HOI-Edit - Spatial Understanding: leaderboard
HOI-Edit (2026) asks image editing models to change human-object interactions, such as picking up, using or moving an object, and scores the edit with grounded questions to Gemini 2.5 Pro about tracked human and object regions. This board is the 0 to 100 interaction score, zeroed when a spatial check fails, on 202 edits that need the right target and destination; higher is better.
Metric: Interaction score gated by a context question (0-100), zero when the spatial check fails, on the 202 Level-2 edits (pick the right target among similar ones, place it at the right destination), scored by HOI-Eval: the human and object boxes are tracked into the edited image and Gemini 2.5 Pro rates the interaction from tagged crops; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nano Banana | 59.6 | |
| 2 | Wan2.2-I2V-A14B | 55.26 | |
| 3 | Qwen-Image-Edit PLUS | 49.28 | |
| 4 | Bagel | 47.81 | |
| 5 | ChronoEdit | 46.08 | |
| 6 | Step1X-Edit | 40.58 | |
| 7 | FLUX.1 Kontext [dev] | 27.8 | |
| 8 | ByteMorph | 22.4 |
Interactive version: theaggregate.ai/benchmark?slug=hoi-edit-spatial-understanding · How It Works · Data refreshed daily, snapshot 2026-11-13.