In the same way that we can set a voice or an adaptive UI, it would be great to be able to tell the R1 what style the default for the Magic Photo should be, instead of always defaulting to studio ghibli.
3 Likes
I agree, I preferred the previous pixel art style. It had its own identity, but with this anime style it has lost its identity 
I know they did it to switch to the new, more powerful model to better reflect reality, but that unpredictability you had before was its beauty, it made you want to take the photo to find out what would come out.
3 Likes
Hopefully this is something they will look at for a future update
1 Like
I agree with all of this, and I’ve provisionally written a PRD (Product Requirement Doc) for Magic Camera V3 to address this exact issue.
It turns out that what was a limitation of the original implementation, for many people, was a feature 
Here’s what happened:
The original implementation wasn’t actually an image-to-image model. What would happen was that the LLM would look at the picture, write a description of the image, and then pass that text description to the image generator. So you could say it was an image-to-text-to-image model.
This had downsides (gender swap, race swap issues) but also upsides (the randomness was delightful).
So when we implemented the new model, it became a full image-to-image model, making it more accurate, but losing the unpredictability or randomness.
So what I’ve proposed is that we re-introduce some randomness back into the system, and I think I have a fairly solid way of doing it. We are pretty strapped on resources right now, so I can’t say when it’ll happen - but just wanted to confirm that I agree with you all!
1 Like
Thanks for the detailed reply and insight! That’s fascinating! Love how a problem becomes a feature! A bit like Dynamic Island for Apple!
What would be amazing would be if we could create a custom prompt for the magic camera in the rabbit hole, a bit like we can already to for magic voice and the magic UI.