidelity than diffusion models. OpenAI now seems to have solved this problem somehow. More than that, they appear far ahead of any available diffusion model, including Midjourney and Imagen 3.

Gemini "integrates" Imagen 3 (a diffusion model) only via a tool that Gemini calls internally with the relevant prompt. So it's not a true multimodal integration, as it doesn't benefit from the advanced prompt understanding of the LLM.

Edit: Apparently Gemini also has an experimental native image generation ability.

If you look at the examples given, this is the first time I've felt like AI generated images have passed the uncanny valley.

The results are ground breaking in my opinion. How much longer until an AI can generate 30 successive images together and make an ultra realistic movie?

i find this “slow” complaint (/observation— i dont view this comment as a complaint, to be clear) to be quite confusing. slow… compared to what, exactly? you know what is slow? having to prompt and reprompt 15 times to get the stupid model to spell a word correctly and it not only refuses, but is also insistent that it has corrected the error this time. and afaict this is the exact kind of issue this change should address substantially.

im not going to get super hyperbolic and histrionic about “entitlement” and stuff like that, but… literally this technology did not exist until like two years ago, and yet i hear this all the time. “oh this codegen is pretty accurate but it’s slow”, “oh this model is faster and cheaper (oh yeah by the way the results are bad, but hey it’s the cheapest so it’s better)”. like, are we collectively forgetting that the whole point of any of this is correctness and accuracy? am i off-base here?

the value to me of a demonstrably wrong chat completion is essentially zero, and the value of a correct one that anticipates things i hadn’t considered myself is nearly infinite. or, at least, worth much, much more than they are charging, and even _could_ reasonably charge. it’s like people collectively grouse about low quality ai-generated junk out of one side of their mouths, and then complain about how expensive the slop is out of the other side.

hand this tech to someone from 2020 and i guarantee you the last thing you’d hear is that it’s too slow. and how could it be? yeah, everyone should find the best deals / price-value frontier tradeoff for their use case, but, like… what? we are all collectively devaluing that which we lament is being devalued by ai by setting such low standards: ourselves. the crazy thing is that the quickly-generated slop is so bad as to be practically useless, and yet it serves as the basis of comparison for… anything at all. it feels like that “web-scale /dev/null” meme all over again, but for all of human cognition.

> it appears to be generating the image tokens and decoding them akin to the original DALL-E

The animation is a lie. The new 4o with "native" image generating capabilities is a multi-modal model that is connected to a diffusion model. It's not generating images one token at a time, it's calling out to a multi-stage diffusion model that has upscalers.

You can ask 4o about this yourself, it seems to have a strong understanding of how the process works.

i mean on free chat an image took maybe 2 seconds?

I’ll just be happy with not everything having that over saturated cg/cartoon style that you cant prompt your way out of.

I was relying on that to determine if images were AI though

Frustratingly the DALL-E API actually has an option for this, you can switch it from "vivid" to "realistic".

This option is not exposed in ChatGPT, it only uses vivid.

Is that an artifact of the training data? Where are all these original images with that cartoony look that it was trained on?

you really have to NOT try to end up with that result in MJ.

It's incredible that this took 316 days to be released since it was initially announced. I do appreciate the emphasis in the presentation on how this can be useful beyond just being a cool/fun toy, as it seems most image generation tools have functioned.

Was anyone else surprised how slow the images were to generate in the livestream? This seems notably slower than DALLE.

I've never minded that an image might take 10-30 seconds to generate. The fact that people do is crazy to me. A professional artist would take days, and cost $100s for the same asset.

I ran stable diffusion for a couple of years (maybe?, time really hasn't made sense since 2020) on my Dual 3090 rendering server. I built the server originally for crypto heating my office in my 1820s colonial in upstate NY then when I was planning to go back to college (got accepted into a university in England), I switched it's focus to Blender/UE4 (then 5), then eventually to AI image gen. So I've never minded 20 seconds for an image. If I needed dozens of options to pick the best, I was going to click start and grab a cup of coffee, come back and maybe it was done. Even if it took 2 hours, it is still faster than when I used to have to commission art for a project.

I grew out of Stable Diffusion, though, because the learning curve beyond grabbing a decent checkpoint and clicking start was actually really high (especially compared to LLMs that seamed to "just work"), after going through failed training after failed fine-tuning using tutorials that were a couple days out of date, I eventually said, fuck it, I'm paying for this instead.

All that to say - if you are using GenAI commercially, even if an image or a block of code took 30 minutes, it's still WAY cheaper than a human. That said, eventually a professional will be involved, and all the AI slop you generated will be redone, which will still cost a lot, but you get to skip the back and forth figuring out style/etc.

Is there any way to see whether a given prompt was serviced by 4o or Dall-E?

Currently, my prompts seem to be going to the latter still, based on e.g. my source image being very obviously looped through a verbal image description and back to an image, compared to gemini-2.0-flash-exp-image-generation. A friend with a Plus plan has been getting responses from either.

The long-term plan seems to be to move to 4o completely and move Dall-E to its own tab, though, so maybe that problem will resolve itself before too long.

4o generates top down (picture goes from mostly blurry to clear starting from the top). If it's not generating like that for you then you don't have it yet.

If you don't have access to it on ChatGPT yet, you can try Sora, which already has access for me.

I've generated (and downloaded) a couple of images. All filenames start with `DALL·E`, so I guess that's a safe way to tell how the images were generated.

My images say "Created with DALLE" below them, and a little info icon tells me they are rolling out the new one soon.

don't enable images on the chat model if your using the site, just leave it all disabled and ask for an image, if you enable dall-e it switches to dall-e is what i've seen

the native just.. works

My experience with these announcements is that they're cherry picking the best results from a maybe several hundred or a thousand prompts.

I'm not saying that it's not true, it's just "wait and see" before you take their word as gold.

I think MS's claim on their quantum computing breakthrough is the latest form of this.

> My experience with these announcements is that they're cherry picking the best results from a maybe several hundred or a thousand prompt

just tried it, prompt adherence and quality is... exactly what they said, it extremely impressive

The examples they show have little captions that say "best of #", like "best of 8" or "best of 4". Hopefully that truly represents the odds of generating the level of quality shown.

Have you tried it? It's crazy good.

it’s rolling out to users on all tiers, so no need to wait. I tried it and saw outputs from many others. it’s good. very good

I got the occasional A/B test with a new image generator while playing with Dall-E during a one month test of Plus. It was always clear which one was the new model because every aspect was so much better. I assume that model and the model they announced are the same.

You should try it out yourself.

The new model in the drop down says something like "4o Create Image (Updated)". It is truly incredible. Far better than any other image generator as far as understanding and following complex prompts.

I was blown away when they showed this many months ago, and found it strange that more people weren't talking about it.

This is much more precise than the Gemini one that just came out recently.

> found it strange that more people weren't talking about it.

Some simply dislike everything OpenAI. Just like everything Musk or Trump.

This is really impressive, but the "Best of 8" tag on a lot of them really makes me want to see how cherry-picked they are. My three free images had two impressive outputs and one failure.

The high five looks extremely unnatural. Their wrists are aligned, but their fingers aren't, somehow?

If that's best of 8, I'd love to see the outtakes.

First AI image generator to pass the uncanny valley test? Seems like it. This is the biggest leap in image generation quality I've ever seen.

How much longer until an AI that can generate 30 frames with this quality and make a movie?

About 1.5 years ago, I thought AI would eventually allow anyone with an idea to make a Hollywood quality movie. Seems like we're not too far off. Maybe 2-3 more years?

>First AI image generator to pass the uncanny valley test?

Other image generators I've used lately often produced pretty good images of humans, as well [0]. It was DALLE that consistently generated incredibly awful images. Glad they're finally fixing it. I think what most AI image generators lack the most is good instruction following.

[0] YandexArt for the first prompt from the post: https://imgur.com/a/VvNbL7d The woman looks okay, but the text is garbled, and it didn't fully follow the instruction.

Ideogram 2.0 and Recraft also create images that looks very much real.

For drawings, NovelAI's models are way beyond the uncanny valley now.

The whiteboard image is insane. Even if it took more than 8 to find it, it's really impressive.

To think that a few years ago we had dreamy pictures with eyes everywhere. And not long ago we were always identifying the AI images by the 6 fingered people.

I wonder how well the physics is modeled internally. E.g. if you prompt it to model some difficult ray tracing scenario (a box with a separating wall and a light in one of the chambers which leaks through to the other chamber etc)?

Or if you have a reflective chrome ball in your scene, how well does it understand that the image reflected must be an exact projection of the visible environment?

I remember literally just two or three years back getting good text was INSANE. We were all amazed when SD started making pretty good text.

am I dumb or every time they release something I can never find out how to actually use it and forget about it. take this for instance I wanted to try out their newton "an infographic explaining newton's prism experiment in great detail" example, but it generated a very bad result but maybe it's because I'm not using the right model? every release of theirs is not really a release, it's like a trailer. right?

This is hilarious. I'm also confused about whether they released it or not because the results are underwhelming.

EDIT: Ok it works in Sora, and my jaw dropped

You're not dumb. They do this for nearly every single major release. I can't really understand why considering it generates negative sentiment about the release, but it's something to be expected from OpenAI at this point.

This is what's so wild about Anthropic. When they release it seems like it's rolled out to all users, and API customers immediately. OpenAI has MONTHS between annoucement and roll out, or if they do it's usually just influencers who get an "early look". It's pretty frustrating.

It's very impressive. It feels like the text is a bit of a hack where they're somehow rendering the text separately and interpolating it into the image. Not always, I got it to render calligraphy with flourishes, but only for a handful of words.

For example, I asked it to render a few lines of text on a medieval scroll, and it basically looked like a picture of a gothic font written onto a background image of a scroll

This doesn't explain the whiteboard example at the link. The handwritten text on the whiteboard does not look like a font at all.

Visual internet content is completely over. Pack it up

For starters, this completely blocks generation of anything remotely related to copy-protected IPs, which may actually be a saving grace for some creatives. There's a lot of demand for fanart of existing characters, so until this type of model can be run locally, the legal blocks in place actually give artists some space to play in where they don't have to compete with this. At least for a short while.

So I spent a good few hours investigating the current state of the art a few weeks ago. I would like to generate a collection of images for the art in a video game.

It is incredibly difficult to develop an art style, then get the model to generate a collection of different images in that unique art style. I couldn't work out how to do it.

I also couldn't work out how to illustrate the same characters or objects in different contexts.

AI seems great for one off images you don't care much about, but when you need images to communicate specific things, I think we are still a long way away.

The character consistency and UI capabilities seem like they open up a lot of new use cases.

https://chatgpt.com/share/67e39ffa-3a98-8011-ab79-fe3ac76632...

Asking it to draw the Balkans map in Tolkien style, this is actually really impressive, geography is more or less completely correct, borders and country locations are wrong, but it feels like something I could get it to fix.

Strange, I'm getting

> I wasn't able to generate the map because the request didn't follow content policy guidelines. Let me know if you'd like me to adjust the request or suggest an alternative way to achieve a similar result.

Are you in the US?

...why are we living in such a retarded sci-fi age

I try this on every new generation:

Generate a photo of a lake taken by a mobile phone camera. No hands or phones in the photo, just the lake.

The hand holding a phone is always there :D

Not when I tried https://i.imgur.com/1cPaXn5.png

Few image generation AI tools understand negatives. If you tell it "not" to do something, that something will always appear somehow.

ditto for me. prompted to remove hands, but no success.

Anyone else frightened by this? Seeing meant believing, and now that isnt the case anymore...

In the short term, yes. Over the long run, I think it's good that we move away from the "seeying is believing" model, since that was already abused by bad actors/propaganda Hopefully, not too much chaos until we find another solution.

No, people have been making photorealistic and convincing images in photoshop for ages.

This specifically? No. We’ve been on this path a while now.

The general idea of indistinguishable real/fake images; yeah

Look closer at the fingers. These models still don’t have a firm handle on them. The right elbow on the second picture also doesn’t quite look anatomically possible.

Nah, I'll maybe start taking them seriously when they can draw someone grating cheese, but holding the cheese and the grater as if they were playing violin.

Not really, but only because literally every person I've met that spends a lot of time on TikTok starts spouting unhinged nonsense at me.

You don't even need deepfakes. https://www.newsweek.com/doug-mastriano-pennsylvania-senator...

The disaster scenario is already here.

it's fine, there will be new jobs.

I enjoy trying to break these models. I come up with prompts that are uncommon but valid. I want to see how well they handle data not in their training set. For image generation I like to use “ Generate an image of a woman on vacation in the Caribbean, lying down on the beach without sunglasses, her eyes open.”

Will be interesting to see how this ranks against Google Imagen and Reve. https://huggingface.co/spaces/ArtificialAnalysis/Text-to-Ima...

ChatGPT Pro tip: In addition to video generation, you can use this new image gen functionality in Sora and apply all of your custom templates to it! I generated this template (using my Sora Preset Generator, which I think is public) to test reasoning and coherency within the image:

Theme: Educational Scientific Visualization – Ultra Realistic Cutaways Color: Naturalistic palettes that reflect real-world materials (e.g., rocky grays, soil browns, fiery reds, translucent biological tones) with high contrast between layers for clarity Camera: High-resolution macro and sectional views using a tilt-shift camera for extreme detail; fixed side angles or dynamic isometric perspective to maximize spatial understanding Film Stock: Hyper-realistic digital rendering with photogrammetry textures and 8K fidelity, simulating studio-grade scientific documentation Lighting: Studio-quality three-point lighting with soft shadows and controlled specular highlights to reveal texture and depth without visual noise Vibe: Immersive and precise, evoking awe and fascination with the inner workings of complex systems; blends realism with didactic clarity Content Transformation: The input is transformed into a hyper-detailed, realistically textured cutaway model of a physical or biological structure—faithful to material properties and scale—enhanced for educational use with visual emphasis on internal mechanics, fluid systems, and spatial orientation

Examples: 1. A photorealistic geological cutaway of Earth showing crust, tectonic plates, mantle convection currents, and the liquid iron core with temperature gradients and seismic wave paths. 2. An ultra-detailed anatomical cross-section of the human torso revealing realistic organs, vasculature, muscular layers, and tissue textures in lifelike coloration. 3. A high-resolution cutaway of a jet engine mid-operation, displaying fuel flow, turbine rotation, air compression zones, and combustion chamber intricacies. 4. A hyper-realistic underground slice of a city showing subway lines, sewage systems, electrical conduits, geological strata, and building foundations. 5. A realistic cutaway of a honeybee hive with detailed comb structures, developing larvae, worker bee behavior zones, and active pollen storage processes.

This works great for many purposes.

One area where it does not work well at all is modifying photographs of people's faces.* Completely fumbles if you take a selfie and ask it to modify your shirt, for example.

* = unless the people are in the training set

> We’re aware of a bug where the model struggles with maintaining consistency of edits to faces from user uploads but expect this to be fixed within the week.

Sounds like it may be a safety thing that's still getting figured out

It just doesn't have that kind of image editing capability. Maybe people just assume it does because Google's similar model has it. But did OpenAI claim it could edit images?

That's to be expected, no? It's a usian product so it will be a disappointment in all areas where things could get lewd.

Is it live yet? Have been trying it out and am still getting poor results on text generation.

I don't think it's available to everyone yet on 4o. Just like you I am getting the same "cartoony" styling and poor text generation.

Might take a day or two before it's available in general.

So far it seems to be the same for me.

It seems like an odd way to name/announce it, there's nothing obvious to distinguish it from what was already there (i.e. 4o making images) so I have no idea if there is a UI change to look for, or just keep trying stuff until it seems better?

This is OpenAI's bread and butter - announce something as though it's being launched and then proceed to slowly roll it out after a couple of days.

Truly infuriating, especially when it's something like this that makes it tough to tell if the feature is even enabled.

You're supposed to generate images, stupid /s

Are these models also based on copyrighted material? Can anyone provide a brief explanation if the datasets are scraped or CC super-public-images ?

I think the biggest problem I still see is the models awareness of the images it generated itself.

The glaring issue for the older image generators is how it would proudly proclaim to have presented an image with a description that has almost no relation to the image it actually provided.

I'm not sure if this update improves on this aspect. It may create the illusion of awareness of the picture by having better prompt adherence.

Something crazy, the old model couldn’t draw a 5.25 drive. I tried this myself at the time.

https://news.ycombinator.com/item?id=42628742

The new one can.

https://chatgpt.com/share/67e36dee-6694-8010-b337-04f37eeb5c...

Here's an example of iterative editing with the new model: https://chatgpt.com/share/67e30f62-12f0-800f-b1d7-b3a9c61e99...

It's much better than prior models, but still generates hands with too many fingers, bodies with too many arms, etc.

For some reason, I can't see the images in that chat, whether I'm signed in or in incognito mode.

I see errors like this in the console:

ewwsdwx05evtcc3e.js:96 Error: Could not fetch file with ID file_0000000028185230aa1870740fa3887b?shared_conversation_id=67e30f62-12f0-800f-b1d7-b3a9c61e99d6 from file service at iehdyv0kxtwne4ww.js:1:671 at async w (iehdyv0kxtwne4ww.js:1:600) at async queryFn (iehdyv0kxtwne4ww.js:1:458)Caused by: ClientRequestMismatchedAuthError: No access token when trying to use AuthHeader

You know the images themselves don’t get shared in links like that, right? (It even tells you so when you make the link.)

Looks about what you'd get with FLUX and attaching some language model to enhance your prompt with eg more text

Flux 1.1 Pro has good prompt adherence, but some of these (admittingly cherry-picked) GPT-4o generated image demos are beyond what you would get with Flux without a lot of iteration, particularly the large paragraphs of text.

I'm excited to see what a Flux 2 can do if it can actually use a modern text encoder.

Exactly. OpenAI isn't going to win image and video.

Sora is one of the worst video generators. The Chinese have really taken the lead in video with Kling, Hailuo, and the open source Wan and Hunyuan.

Wan with LoRAs will enable real creative work. Motion control, character consistency. There's no place for an OpenAI Sora type product other than as a cheap LLM add-in.

Flux doesn't do text that good

The real test for image generators is the image->text->image conversion. In other words it should be able to describe an image with words and then use the words to recreate the original image with a high accuracy. The text representation of the image doesn't have to be English. It can be a program, e.g. a shader, that draws the image. I believe in 5-10 years it will be possible to give this tool a picture of rainforest, tell it to write a shader that draws this forest, and tell it to add Avatar-style flying rocks. Instead of these silly benchmarks, we'll read headlines like "GenAI 5.1 creates a 3D animation of a photograph of the Niagara falls in 3 seconds, less than 4KB of code that runs at 60fps".

Why is that “the real test for image generators”? I mean, most image generators don't inherently include image->text functionality at all, so this seems more of a test of multimodal modals that include both t2i and i2t functionality, but even then, I don't think humans would generally pass this test well (unless the human doing the description test was explicitly told that the purpose was reproduction, but that's not the usual purpose of either human or image2text model descriptions.)

> ChatGPT’s new image generation in GPT‑4o rolls out starting today to Plus, Pro, Team, and Free users as the default image generator in ChatGPT, with access coming soon to Enterprise and Edu. For those who hold a special place in their hearts for DALL·E, it can still be accessed through a dedicated DALL·E GPT.

> Developers will soon be able to generate images with GPT‑4o via the API, with access rolling out in the next few weeks.

That's it folks. Tens of thousands of so-called "AI" image generator startups have been obliterated and taking digital artists with them all reduced to near zero.

Now you have a widely accessible meme generator with the name "ChatGPT".

The last task is for an open weight model that competes against this and is faster and all for free.

> Tens of thousands of so-called "AI" image generator startups have been obliterated and taking digital artists with them all reduced to near zero. Now you have a widely accessible meme generator with the name "ChatGPT".

ChatGPT has already had a that via Dall-E. If it didn't kill those startups when that happened this doesn't fundamentally change anything. Now its got a new image gen model, which — like Dall-E 3 when it came out — is competitive or ahead of other SotA base models using just text prompts, the simplest generation workflow, but both more expensive and less adaptable to more involved workflows than the tools anyone more than a casual user (whether using local tools or hosted services) is using. This is station-keeping for OpenAI, not a meaningful change in the landscape.

Yep. The coherence and text quality is insanely good. Keen to play with it to find it's "mangled hands" style deficiencies, because of course they cherry picked the best examples.

Really liked the fact that the team shared all the shortcomings of the model in the post. Sometimes products just highlights the best results and isn't forthcoming in areas that need improvement. Kudos to the OpenAI team on that.

For those who are still getting the old DALL-E images inside ChatGPT, you can access the new model on Sora: https://sora.com/explore/images

Has the meaning of the words "available today" changed since I learned them?

They're doing "vibe writing", where you just use words to mean random things regardless of whether they actually mean that.

Just curious, where do you see the words "available today"?

It is now available. It is still the same day of this release.

I wanted to use this to generate funny images of myself. Recently I was playing around with Gemini Image Generation to dress myself up as different things. Gemini Image Generation is surprisingly good, although the image quality quickly degrades as you add more changes. Nothing harmful, just silly things like dressing me up as a wizard or other typical RPG roles.

Trying out 4o image generation... It doesn't seem to support this use-case at all? I gave it an image of myself and asked to turn me into a wizard, and it generate something that doesn't look like me in the slightest. A second attempt, I asked to add a wizard hat and it just used python to add a triangle in the middle of my image. I looked at the examples and saw they had a direct image modification where they say "Give this cat a detective hat and a monocle", so I tried that with my own image "Give this human a detective hat and a monocle" and it just gave me this error:

> I wasn't able to generate the modified image because the request didn't follow our content policy. However, I can try another approach—either by applying a filter to stylize the image or guiding you on how to edit it using software like Photoshop or GIMP. Let me know what you'd like to do!

Overall, a very disappointing experience. As another point of comparison, Grok also added image generation capabilities and while the ability to edit existing images is a bit limited and janky, it still manages to overlay the requested transformation on top of the existing image.

It's not actually out for everyone yet. You can tell by the generation style. 4o generates top down (picture goes from mostly blurry to clear starting from the top).

Is anyone else getting wild rejections on content policy since this morning? I spent about 20 minutes trying to get it to turn my zoo photos into cartoons and could not get a single animal picture past the content moderation....

Even when I told it to transform it into a text description, then draw that text description, my earlier attempt at a cat picture meant that the description was too close to a banned image...

I can't help but feel like openAI and grok are on unhelpful polar opposites when it comes to moderation.

One very neat thing the interwebs are talking about is the ghiblification of family pictures. It’s actually pretty cute: https://x.com/grantslatton/status/1904631016356274286

In the coming days, people will Anime all sorts of images, for example historical images: https://x.com/keysmashbandit/status/1904764224636592188

To quote myself from a comment on sora:

Iterations are the missing link. With ChatGPT, you can iteratively improve text (e.g., "make it shorter," "mention xyz"). However, for pictures (and video), this functionality is not yet available. If you could prompt iteratively (e.g., "generate a red car in the sunset," "make it a muscle car," "place it on a hill," "show it from the side so the sun shines through the windshield"), the tools would become exponentially more useful.

I‘m looking forward to try this out and see if I was right. Unfortunately it’s not yet available for me.

You can do that with Gemini's image model, flash 2.0 (image generation) exp.[1] It's not perfect but it does mostly maintain likeness between generations.

[1]https://aistudio.google.com/prompts/new_chat

DALLE-3 with ChatGPT has been able to approximate this for a while now by internally locking the seed down as you make adjustments. It's not perfect by any means but can be more convenient than manual inpainting.

Ditto Instruct Pix2Pix https://www.timothybrooks.com/instruct-pix2pix

Reading other comments in other threads on HN has left me with the impression that iterative improvement within a single chat is not a good idea.

For example, https://news.ycombinator.com/item?id=43388114

Still can't show me a clock that isn't 10:10.

Otherwise impressive.

Is there a technical paper released about the model architecture? Great resolution points to diffusion style generation rather than just token based?

Wow this works really well at editing existing photos.

> we see the photographer's reflection

Am I the only one immediately looking past the amazing text generation, the excellent direction following, the wonderful reflection, and screaming inside my head, "That's not how reflection works!"

I know it's super nitpicky when it's so obviously a leap forward on multiple other metrics, but still, that reflection just ain't right.

Could you explain more? I'm having trouble seeing anything weird in the reflection.

Edit: are we talking about the first or second image? I meant to say the image with only the woman seems normal. Image with the two people does seem a bit odd.

Haha it's almost like early 2000's game logic, the logic is that it's the opposite of the viewpoint then reversed.

In games they did it by creating a duplicate then reversing it, I wonder if this is the same idea.

Like it's predecessor this has most of its utility within the first response, and after that the quality rapidly degrades.

I think it is too biased to use heuristics discovered in the first response to apply the same level of compute to subsequent requests.

It makes me kind of want to rewrite an interface that builds appropriate context and starts new chats for every request issued..

To avoid confusion, why not always use a general AI model upfront, then depending on the user's prompt, redirect it to a specific model?

The models are noticeably different — for example, o1 and o3 have reasoning, and some users (eg. me) want to tell the model when to use reasoning, and when not.

As to why they don't automatically detect when reasoning could be appropriate and then switch to o3, I don't know, but I'd assume it's about cost (and for most users the output quality is negligible). 4o can do everything, it's just not great at "logic".

The fact that it nailed the awkward engineer high five (image 2) is pretty impressive as someone who only gives awkward high fives.

One of the fingers is the wrong way around… it’s a big improvement but it’s easy to find major problems, and these are the best of 8 images and presumably cherry picked.

They still all have a somewhat cold and sterile look to them. Probably that 1% the next decade will be spent working out.

Edit: Please ignore. They hadn't rolled the new model out to my account yet. The announcement blog post is a bit misleading saying you can try it today.

Comparison with Leonardo.Ai.

ChatGPT: https://chatgpt.com/share/67e2fb21-a06c-8008-b297-07681dddee...

ChatGPT again (direct one shot): https://chatgpt.com/share/67e2fc44-ecc8-8008-a40f-e1368d306e...

ChatGPT again (using word "photorealistic instead of "photo"): https://chatgpt.com/share/67e2fce4-369c-8008-b69e-c2cbe0dd61...

Leonardo.Ai Phoenix 1.0 model: https://cdn.leonardo.ai/users/1f263899-3b36-4336-b2a5-d8bc25...

Is the ”2D animation style" part you put at the beginning and then changed an attempt to see how well the AI responds to gas lighting?

The ChatGPT examples don't look like the new Image Gen model yet. The text on the dog collar isn't very good.

What did the prompt look like for Leonard.ai?

I'm curious if you said 2d animation style for both or just for chatgpt.

Edit: Your second version of chatgpt doesn't say photorealistic. Can you share the Leonard.ai prompt?

Yeah, its just not good enough. The big labs are way behind what the image focused labs are putting out. Flux and Midjourney are running laps around these guys

In all fairness you _did_ say 2D animation style

For the first time ever, it feels like it listens and actually tries to follow what I say. I managed to actually get a good photo of a dog in the beach with shoes, from a side angle, by consistently prompting it and making small changes from one image to another till I got my intended effect

The pre-recorded short videos are a much better form of presentation than live-streamed announcements!

I created an app to generate image prompts specifically for 4o. Geared towards business and marketing. Any feedback is welcome. https://imageprompts.app/

It does extremely well at creating images of copyrighted characters. Dall-e couldn't generate images of Miffy, this one can. Same for "Kikker en vriendjes" - a dutch children's book. There seems to be copyright protection at all?

Just curious if it works for creating a comic strip? I.e. will it maintain the consistency of the characters? I watched a video somewhere they demo'ed it creating comic panels, but I want to create the panels one by one.

I believe so! Since it is good at consistency and can be feed reference images, you can generate character references and deed those, along with the previous panels, to the model working one panel at a time.

Interesting that in the second image the text on the whiteboard changes (top left)

It seems this is because the string "autoregressive prior" should appear on the right hand side as well, but in the second image it's hidden from view, and this has confused it to place it on the left hand side instead?

It also misses the arrow between "[diffusion]" and "pixels" in the first image.

> I wasn’t able to generate the image because the combination of abstract elements and stylistic blending [...] may have triggered content filters related to ambiguous or intense visuals.

nah. i pass and stick with midjourney.

So what's the lore with why this took over a _year_ to launch from the first announcement. It's fairly clear that their hand was forced by Google quietly releasing this exact feature a few weeks back though.

I would love to see advancement in the pixel art space, specifying 64x64 pixels and attempting to make game-ready pixel art and even animations, or even taking a reference image and creating a 64x64 version

The easy infographic generation scares me on the implications for society.

It’s pretty good, the interesting thing is when it fails it seems to often be able to reason about what went wrong. So when we get CoT scaffolding for this it’ll be incredibly competent.

So did they deprecate the ability to use DALL-E 3 to generate images? I asked the legacy ChatGPT 4 model to generate an image and it used the new 4o style image generator.

I think you can still use it on Bing.

Can you specify the output dimensions?

EDIT: Seems not, "The smallest image size I can generate is 1024x1024. Would you like me to proceed with that, or would you like a different approach?"

I suspect you can prompt aspect ratios.

I tried a few of the prompts and the results I see are far worse than the examples provided. Seems like there will be some room for artists yet in this brave new world.

It hasn't actually rolled out to everyone yet in chat. You'll have to try it on sora to be sure.

> All generated images come with C2PA metadata

How easy is this to remove? Is it just like exif data that can be easily stripped out, or is it baked in more permanently somehow

Yeah it's just metadata.

My version of the full glass of wine challenge is "clock face with 13 hour divisions". Nothing I've tried has been able to do it yet.

I am blown away by the hyperrealistic renderings, especially of humans. It is getting to the point where I can no longer distinguish ai ones.

The garbled text on these things always just makes them basically useless, especially it often text without being told to like previous models.

you are being served the old model

A real improvement, but it still drew me a door with a handle where the should be one and an extra knob on the side where hinges are.

Can anyone tell me when this will be available in the API? Or is it already available?

I couldn't find anything on the pricing page.

The page says in the following week, which is disappointing. It’s likely we will see openAI favor their own product first more and more, an inversion of their more developer oriented start.

Tried it, the "compise armporressed" and "Pros: made bord reqotons" didn't impress me in the slightest.

Are you sure you were even using the model from the post?

It produces amazing results for me! But the wow effect would have been greater if they had released it a few months ago.

Can someone explain what is going with 4o and Anime with Ghibli Style? Why is it suddenly all over x/twitter?

it isn't Ghibli style in particular, just any style as 4o image gen is much better at maintaining a particular art style, the ghibli ones just stand out due to one tweet that blew up and people followed along

Someone posted a mildly amusing tweet that blew up and went really viral. Now lots of other people are copying the idea.

This dynamic happens on Twitter every day. Tomorrow it'll be a different craze.

fwiw i've added a summary of this discussion here (https://extraakt.com/extraakts/gpt-4o-image-generation-discu...) to keep track of the main points

So Google released Gemini 2.5 and one hour later OpenAI comes with this. It’s almost childish at this point.

its other way around, openai was gonna announce it so Google upstaged

The best thing about this is how the still of the livestream at the bottom is the most uncanny valley image.

saw this thread on X. here are some incredible use cases of 4o image generation: https://x.com/0xmetaschool/status/1904804251148443873

Thankfully. It was outrageous how inferior DALL-E 3 was to any other image generation system.

Could they have switched to *both* image and text generation via diffusion, without tokens?

It is amazing how far text generation in images has come over the past 1-2 years

It bothers me to see links to content that requires a login. I don't expect openai or anyone else to give their services away for free. But I feel like "news" posts that require one to setup an account with a vendor are bad faith.

If the subject matter is paywalled, I feel that the post should include some explanation of what is newsworthy behind the link.

The linked post is not paywalled, did you click something else?

Not a criticism, but It stands out how all the researchers or employees in these videos are non native English speakers (i.e. not American). Nothing wrong with that, on the contrary, it just seems odd that the only American is Altman. Same thing with the last videos from Zuck, if I recall correctly. Especially in this Trump era of MAGA.

The reflections in the whiteboard are all off. Do they address this?

Still can't generate an analog clock face with a given time.

Can it draw the notorious glass of wine filled to the brim yet?

It can indeed

Attention to every detail, even the awkward nerd high-five.

Still seems to have problems with transparent backgrounds.

That's expected with any image generating models because they aren't trained with an alpha channel.

It's more pragmatic to pipeline the results to a background removal model.

EDIT: It appears GPT-4o is different as there is a video demo dedicated to transparancy.

> we’ve built our most advanced image generator yet into GPT‑4o. The result—image generation that is not only beautiful, but useful.

Sorry, but how are these useful? None of the examples demonstrate any use beyond being cool to look at.

The article vaguely mentions 'providing inspiration' as possible definition of 'useful'. I suppose.

question i have is when do we get an opensource version of this form of image generation will we see diffusion models moving to this space

The periodic table poster under "High binding problems" is billed as evidence of model limitations, but I wonder if it just suggests that 4o is a fan of "Look Around You".

This technology should have never existed. Thank you OpenAI for being a contributing factor to destroying politics in future.

And I hope that people who worked on this know this. They are pure evil.

I wish AI companies would release new things once a year, like at CES or how Apple does it. This constant stream of releases and announcements feels like it's just for attention.

Apple held three big keynotes in 2024 plus multiple product announcements via press releases:

May 7, 2024 - The “Let Loose” event, focusing on new iPads, including the iPad Pro with the M4 chip and the iPad Air with the M2 chip, along with the Apple Pencil Pro.

June 10, 2024 - The Worldwide Developers Conference (WWDC) keynote, where Apple introduced iOS 18, macOS Sequoia, and other software updates, including Apple Intelligence.

September 9, 2024 - The “It’s Glowtime” event, where Apple unveiled the iPhone 16 series, Apple Watch Series 10, and AirPods 4.

Via Press releases: MacBook Air with M3 on March 4, the iPad mini on October 15, and various M4-series Macs (MacBook Pro, iMac, and Mac mini) in late October.

Attention is all they need.

Why won't they add benchmarks against o1?

someone try to make an open source version of it. I need to know the inner workings of this cool thing.

Still failing the wine glass test,

https://imgur.com/a/aS8e0UY

Worked for me with a clarification. Looks pretty great, actually. https://imgur.com/a/V8eQWi6

It was easy to fix though, I just said "all the way full" and it got it on the next try. Which makes sense, a full pour is actually "overfull" given normal standards.

that's probably dalle

This to me is the clearest example of why there is no actual physical simulation understanding of the world in these models

that "best of 8" is doing a lot of work. i put in the same input and the image is awful.

literally spent all day playing with this until I ran out of image gen capacity a lil while ago.

so much fun.

Everyone should try running their prompts and see how over hyped this is. The results I get are terrible comparatively.

I don't think the new model is rolled out to all users yet.

This is literally the biggest leap in image generation in a long time.

What is the api price?

This is incredibly impressive, but it's still theft of assets.

drug repurposing for sarcopenia

drug repositiong for sarcopenia

Where are the lawyers where you need them?

well it failed on me, after many tries:

...Once the wait time is up, I can generate the corrected version with exactly eight characters: five mice, one elephant, one polar bear, and one giraffe in a green turtleneck. Let me know if you'd like me to try again later!

Whelp. That's terrifying.

They say it must be an important OpenAI announcement when they bring out the twink.

LPT: while the benchmarks don't show it, chatGPT4>4o. It amazes me people use 4o at all. But hey its the brand name and its free.

ofc 4.5 is best, but its slow and I am afraid I'm going to hit limits.

OpenAI themselves discourages using GPT-4 outside of legacy applications, in favor of GPT-4o instead (they are shutting down the large output gpt-4-32k variants in a few months). GPT-4 is also an order of magnitude more expensive/slower.

Similar to regular LLM plagarism, it's pretty obvious that visual artefacts like the loadout screen for the rpg cat (video game heading) which is inspired by diablo, aren't unique at all and just the result of other peoples efforts and livelihoods.

Garbage compared to Midjourney. I don't even know why you'd market this. It's takes a minute or more and the results are what I'd say Midjourney looked like 1.5 years ago.

Prompt adherence is far, far ahead of midjourney. FAR.

Did they time it with the Gemini 2.5 launch? https://news.ycombinator.com/item?id=43473489

Was it public information when Google was going to launch their new models? Interesting timing.

"Interesting timing" It's like the 4th time by my counting they've done this

OpenAI was started with the express goal of undermining Google's potential lead in AI. The fact that they time launches to Google launches to me indicates they still see this as a meaningful risk. And with this launch in particular I find their fears more well-founded than ever.

Reply on news.ycombinator.com