It seems some new update might have taken place concerning how pictures files are shown in chats. I cannot say if that update is related, but now when I ask chatgpt to generate a new image based on the one I just uploaded, it now generates something entirely different instead of referencing the picture I uploaded. Everything was working fine until this new update or whatever appeared and I am using my own content, not a copyrighted character.
Okay, I posted this elsewhere, but I’ll quote myself as I discovered this along with another user:
'Yes! I also noticed that, too. Chatgpt recommended that to me, itself. I noticed that after going to the library, uploading the image there and then asking for updates made the image upload like before in the chat: So, one big image instead of showing me a tiny image with the file name next to it.
Edit: Okay…wait, I just uploaded a picture without going through that library process and the picture uploaded like before. I need to test if the model anchors to the picture.
Edit: I just tested it in regular chat, just uploading the picture in there and asking for an edit…now it seems to be working like before, again. The model referenced the uploaded picture and generated based on my instructions instead of making a brand new character. I do see new options for editing the picture when I click on it directly.’
I’m experiencing a very similar issue recently.
I create four-panel comics in ChatGPT using fixed reference images for recurring characters. I also provide detailed specifications for their faces, hairstyles, clothing, age, and body type.
This workflow used to work considerably better, but recently the generated characters seem much less faithful to the reference images.
In my latest example, there are only two characters, and all four panels were generated together as a single image. Even then, the same characters’ faces drift noticeably from panel to panel. Their general features such as glasses, hairstyle, and clothing may remain, but their facial structure, apparent age, eyes, and jawline change.
So I also suspect that something may have changed in how uploaded reference images are being used during image generation, rather than this being only the usual difficulty of character consistency.
I’m attaching an example because the drift can be seen even within a single four-panel image.
Additional report from Japan: uploaded images are understood correctly, but image generation ignores them
Location: Japan
Observed: August 29, 2026, approximately 14:50–17:00 JST
I observed what appears to be a serious problem in the hand-off between ChatGPT’s image understanding and image generation/editing.
What I observed
-
I uploaded four real exterior photos of a ramen restaurant and asked ChatGPT to create a corrected/front-facing exterior image based on them.
Instead of editing or referencing the restaurant photos, it generated a completely unrelated black-hole infographic.
A subsequent attempt generated a photosynthesis infographic.
-
To determine whether this was related to a Project’s custom instructions or context, I repeated the test in a new chat outside the Project, using the same restaurant photos.
This time it generated an invented ramen restaurant interior that looked nothing like the uploaded exterior.
-
I then tested with only one exterior photo.
Again, it generated a fictional ramen restaurant interior from another angle rather than using the uploaded image.
-
I also tested with an actual interior photo.
The result was another unrelated, AI-generated ramen restaurant interior, including fabricated restaurant names, prices, signs, and other details that did not exist in the source image.
-
I tested a separate infographic workflow using real photographs as references.
The generated infographic:
-
did not use the actual photographs,
-
ignored an explicit 1:1 square aspect ratio instruction,
-
ignored an explicit colored-pencil illustration style instruction,
-
and instead created its own unrelated visual content.
-
Important control test: image understanding itself appears to work normally
Using eight photographs from the same restaurant, I asked ChatGPT only to analyze them, without generating an image.
The analysis was very accurate. It correctly extracted, among other things:
-
restaurant name: 神儺祁 / TSUKEMEN JINDAGI
-
menu prices
-
noodle serving sizes
-
text from signs and the ticket machine
-
exterior characteristics
-
interior characteristics
-
food presentation
-
instructions explaining how to eat the curry tsukemen
This suggests that ChatGPT can see and understand the uploaded images correctly.
The failure appears specifically when the task is handed over to image generation/editing.
Reproduction pattern
The problem reproduced with:
-
multiple reference images,
-
a single reference image,
-
exterior photographs,
-
interior photographs,
-
a Project chat,
-
a new chat outside the Project,
-
image-editing requests,
-
and infographic-generation requests.
In every generation test, the source image was either ignored or replaced by invented visual content.
Current hypothesis
From the user side, this looks less like an image-recognition failure and more like a failure somewhere between:
ChatGPT image understanding → generation context/reference-image hand-off → image generation model
The text/vision side appears to understand the source images, but the generation side behaves as though it has received a generic text-to-image request instead of the actual reference image and current instructions.
I am reporting this because it appears reproducible and is severe enough to make reference-image editing unusable.