i think the mask is not used at all, we almost get similar results without a mask
Yep, same here. The feature is completely useless with a soft mask. Who came up with that?
Even their example is pointless, you can just say “put a flamingo in the pool” and it will give the exact same result
I think it’s because the underlying tech is different. DALLE was pure image model, and 4o is multi-modal… Pretty sure it is a soft mask. I’m trying to get more details, so stay tuned…
You don’t have to be pretty sure, one api call and you are 100% sure. The whole image gets regenerated
2 Likes
Yeah, I’m not sure if the masked area is weighted or what, though. Like I said, trying to get an OpenAI employee to give us some more details if possible.
I wonder if they would consider adding an edit endpoint for DALLE3 in the meantime… I’m still using that model for quite a bit personally.
1 Like
since it’s calling the same underlying image endpoint as ChatGPT, lessons learned from there apply (you’re just providing your own prompt, which ChatGPT does for you, but the tool invocation has been identical for me).
The mask doesn’t do anything, and in my experience, NOT providing a mask, and instead telling the generator what to change, and to change nothing else, will frequently result in better outcomes than even providing a mask. It’s not built like inpainting at all.
I know this is “API Forum” but this particular tool is different. Both ChatGPT and Image-1 appear to call the same underlying remote tool, so anything we learn on one is applicable to the other:
2 Likes
I’m echoing others in this thread. I’m also experiencing issues with the Image Generation API when providing a mask. It seems the mask is being ignored and the entire image is being edited.
3 Likes
Let’s be honest: this isn’t a bug — it’s false advertising.
OpenAI’s API does not currently support true inpainting, despite what the documentation claims.
According to the docs, only the transparent areas of the mask should be modified, while the filled areas remain unchanged. But that’s clearly not what happens.
The official examples use small images where the issue is easy to miss — but if you view them at full resolution, the problem becomes obvious.
Even in the official example, the entire image is regenerated — not just the masked (transparent) area — directly contradicting the claim that “the filled areas will be left unchanged.”
1 Like