Read the text from the old conversation and the new one from _j…

The point is whether “do not change the text, send the prompt as it is” is still necessary. I think it is. We talked about exactly this article back then.

It also means that you cannot infer 1:1 from the billed token count how many tokens were actually used for the image. The tokens used for the re-captioning are also included in the bill.

You said a simple prompt is enough. I speculated that GPT still performs prompt improvement. As I wrote in the text, the data in conversations.json points to that.

So again: GPT probably improves prompts when they are too short, and additionally creates a re-captioning that describes the image as part of the safety system.

That probably proves two things:

1. “Do not change the prompt” can still be necessary for testing, though not for simply creating images.
2. The token count used for the image is not only for the image itself, but also for the re-captioning. If that is complex, the image becomes more expensive. (I don’t have API access, so it could be that the token count shows exactly which tokens are image tokens and which are used for re-captioning. _j can correct me if this argument is wrong.)

I checked this with the data from back then.
conversations.json contains the data GPT creates for conversations, so everything should be in there. You can also see, for example, that the new generator no longer has a seed, or at least there is none in conversations.json. All you see in the post, is form the conversations.json back then.

(but conversations.json is complex. I did not download a new version, it is very large. I used the notes from back then, for the token count argument.)

I really don’t care whatever you have to say about this. I object to you reverse engineering my images no matter what - especially when you are conversing with someone else. Do you do this type of thing often?

At least have the courtesy of letting me know in advance what you want to do with it.

Did the image generator regress? Photo realism is still fine, but anatomy is just gone. Look at the blonde in these examples. Her arms are 5 feet long. One (the one around the brunette’s shoulder) is smaller than the other. Her torso is utterly wacked in the top image and she has six fingers. Am I taking crazy pills or was the generator better at this up until recently?

Hi @Rafaelis and welcome to the community.

To better assist you, a couple of things are needed from you:

• What OpenAI model did you use to generate the image?
• What was the prompt you used to generate the image?
• When did you generate the image?

The OpenAI image verification tool reported the following:

Note: Image verification tool results can produce false negatives for older OpenAI image generation models.

Hi Jeff.

Model used: GPT-5.5 Thinking.

Prompt:

Photorealistic candid lifestyle photo of two young women sitting closely together on a white couch in a bright coastal living room in Newport, Rhode Island. The brunette is based on the reference image: short dark tousled hair, warm brown eyes, natural makeup, soft oval face, calm slightly humorous expression. Beside her is a blonde friend with a natural sun-kissed look, relaxed and affectionate, with her arm wrapped around the brunette’s shoulders. The blonde is based on the reference image: long hair in twin braids, smiling. They are seated casually, close together, looking comfortable and natural rather than posed.

Behind them is a large window with a clear summer afternoon view of Easton Bay, with blue water and several small sailboats visible in the distance. The interior is elegant New England beach-cottage style, with white and cream furnishings, light earthy tones, linen textures, pale wood, and subtle coastal décor. The room feels airy, refined, and lived-in. Natural summer daylight pours through the window, creating soft realistic highlights and gentle shadows.

The image should feel like a real handheld phone or casual DSLR snapshot: authentic skin texture, natural facial proportions, subtle detail, soft depth, true-to-life color, and an intimate candid atmosphere. Keep the composition relaxed and believable.

Style cues photorealistic candid lifestyle photography summer afternoon light upscale Newport beach house natural expressions realistic skin texture soft coastal New England palette

Generated the image a few hours ago.

I suspect the OpenAI image verification tool didn’t detect any OpenAI signals because I’m limited to sharing only one image due to being a n00b, so I used Photoshop to place both in one image. Here’s the original of the top one:

So you used two reference images which I assume were photos, correct?

But, what ImageGen model did you use? GPT-image-1, GPT-image-1.5, or GPT-image-2?

ChatGPT Images 2.0

The reference images are two photorealistic AI generated face portraits.

One issue that a few of us have noticed with GPT-image-2: It can screw-up hands sometimes for no reason

Yes, her right arm is a bit out of whack…

But your prompt is way too complex - too much unnecessary verbiage. Examples that the model does not care about:

Newport, Rhode Island or Easton Bay or New England palette or

The image should feel like a real handheld phone or casual DSLR snapshot: authentic skin texture, natural facial proportions, subtle detail, soft depth, true-to-life color, and an intimate candid atmosphere. Keep the composition relaxed and believable.

Just as an example.

The more simple the prompt, the better. Best to start as simple as possible and then build from there.

Fair enough. Keep prompts simple. But not sure it has any say on anatomy, though. Here’s about as simple as I can get the prompt while maintaining what I want out of it.

Photorealistic candid photo of two young women sitting close together on a white couch in a bright beach house. The brunette has short dark tousled hair, warm brown eyes, natural makeup, and a calm, amused expression. The blonde has long twin braids, a sun-kissed look, and a warm smile. The blonde has one arm naturally around the brunette’s shoulders. Large window behind them with a summer afternoon view blue water, and sailboats. Beach-cottage interior, white and cream tones, linen, pale wood, subtle coastal décor.

Result:

Six fingers again. Arm a bit more acceptable. But see image in next post. Upper arms are way too long. Anatomy is off in all of these.

In this example, her elbow would bend right behind the brunette’s right ear, making the upper arm unnaturally long. If she stood up, her hands would reach her knees.

Damned fingers…

Prompt is much better. Maybe include “five fingers only” and “correct anatomy” in a sentence?

I did try it. Telling it to use believable anatomy and proportions in one prompt and in another Keep anatomy natural, especially arms, shoulders, hands, and torso proportions.

GPT has never been great at amodal completion. No generative models are, which explains why an arm going behind an occluding object, like the brunette’s head in these examples, can get all out of anatomical whack. But I find it especially bad in 2.0. And the finger issue is ‘new’. Overall, 2.0 feels like a regression in realism compared to 1.5.

I think you are right. Glad you brought this up - maybe imageGen Dev will see this…

Yes, topology is still a source of errors, and that’s exactly why it’s difficult to fix through prompts alone. The editing tools may help in some cases.

It will still take some time before the models understand topology better and produce even more realistic images.

Another recommendation is to build prompts step by step, adding one detail at a time, so you can identify which elements are causing the errors and then remove those specific parts. Unfortunately, there’s no real alternative to trial and error. Creating variations can also help, as can changing the order of the prompt. GPT has become quite good at helping with that.

The internal workings of the models are not publicly documented, but this model also appears to assemble images from individual parts. Those parts may work well within a local section, yet fail in the overall composition, probably because the model does not yet fully understand the image as a whole.

With photorealism, there’s also a psychological factor: the Uncanny Valley effect Uncanny valley - Wikipedia. People have a very precise understanding of what reality is supposed to look like, so even small deviations tend to stand out. Stylized images on the other hand, establish their own visual style that people can easily accept as a different kind of reality. Achieving truly perfect photorealism will likely require quite a bit more development time.