ImagingBench · Results

Leaderboard

Unified scores (0–100, higher is better) for six image models across the 20 image tasks, with specialized neural-network baselines where a pretrained model exists.

Protocol
View

Loading…
Click a column header to sort. Bold blue = best model in column. “–” = not available (e.g. all HunyuanImage-3.0 ToF jobs failed).

How the score works

Each signal is normalized to [0, 1]: PSNR / 40 (clipped), 1 − LPIPS, VLM-judge score / 100, and so on. The unified score is a weighted mean of the signals available per instance, with weights fitted to 2,665 blind pairwise human judgments. CGH, ToF depth and Forward outputs use the unweighted mean of four full-reference metrics. Overall is the flat mean over a protocol's image tasks, not the mean of the category columns, and a refused or failed job scores 0.

Models & planners

Each editor uses a planner from its own vendor where one exists. FLUX.2-Pro executes Gemini 3.1 Pro's plans; HunyuanImage-3.0 executes GPT-5's and runs locally as a community INT8 quantization of the 8-step distilled Instruct checkpoint.