An example of reasoning in 'pixel space':

Notably we pass the 'room without an elephant' test (medium.com/@avanib28264...)

Imagen 3 (deepmind.google/technologies...) is now the top ranking model on the lmsys image generation arena, by a significant amount. Proud to have been part of the team that built it (and there's even more to come soon !).

Surprisingly effective. The problematic parts are changed but everything else remains the same in the fine-tuned model. This is different from an editing model, where 2 rounds of inference are needed to fix the problematic parts.

Then you fine-tune using a combination of DRAFT (arxiv.org/html/2309.17...) and our custom region-aware fine-tuning objective.

you generate a heatmap highlighting the problematic region (e.g. using our previous work on Rich Human Feedback for T2I): arxiv.org/pdf/2312.10240

The idea is simple. If the image from the base model has a region that's (say) NSFW: