Editing any part of an image degrades the quality of the entire image.
Try it yourself. Generate a scene with detailed textures, such as walls, mountains, grass, etc. Then ask ChatGPT to edit a small part of the image, like adding or changing some text. Compare the 2 images: The next image will have reduced detail in the textures, as if a noise reduction/smoothing/blurring filter was applied to it. This degradation compounds (gets worse) with every successive generation, and it makes the editing features useless.
Exact resolutions requests are ignored. ChatGPT should warn us when a requested resolution cannot be generated exactly.
Hands, feet and limbs are still sometimes disfigured and distorted. Sigh.
Intermediate/temporary iterations are sometimes better than the final output.
For example, when I ask for an infographic or UI mockup etc, the chat and ChatGPT’s local working folders sometimes contain images while the chat is still running. Those look pretty good. But then when the chat ends, it offers me an output that is worse than the previous intermediary image in some ways, like the hands and arms in symbols of people turn into weird arrows etc.
Could this be caused by some plugins? I only use a few OpenAI-made plugins, including “Creative Production” and “Product Design” but my prompts never explicitly request their use.
I hope these flaws are not intentionally neglected or given a lower priority, to make us spend more tokens
I saw this problem since the first day and i always ask the model/agent to make a new separate cut and paste into the original image. This way it retains pixel-perfect on the original image.
Regarding resolution, yes it’s 100% ignored no matter how you ask, at least in my experience. Only work-around i can think of is creating 4 separate images and asking it to do a montage, but this adds a lot of complications and power consumption, so I’m still thinking if i should even bother.
I wonder if one day we could make an image higher than 1536px (the 4K the agent says it’s possible is a lie).
Exactly. Why offer “local editing” when the image gets mangled beyond usability, accuracy, and aestethic value?
The announcement of 2.5 made it sound like this issue was finally resolved, but I genuinely cannot see a difference myself.
If the chair I wanted to be moved further to the left was sucessfully moved and even got to keep its identity, there will still be the issue of:
The character occupying the same image has now lost established facial identity, the image itself its overall fidelity and unique interpretation of details. Color values and even small enviroment details have been contaminated by and grouped into a visible “noise” pattern.
The lack of inpainting turns the existing flaws into a dealbreaker for me.
It’s just an extremely inefficient way to work and needlessly taxing.
I am extremely frustrated with the image generation on ChatGpt. Today to most of my requests to edit or create images, here has been the reply :“We’re so sorry, but the image we created may violate our guardrails around violence. If you think we got it wrong, please retry or edit your prompt.” it doesn’t edit the images it created itself !!! and I have the Pro service !!!
