Back to Discussions
Help· edited

GPT Image 2 stop messing around

日日本及大和號戰艦的应援会™
Translated to English·

While using AIGAZOU (GPT Image 2), I wondered why 1K output seems to have higher prompt understanding and fidelity compared to generating in high quality like 4K. What is the technical reason behind this? 20 credits please return.
Supplement
After verifying multiple times under different environments and conditions, I found that with AIGAZOU (GPT Image 2), 1K output clearly tends to have higher fidelity to prompt details than 4K output. I speculate that when computational resources are allocated to high resolution and detailed rendering, the processing dedicated to accurately interpreting text information and maintaining composition decreases as a trade-off, making deviations from instructions more likely. Does anyone know the technical background of this inverse relationship between resolution and prompt understanding from the perspective of AI processing models?
Supplement to the supplement
When outputting at high resolution (4K), the model's computational resources (attention mechanism) are heavily pulled toward generating vast pixel spaces and completing details, resulting in a relative decrease in the weight for maintaining text embedding (linguistic understanding of the prompt).

4 replies

ま
まれん2·

#Fictional Anime
Normal↓
https://muryou-aigazou.com/ja/gallery/4yblj0s1c8wd
4K↓
https://muryou-aigazou.com/ja/gallery/ovm8mrjvf7o4
I have also compared them in the past.
In one word, the 4K version is beautiful, while the normal version has more information.
But if you look closely, in the 4K version, both characters have six fingers each. Can you tell?
This result can be related to the speculation that the supporter mentioned: "a trade-off between information and quality."

日
日本及大和號戰艦的应援会™OP·

Thank you for sharing the comparison results!

It's very interesting that while the 4K version excels in visual sharpness, errors can be seen in the number of fingers.
With the standard version and the 4K version, simply having a higher resolution doesn't mean all depictions will be accurate—it seems that detail sharpness and shape accuracy may be separate issues.

The idea you mentioned, "the trade-off between information volume and quality," could also be related to this. However, since the cause cannot be determined from this single result, repeating the comparison multiple times with the same prompt could make for an even more intriguing investigation.

Thank you for sharing such a valuable comparison example! It is very informative.

👍1
O
Otakosan·

Or rather, am I the only one who feels that GPT Image 2's generation ability has definitely declined compared to before? The character drawings are badly distorted, there are many bugs, the composition is half-baked, and now you have to make at least four images (32 credits) to get a decent picture. The sense of satisfaction when you get a good result has definitely dropped compared to before. That's why lately I've been hesitating to use GPT Image 2.

O
Otakosan·

I asked the AI about the cause, and it said something about a problem occurring when OpenAI updated to version 2.5, but I don't know how true that is.

Leave a reply

Join the community to share your thoughts and ideas.

Sign in to join the discussion