s-aniruddha's picture
Card, photo credits, model settings, standings and settings evidence for the 19 ranked models
23ec548 verified
|
Raw History Blame
5.59 kB
---
license: cc-by-sa-4.0
pretty_name: Multi-step Image Edit (identity drift at step 5)
tags:
- image-editing
- human-preference
- benchmark
- identity
size_categories:
- 10K<n<100K
---
# Multi-step Image Edit: identity drift at step 5
Human-preference benchmark by [Rapidata](https://www.rapidata.ai) on whether a person still looks like themselves after
five everyday edits that never touch the face. 19 image-editing models, 199 photographs of real people, 292,017 votes.
## What it measures
Each photograph gets a chain of five instructions about clothing, objects or text in the scene. Step 1 edits the
photograph; every later step gets only the model's own output from the step before as its input image, with that step's
instruction, so the model never sees the original again. An automatic check (`pipeline/judge_prompt.txt`) confirms that
each edit was made; a step has three attempts, and a model is ranked only if at least 80% of its chains reach step 5.
Annotators then see two models' faces after the fifth edit, with the original face as reference, and answer "Which
**edited face** looks **more similar** to the **original face**? Ignore changes to hair, clothes and background." Every
pair of models meets in 320 matches on photographs both completed; each match is decided by the majority of its votes,
and the matches give Bradley–Terry scores.
## How the images were made
Every setting is the lab's own API default, sent explicitly, except: highest quality, reasoning, thinking and precision;
watermarks off; the least restrictive moderation that can be sent; a fixed seed per model, photograph, step and attempt
where the model takes one. Where a model can set the aspect ratio or keep the input's, that is set and the resolution
stays at its default; Wan 2.7 Pro and the Qwen models get the input's exact size. Each model returns the format its lab's
API returns by default (JPEG for Seedream 5.0 Pro and Lite and Grok Imagine Image 2.0, WebP for Muse Image, PNG for the
rest), and each output is passed to the next step unchanged. Every setting sent per model, and why, is in
`metadata/models.json`; the evidence for the defaults is in `pipeline/fal_schemas/` (fal's endpoint definitions) and
`pipeline/provider_docs/` (the labs' API docs as saved on 2026-10-30, with Wayback Machine snapshot links in
`fetch_log.tsv` and `wayback_retry.tsv`).
## Standings
| # | Model | Score | Called through | List price per edit |
|---|---|---|---|---|
| 1 | Ideogram 4.5 (High Quality, High Precision) | 1901 ±33 | fal `ideogram/v4.5/edit` | $0.22 |
| 2 | FLUX 3 Image | 1742 ±29 | fal `blackforestlabs/flux-3/edit-image` | $0.048 |
| 3 | Muse Image | 1490 ±24 | fal `meta/muse-image/edit` | $0.01 |
| 4 | Qwen-Image-2.1-Pro (Thinking) | 1373 ±23 | Alibaba Model Studio `qwen-image-2.1-pro` | $0.04 |
| 5 | Seedream 5.0 Pro | 1247 ±23 | fal `bytedance/seedream/v5/pro/edit` | $0.09 |
| 6 | MAI-Image-2.5-Pro | 1209 ±22 | fal `microsoft/mai-image-2.5-pro/edit` | $0.117 |
| 7 | Qwen-Image-3.0-Pro (Thinking) | 1079 ±22 | Alibaba Model Studio `qwen-image-3.0-pro` | $0.043 |
| 8 | GPT Image 2.6 Sunburst (Max Quality, High Reasoning) | 1061 ±22 | fal `openai/gpt-image-2.6/sunburst/edit` | $0.18 |
| 9 | Nano Banana Pro | 1054 ±22 | fal `fal-ai/nano-banana-pro/edit` | $0.148 |
| 10 | Nano Banana 2.1 (High Thinking) | 984 ±22 | fal `google/nano-banana-2.1/edit` | $0.043 |
| 11 | Nano Banana 2 (High Thinking) | 955 ±22 | fal `fal-ai/nano-banana-2/edit` | $0.071 |
| 12 | Qwen-Image-Edit-2511 | 954 ±22 | fal `fal-ai/qwen-image-edit-2511` | $0.03 |
| 13 | GPT Image 2.5 Sunburst (Max Quality) | 944 ±22 | fal `openai/gpt-image-2.5/sunburst/edit` | $0.18 |
| 14 | Grok Imagine Image 2.0 (Medium Quality) | 888 ±22 | fal `xai/grok-imagine-image/v2.0/edit` | $0.04 |
| 15 | Wan 2.7 Pro | 716 ±23 | fal `fal-ai/wan/v2.7/pro/edit` | $0.075 |
| 16 | GPT Image 2.6 Flare (Max Quality, High Reasoning) | 613 ±24 | fal `openai/gpt-image-2.6/flare/edit` | $0.18 |
| 17 | GPT Image 2.5 Flare (Max Quality) | 594 ±24 | fal `openai/gpt-image-2.5/flare/edit` | $0.18 |
| 18 | Kling Image O1 | 399 ±27 | fal `fal-ai/kling-image/o1` | $0.028 |
| 19 | Seedream 5.0 Lite | −203 ±44 | fal `bytedance/seedream/v5/lite/edit` | $0.035 |
## Files
| Path | What |
|---|---|
| `data/` | one row per matchup: the two models (`model1`, `model2`), their face crops after the fifth edit (`image1`, `image2`), the reference face crop (`reference_face`), the whole reference image (`reference_image`), the five instructions (`step_1` … `step_5`), and the votes, weighted and individual |
| `metadata/photos.jsonl` | per photograph: the five instructions with their kind and the item each edits, people visible, tags, face box, how the reference image was made, source page, creator and licence |
| `metadata/models.json` | per model: endpoint, every setting sent and why, list price and its basis |
| `metadata/standings.json` | the standings with confidence intervals |
| `metadata/attribution.csv`, `ATTRIBUTION.md` | photo credits |
| `pipeline/` | the step-check prompt, fal's endpoint definitions and the labs' API docs |
## Credits and licence
The photographs come from Wikimedia Commons (179) and from StockSnap and Rawpixel through Openverse (20): 106 CC BY-SA
4.0, 57 CC BY-SA 3.0, 1 CC BY-SA 2.0 fr, 8 CC BY 4.0, 27 CC0. The creator, licence and source of each are in
`ATTRIBUTION.md` and `metadata/attribution.csv`. The edited images are adaptations of these photographs; this dataset is
released under CC BY-SA 4.0, and the photographs keep their own licences.