---
license: cc-by-sa-4.0
pretty_name: Multi-step Image Edit (identity drift)
tags:
- image-editing
- human-preference
- benchmark
- identity
size_categories:
- 10K<n<100K
---

# Multi-step Image Edit: identity drift

Human preference data from Rapidata's [Multi-step Image Editing benchmark](https://www.benchmark.ai/multi-step-image-edit) on benchmark.ai.

Each photograph of a real person gets a chain of five everyday edits on clothing, objects or text in the scene. None of them concerns the face. Every model edits its own output from the step before. Annotators then see two models' faces after the fifth edit next to the original face and pick the one that looks more like the original.

The benchmark page has the method, the settings of every model and the standings.

## Columns

| Column | Content |
| --- | --- |
| `reference_face` | The original face, shown to annotators as the reference |
| `reference_image` | Link to the full reference photograph (PNG) |
| `step_1` … `step_5` | The five edit instructions |
| `prompt_tags` | Edited items, people in the photo, face size and photo shape |
| `image1`, `image2` | The two models' faces after the fifth edit |
| `model1`, `model2` | The models that made them |
| `weighted_results_image1_identity_drift_at_step_5`, `weighted_results_image2_identity_drift_at_step_5` | Votes for each face, weighted by annotator reliability |
| `detailed_results_identity_drift_at_step_5` | Every vote with the annotator's country, language, demographics and reliability score |

## Licence and credits

The photographs are by their authors under CC0, CC BY or CC BY-SA licences. The faces and edits made from them are shared under CC BY-SA 4.0. Every photograph's author, licence and source page are listed in `ATTRIBUTION.md` and `metadata/attribution.csv`.