yanlinli/image-editing-evaluation-handoff

Image editing evaluation handoff

This folder contains the PIE-Bench main set (700 cases), the ICE-Bench Reference Editing subset (518 cases), edited images for 12 methods on each benchmark, the per-case metric CSVs, and the metric implementation used for the reported results. PIE and ICE are reported separately. The ICE results use the PIE-style metrics plus reference-image CLIP similarity; they are not the official ICE-Bench aggregate score.

Files

Path	Contents
archives/image_editing_piebench_inputs.zip	700 source images and a compact sample list
archives/image_editing_icebench_reference_inputs.zip	518 source images, reference images, available masks, and a compact sample list
archives/image_editing_ours_results.zip	Seven variants on both benchmarks; edited images only
archives/image_editing_baselines_results.zip	Five baselines on each benchmark; edited images only
evaluation/PIE-Bench_DirectInversion/data/PIE-Bench_v1/mapping_file.json	Full PIE descriptions, instructions, categories, and encoded masks required by the evaluator
evaluation/ICE-Bench/data/data_reference.jsonl	Full ICE sample metadata and captions required by the evaluator
evaluation/PIE-Bench_DirectInversion/evaluation/	PIE metric evaluator and shared metric implementation
evaluation/ICE-Bench/evaluate_reference_pie_metrics.py	ICE reference-editing metric evaluator
metrics/	Existing per-case metric CSVs for all 24 benchmark/method combinations
image_editing_evaluation_protocol_zh.md	Detailed metric and aggregation protocol in Chinese
SHA256SUMS	Checksums for the four archives
environment.yml	CUDA 12.1 conda environment for both image metric evaluators

The seven variants are our_geo, our_imgasvid, ours_7k, ours_10k, ours_3d_9k, ours_imgasvid_v2, and ours_geo_full_modality. PIE baselines are ACEPlus_LoRA, ChordEdit_sd_turbo, FLUX1_Kontext_dev, ICEdit_normal_lora, and VAREdit_8B_512. ICE baselines are ACEPlus_LoRA, AnyDoor_auto_ref_mask, DreamOmni2_edit, FLUX1_Kontext_dev, and MimicBrush.

Prepare the evaluation layout

From this folder, run:

conda env create -f environment.yml
conda activate image-editing-eval
sha256sum -c SHA256SUMS
python prepare_eval_layout.py

The preparation script extracts the images into the paths expected by the evaluators, verifies the 700 PIE source images and 518 ICE output images per method, and creates portable ICE gen_info.json files. The original ICE result mappings contained an absolute path from the generating machine, so use the regenerated mappings. Allow space for the extracted images in addition to the approximately 5.8 GB of compressed archives.

The datasets/ folder created during preparation keeps the compact sample lists from the input archives. The full evaluation metadata is already in evaluation/.

Recompute one method's metrics

Use the included environment.yml for both evaluators. The structure metric loads DINO ViT-B/8 through torch.hub; CLIP ViT-L/14 and the SqueezeNet LPIPS weights are also required. The scripts default to CUDA. Model downloads or local caches must be available when they first run. Model weights are not bundled.

PIE-Bench example, from this folder:

cd evaluation/PIE-Bench_DirectInversion
python -m evaluation.evaluate \
  --annotation_mapping_file data/PIE-Bench_v1/mapping_file.json \
  --src_image_folder data/PIE-Bench_v1/annotation_images \
  --tgt_methods 9_ours_7k \
  --result_path rerun_ours_7k.csv

ICE-Bench example, from this folder:

cd evaluation/ICE-Bench
PIE_BENCH_ROOT=../PIE-Bench_DirectInversion python evaluate_reference_pie_metrics.py \
  --meta_file data/data_reference.jsonl \
  --result_file results/ours_7k/gen_info.json \
  --output_csv rerun_ours_7k.csv

For PIE, change 9_ours_7k to the method key in evaluation/evaluate.py. For ICE, change the result directory name. The existing CSVs in metrics/ provide the recorded scores without rerunning the GPU evaluation. See the protocol document for metric definitions, valid sample counts, and aggregation details.