Hi,
Today in the second half of the day, I was generating a lot of images in Sora and noticed something strange. At some point, the generation process suddenly became 3–4 times faster than usual. Typically, generating two images takes about 1.5 to 2 minutes, but now they started completing in just 20–40 seconds.
The problem is that the quality of these faster-generated images is significantly worse than what I usually get. This concerns me because there was no announcement from the developers about any changes, and I absolutely do not want lower-quality images. I would rather wait longer if it means getting higher-quality results.
To illustrate the issue, I’ve uploaded a screenshot. All four images were generated using the exact same prompt. The top two images were generated at the usual pace, and the bottom two were generated much faster. The difference in quality is immediately obvious.
I understand that servers may be under heavy load, but introducing such a method without notice feels disrespectful to users. I deeply respect OpenAI for providing the ability to generate high-quality images for just $20 a month—that’s amazing. But I would really prefer not to encounter these kinds of unpleasant surprises.
Is this behavior intentional or a bug? I’d appreciate any clarification.
Yes, but not always—only when I’m giving prompts that aren’t very complex. Still, even when I provide simple prompts, I want to get the best result, not something low-quality.
I have noticed that generation time is massively tied to prompt complexity, that’s nothing new.
Addendum - I don’t think that’s a lower quality thing. In the API, if you lose the quality, you don’t just get a little worse, it’s utter trash. The model can’t generate “slightly lower quality”. Medium is awful. Low is like .. blobs.
I signed up for Plus and happily generated dozens of very high-quality images, and then suddenly since last night all of my images are coming out potato-quality.
No idea why this happened or when or how I can get high-quality images back. ChatGPT cannot offer any insight. Be nice if there was a way to get some answers.
Hi!
Sorry to drop in just like this, I had this issue a bit, too, but not much, sometimes, yes, sometimes, no.
It seems it’s slowly improving at least for now! (also for me quality is really difficult for myself to see, maybe, it’s just the prompt that was a bit changed (I love to change my prompts so ) Maybe I changed 1 prompt to something very similar, and thought - 1 week later - I hadn’t changed it. I don’t know! I do want to say this to OpenAI and the community:
Thanks a lot for everything you all do! Really; Warmest wishes!
quick edit I’m still new so, point out any fuzzy/wrong wording!
Sorry for a bit long post, I wish everyone a beautiful day! Because You who reads this deserves that!
Jane
After some help I got from OpenAI by mailing, It did start to:
Progress
It Generates faster now (I don’t know, maybe it really is the so called during peak hours cuz.. I do LOVE to Use Sora!
It Generating More higher quality more often, from now at least
2.2 I also wondered if the change in quality might be because Sora has been trained on a much larger dataset? Maybe now it’s pulling from a wider mix of high- and low-quality data, so 50%+ and 50%+ for bad and good respectively? so it’s higher chance to get both good and bad now? what led me to this "thought": I noticed that an old prompt (like a lake or river) used to give really beautiful results, but now it sometimes looks flatter or more “clay-like,”even though I haven’t changed the prompt.
That’s what made me think of the “more data = more variance” idea in the first place.
TLDR
delays has been improving since a few days, probably was peak hours in my case.
quality has been improving, maybe Sora got “improved/changed” or learned on more (good+bad) data so it maybe has more chance to give bad & good results
Thanks for reading! If I can improve, do tell me! or, if I have written something completely downright wrong, please tell me that also
TLDR of what I “did”
I didnt do much, I waited after contacting support
(Oh and I try to make it easy for readers to see what I had for problem and what solution I tried although here it was just 1. contact support after waiting a few days,2. wait
I’m happy that I’m not the only one who’s been experiencing a massive decline in image quality. I tend to generate images of plants (almost entirely seedlings). They used to look great, but now, it doesn’t look right. Here’s an image I generated from July 11th.
It looks awful, with weird looking anatomy. The worst part is that the imagery just keeps getting worse. Even when I started using more advanced descriptions, it still doesn’t look like what I want it to do almost all of the time. If there’s anyone who works for OpenAI, please help fix the quality issue, even if that means expanding generation time a bit. I’d much rather wait 2 minutes for a great image than 30 seconds for a mediocre image.
thank you so much for replying and showing; and yes i agree kind of; and it has this clay like thing i was also noticing; idk if clay like is right word,also writing this on 24h no sleep but, i hope sora gets better, and take care!
) Maybe I changed 1 prompt to something very similar, and thought - 1 week later - I hadn’t changed it. I don’t know! I do want to say this to OpenAI and the community:
Because You who reads this deserves that!