Skip to content

feat: add ImgEdit benchmark with edit type subsets#517

Open
davidberenstein1957 wants to merge 9 commits intomainfrom
feat/add-imgedit-benchmark
Open

feat: add ImgEdit benchmark with edit type subsets#517
davidberenstein1957 wants to merge 9 commits intomainfrom
feat/add-imgedit-benchmark

Conversation

@davidberenstein1957
Copy link
Member

@davidberenstein1957 davidberenstein1957 commented Jan 31, 2026

Closes #510

Summary

  • Add ImgEdit benchmark for image editing evaluation with 8 edit type subsets
  • Fetch instructions and judge prompts from GitHub (PKU-YuanGroup/ImgEdit)
  • Support subset filtering: replace, add, remove, adjust, extract, style, background, compose

Usage

from pruna.data import PrunaDataModule

# Load all edit types
dm = PrunaDataModule.from_string("ImgEdit")

# Load specific edit type
dm = PrunaDataModule.from_string("ImgEdit", subset="replace")

Test plan

  • PrunaDataModule.from_string("ImgEdit") works
  • Subset filter works for all 8 edit types
  • Auxiliaries include subset, image_id, judge_prompt fields
  • Docstring tests pass

davidberenstein1957 and others added 6 commits January 22, 2026 10:58
…mpts benchmark

- Introduced `from_benchmark` method in `PrunaDataModule` to create instances from benchmark classes.
- Added `Benchmark`, `BenchmarkEntry`, and `BenchmarkRegistry` classes for managing benchmarks.
- Implemented `PartiPrompts` benchmark for text-to-image generation with various categories and challenges.
- Created utility function `benchmark_to_datasets` to convert benchmarks into datasets compatible with `PrunaDataModule`.
- Added integration tests for benchmark functionality and data module interactions.
…filtering

- Remove heavy benchmark abstraction (Benchmark class, registry, adapter, 24 subclasses)
- Extend setup_parti_prompts_dataset with category and num_samples params
- Add BenchmarkInfo dataclass for metadata (metrics, description, subsets)
- Switch PartiPrompts to prompt_with_auxiliaries_collate to preserve Category/Challenge
- Merge tests into test_datamodule.py

Reduces 964 lines to 128 lines (87% reduction)

Co-authored-by: Cursor <cursoragent@cursor.com>
Document all dataclass fields per Numpydoc PR01 with summary on new line per GL01.

Co-authored-by: Cursor <cursoragent@cursor.com>
- Add list_benchmarks() to filter benchmarks by task type
- Add get_benchmark_info() to retrieve benchmark metadata
- Add COCO, ImageNet, WikiText to benchmark_info registry

Co-authored-by: Cursor <cursoragent@cursor.com>
Update benchmark metrics to match registered names:
- clip -> clip_score
- clip_iqa -> clipiqa
- Remove unimplemented top5_accuracy

Co-authored-by: Cursor <cursoragent@cursor.com>
Closes #510

- Add setup_imgedit_dataset in datasets/prompt.py
- Support subset filter (replace, add, remove, adjust, extract, style, background, compose)
- Fetch instructions and judge prompts from GitHub (PKU-YuanGroup/ImgEdit)
- Register ImgEdit in base_datasets
- Add BenchmarkInfo entry with accuracy metric, task_type image_edit
- Add test for loading with subset filter

Co-authored-by: Cursor <cursoragent@cursor.com>
Copy link

@cursor cursor bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Cursor Bugbot has reviewed your changes and found 2 potential issues.

Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.

Comment @cursor review or bugbot run to trigger another review on this PR

davidberenstein1957 and others added 3 commits January 31, 2026 17:19
…nting

- Rename subset parameter to category in setup_imgedit_dataset
- Add empty dataset guard before ds.select([0])
- Fix trailing newlines linting issue
- Update tests to use category parameter

Co-authored-by: Cursor <cursoragent@cursor.com>
Prevents crash when category filter produces empty dataset.

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
@github-actions
Copy link

This PR has been inactive for 10 days and is now marked as stale.

@github-actions github-actions bot added the stale label Feb 13, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

feat: add ImgEdit benchmark with edit type subsets

1 participant