# AI use cases for Edit Drift (MegaTroll222/edit-drift)

Repo: [MegaTroll222/edit-drift](https://github.com/MegaTroll222/edit-drift) | Use cases: 3 | Page: https://dedupedai.com/use-cases/megatroll222-edit-drift/

Edit Drift is a chained-edit benchmark and scorer that measures how much image models alter the untouched parts of a photo across 40 consecutive edits, with a CLI and a Claude Code plugin.

This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL.

- Score how much an image model changes what it shouldn't: https://dedupedai.com/use-cases/megatroll222-edit-drift/score-image-model-edit-fidelity/
- Compare image models with a leaderboard and video: https://dedupedai.com/use-cases/megatroll222-edit-drift/compare-models-side-by-side/
- Run the benchmark from Claude Code: https://dedupedai.com/use-cases/megatroll222-edit-drift/run-edit-drift-from-claude-code/

## How to score how much an image model changes what it shouldn't with Edit Drift

Repo: [MegaTroll222/edit-drift](https://github.com/MegaTroll222/edit-drift) | Difficulty: Advanced | Time: 2 hours | Basis: From the README

Run the 40-edit benchmark with your model and compute its Edit Fidelity Score.

**Who it is for:** An ML engineer picking an image editing model for a product.

**The problem:** Models quietly alter areas they were told to leave alone.

**The outcome:** Per-step drift and an Edit Fidelity Score in scores.json.

### Steps

1. Install with `pip install git+https://github.com/MegaTroll222/edit-drift`
2. List the edits with `editdrift steps  benchmarks/milk`
3. Generate each edit with your model, applying it to the previous result
4. Score with `editdrift score  benchmarks/milk results/milk/my-model`

### Prompt to paste into your agent

```text
Using MegaTroll222/edit-drift, run the milk benchmark edits with <model>, score them, and report the Edit Fidelity Score and final drift.
```

### Caveat

The score averages all 40 steps, so it can rank models differently from final drift alone.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/megatroll222-edit-drift/score-image-model-edit-fidelity/

## How to compare image models with a leaderboard and video with Edit Drift

Repo: [MegaTroll222/edit-drift](https://github.com/MegaTroll222/edit-drift) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README

Score each model's chained edits, build a leaderboard with drift curves, and render a side-by-side MP4 to share.

**Who it is for:** A product team choosing an image editing API that needs evidence to show non-technical stakeholders.

**The problem:** A number in a table does not convince anyone; people need to see which model wrecks the untouched parts of a photo.

**The outcome:** A leaderboard, per-step curves and a side-by-side video of two models across all 40 edits.

### Steps

1. Run `editdrift steps  benchmarks/milk` to see the 40 edits and generate them with each model
2. Score each model with `editdrift score  benchmarks/milk results/milk/my-model`
3. Build the leaderboard with `editdrift report benchmarks/milk results/milk --out report`
4. Render two models side by side with `editdrift video  benchmarks/milk results/milk/model-a results/milk/model-b out.mp4`
5. For every model at once use `editdrift grid   benchmarks/milk results/milk grid.mp4`

### Prompt to paste into your agent

```text
Run edit-drift's score, report and video commands for <model-a> and <model-b> in results/milk. Output report/ and out.mp4, and confirm both models have the same number of steps before comparing.
```

### Caveat

Drift is measured on the regions that should stay unchanged, so a model can look good here and still miss the edit you asked for.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/megatroll222-edit-drift/compare-models-side-by-side/

## How to run the benchmark from Claude Code with Edit Drift

Repo: [MegaTroll222/edit-drift](https://github.com/MegaTroll222/edit-drift) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README

Add the plugin to Claude Code and let it run the 40-step chained-edit test, score drift and write the report.

**Who it is for:** An ML engineer comparing image editing models who wants the benchmark run from inside the coding agent, not by hand.

**The problem:** Scoring 40 consecutive edits per model with masks and reports is tedious to script correctly.

**The outcome:** scores.json per model plus a leaderboard report, produced by Claude Code from your results folder.

### Steps

1. Install the CLI with `pip install git+https://github.com/MegaTroll222/edit-drift`
2. In Claude Code run `/plugin marketplace add MegaTroll222/edit-drift` then `/plugin install edit-drift@edit-drift`
3. Put each model's outputs under results/milk/<model> (one image per step)
4. Ask Claude Code to score them; it runs `editdrift score  benchmarks/milk results/milk/my-model` per model
5. Ask for the report; it runs `editdrift report benchmarks/milk results/milk --out report`

### Prompt to paste into your agent

```text
Using the edit-drift plugin, score every model folder under results/milk and build the report in ./report. Tell me which model drifted least by step 40 and check every folder has all 40 steps.
```

### Caveat

The benchmark scores images you already generated; calling the image models themselves is up to you.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/megatroll222-edit-drift/run-edit-drift-from-claude-code/