Zurück zu Diskussionen
Hilfe· bearbeitet

GPT Image 2, hör auf zu scherzen.

日日本及大和號戰艦的应援会™
Übersetzt in Deutsch·

While using AIGAZOU (GPT Image 2), I wondered: it seems that 1K output has higher prompt comprehension and fidelity than generating in high quality like 4K. What is the technical reason for this in terms of the mechanism?
20 credits please return・…•
Supplement
After verifying multiple times under various environments and conditions, 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 for accurate interpretation of text information and maintaining composition decreases as a trade-off, making deviations from instructions more likely. Does anyone know the technical background from the perspective of AI processing models regarding this inverse relationship between resolution and prompt comprehension?
Supplement to the supplement
When outputting at high resolution (4K), the model's computational resources (attention functions) are heavily pulled toward generating vast pixel spaces and complementing details, potentially causing the weight for maintaining text embedding (linguistic understanding of the prompt) to relatively decrease.

4 Antworten

ま
まれん2·
Fiktionale Anime

Normal ↓
https://muryou-aigazou.com/ja/gallery/4yblj0s1c8wd
4K ↓
https://muryou-aigazou.com/ja/gallery/ovm8mrjvf7o4
Ich habe auch schon früher Vergleiche angestellt.
Um es kurz zu fassen: 4K ist schöner, Normal bietet mehr Informationen.
Wenn man jedoch genau hinsieht, haben beide Personen in der 4K-Version sechs Finger. Erkennen Sie das?
Dieses Ergebnis lässt sich mit der von der Unterstützergruppe geäußerten Vermutung des „Trade-offs zwischen Information und Qualität“ in Verbindung bringen.

日

Thank you for sharing the comparison results!
It's very interesting that the 4K version excels in visual sharpness, yet shows errors in the number of fingers.
Between the standard version and the 4K version, it seems that simply having higher resolution doesn't guarantee accuracy in all depictions—clarity of detail and precision of shape may be separate issues.
Your point about a "trade-off between information volume and quality" might also be relevant here. However, since we can't determine the cause from this single result, it would be fascinating to run multiple comparisons with the same prompt.
Thank you again for sharing such a valuable comparison—it's very informative!

O
Otakosan·

Actually, am I the only one who thinks that recently the generation ability of GPT Image 2 itself has definitely declined compared to before? The distortion in figure drawing is terrible, there are many bugs, the composition is half-baked, and now I have to make at least four images (32 credits) to get a decent result. The satisfaction when it’s done has definitely dropped compared to before. So lately, I’ve been hesitant to use GPT Image 2.

O
Otakosan·

Ich habe die KI nach dem Grund gefragt, und sie sagte, dass es ein Problem gab, als OpenAI auf Version 2.5 aktualisiert hat, aber wie viel davon wahr ist, weiß ich nicht.

Antwort hinterlassen

Treten Sie der Community bei, um Ihre Gedanken und Ideen zu teilen.

Melde dich an, um an der Diskussion teilzunehmen