Rapidata

Gemini Omni Flash 1.2 is #1 on both camera benchmarks: Relative Camera Movement, published yesterday (moving through the scene, "between the doors, up to the lantern"), and now Camera Movement (copying an exact move like "pan 40° left and back" from a reference clip).

Some match-ups below:

Gemini Omni Flash 1.2 debuts at #1 on our Relative Camera Movement benchmark! 🥇

Backed by 900K+ human judgments, this benchmark evaluates how well models navigate from a single image to scene-relative viewpoints, like behind a sofa, into the driver’s seat, or from a dog’s perspective.

Ranking, votes, and outputs are transparent on benchmark.ai/relative-camer…

Our human preference datasets just crossed 400K Downloads on our @huggingface 🥳.

GPT Image 2.6 is here and as to expect from the @OpenAI Image team it is right at #1 based on over 58'000 human responses collected in less than an hour! No overlapping error bars.

Check out full results: benchmark.ai/image
All data public on @huggingface as always.

The best models are the worst at diversity: the better models get at prompt alignment, the less diverse their outputs tend to be (r = −0.77).

Qwen and SeeDream seem to be different in their post-training method, being able to maintain diversity.

Compare outputs and explore the study at benchmark.ai/output-diversi…