FED-Bench (Simple Instructions): leaderboard

Metric: FED-Score (x100): product of the alignment (mean of instruction consistency and expression match to the ground-truth image, both judged by Gemini-2.5-Pro), fidelity (mean of ArcFace identity similarity, background error outside the face and judged perceptual quality) and relative expression gain (face LPIPS change over the ground-truth change, a Gaussian penalty centred on 1 that punishes too little and too much change) scores, over FED-Bench's 747 in-the-wild face editing triplets (source image, instruction, ground-truth target) across seven basic expressions, default inference settings; simple instructions such as change the expression from neutral to happy; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 17 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=fed-bench-simple-instructions · How It Works · Data refreshed daily, snapshot 2026-10-11.