[Updated] One image, infinite re-rolls
loading
Comments
Endless is brand new and I'm still polishing it. Found a bug or got a feature idea? Drop it in the comments, I read every one and I'm actively iterating on this. 🙏
It sounds beneficial for normal models, but gpt-image-2 is an oddball and so it probably won't work for that one in particular. I find that changing the prompt a little does nothing and regenerating with the same prompt hardly changes anything at all, in most cases. These super modern models appear to rewrite your prompt for you, so the details of what you type is partially ignored. I end up with very similar images on different seeds, unlike normal models. At least in my experience. If you find that it works well with that model that would be great but I've had to vary my prompts a lot to get changes out of it. Personally, I would like it if online generators had a draft mode. A quick, cheap, low-cost preview would be real handy for experimenting. It's what I do with offline models: set the step count low and blitz 16 at a time. (On that note, gpt-image-2's low quality mode is very close to medium and high in some cases so that may be a useful option at some point).
I think it depends on how impressionist or subjective your prompt is. things that are open to interpretation can change from Ai slop to cohesive image from gen to gen with the same wording
Thanks for the feedback. You're right, this makes more sense with SDXL and Z-Image type models. gpt-image-2 does its own thing with the prompt rewriting, so changing the seed barely moves anything. One trick that can help get more variation: you can write prompts with alternatives using [option1|option2|option3], and it'll pick between them on each run. Noted on the quick preview / draft mode. It's on my list, just not a priority right now.
- Comment deleted
