Face-Edit-Bench

Face-Edit-Bench is a large-scale benchmark for evaluating text-guided facial image editing models. It is introduced in:

A Large-scale Evaluation of Text-guided Models for Facial Editing ACMMM 2026 GitHub: https://github.com/rahul1801/Face-Edit-Bench

The benchmark evaluates models on three axes — identity preservation, edit fidelity, and unintended attribute changes — with fine-grained breakdowns by demographic subgroup.

---

Models Evaluated

Model	Source
Gemini Image 2.5 Flash	Google DeepMind
BAGEL-Edit	via Replicate
FLUX.1-dev	via Replicate
FLUX Kontext Pro	via Replicate
Qwen-Edit	via Replicate
SeedEdit v3	via Replicate

---

Edit Axes

Each model is evaluated across four edit categories:

Axis	Description
accessories	Adding or changing glasses, hats
hair	Hair color, type, style, beard, mustache
poses	Head position / gaze direction
multi_axis	Combined edits spanning multiple axes

---

Dataset Files

File	Size	Description
face_edit_results_metadata.tar.gz	—	All evaluation metadata (scores, state dicts) per session. Sufficient to reproduce all paper results.
face_edit_attributes.json	—	Edit candidate database. Required for OER (unintended change) analysis.
list_attr_celeba.txt	—	Standard CelebA attribute annotations. Required for demographic analysis.
celebset_unique_images.csv	—	CelebSet identity annotations (ethnicity, gender, age). Required for demographic analysis.
face_edit_image_data/*.tar.gz	~958 GB total	One tar per model×dataset×axis. Contains all edited images. Not needed to reproduce paper results, but critical for advancing image editing research.

---

Reproducing Paper Results

Download and extract the metadata archive — this is the only file required to reproduce all FSS, SC, Drop_FSS, and OER results from the paper:

pip install huggingface_hub

huggingface-cli download rnair21/Face-Edit-Bench \
    face_edit_results_metadata.tar.gz \
    face_edit_attributes.json \
    list_attr_celeba.txt \
    celebset_unique_images.csv \
    --repo-type dataset \
    --local-dir ./face_edit_bench_data

Extract the archive:

tar -xzf face_edit_bench_data/face_edit_results_metadata.tar.gz -C face_edit_bench_data/

After extraction, your directory should look like:

face_edit_bench_data/
├── face_edit_bench_results/       ← extracted metadata
├── face_edit_attributes.json
├── list_attr_celeba.txt
└── celebset_unique_images.csv

Then follow the instructions in the GitHub repo to run the analysis scripts.

---

Metadata Structure

Extracting face_edit_results_metadata.tar.gz produces the directory face_edit_bench_results/ with 44 subfolders, one per model × dataset × edit axis combination:

face_edit_bench_results/
├── celeba_full_gemini-image_2.5-flash_test_set_accessories/
├── celeba_full_gemini-image_2.5-flash_test_set_hair/
├── celeba_full_gemini-image_2.5-flash_test_set_multi_axis/
├── celeba_full_gemini-image_2.5-flash_test_set_poses/
├── celeba_full_replicate_bagel-edit_test_set_accessories/
│   ...
├── celebset_full_replicate_seededit-v3_test_set_accessories/
│   ...

Folder naming convention: {dataset}_{model}_test_set_{edit_axis}

• dataset: celeba_full or celebset_full
• model: e.g., gemini-image_2.5-flash, replicate_flux-kontext-pro
• edit_axis: accessories, hair, multi_axis, poses

Each experiment folder contains thousands of session subfolders, one per identity × sequence:

celeba_full_gemini-image_2.5-flash_test_set_accessories/
├── 000245_Accessories_seq0/
│   └── session_results.json
├── 000245_Accessories_seq1/
│   └── session_results.json
├── 000245_Accessories_seq2/
│   └── session_results.json
│   ...

Session folder naming: {image_id}_{Axis}_seq{N}

• image_id: CelebA/CelebSet image identifier
• Axis: primary edit axis for this session
• N: sequence index (multiple sequences may exist per identity)

---

Understanding session_results.json

Each session folder contains a single session_results.json that captures everything about one sequential editing session: the prompts used, the per-step metric scores, and a full attribute state snapshot at each step.

Top-level structure

{
  "sequence_info":       { ... },   // what was edited and how
  "step_by_step_metrics": [ ... ],  // scores at each step (base + 4 edits)
  "summary_statistics":  { ... },   // aggregate scores for the session
  "state_tracking":      { ... }    // attribute state evolution across steps
}

---

sequence_info — the edit plan

"sequence_info": {
  "sequence_id":    "single_axis_Accessories",
  "identity_id":    "4686",
  "edit_method":    "single_axis",
  "primary_axis":   "Accessories",
  "sub_axes_used":  ["glasses", "hats"],
  "prompts": [
    "Add square glasses",
    "Add a wool hat",
    "Replace their wool hat with a knit hat.",
    "Change to rectangular sunglasses"
  ],
  "total_steps": 5,    // 1 base image + 4 sequential edits
  "edit_steps":  4,
  "instructions_metadata": [
    {
      "instruction":    "Add square glasses",
      "axis":           "accessories",
      "sub_axis":       "glasses",
      "base_candidate": "rectangle-shaped glasses",
      "synonym_used":   "square glasses",
      "edit_index":     0
    },
    ...
  ]
}

• prompts — the exact text instructions sent to the model, in order
• instructions_metadata — the ground-truth target attribute for each step, along with the synonym used in the prompt (e.g., "square glasses" is a synonym for the canonical "rectangle-shaped glasses")

---

step_by_step_metrics — scores at each step

One entry for step=0 (the base image) and one for each of the 4 edits (step=1 through step=4).

{
  "step": 1,
  "image_path": "face_edit_bench_results/.../output_0.png",
  "is_base_image": false,
  "prompt": "Add square glasses",
  "metrics": {
    "identity_preservation": {
      "face_id_score": 0.546,   // ArcFace cosine similarity to the original face
      "dino_i_score":  0.0      // DINOv2 image-level similarity (null for base)
    },
    "edit_alignment": {
      "clip_score":                 0.263,  // CLIP text-image alignment
      "semantic_consistency_score": 1.0    // whether the target attribute is present (0 or 1)
    },
    "naturalness": {
      "pq_score": 1.0   // perceptual quality score
    }
  },
  "state_dict": {
    // full VLM probability distribution over all attribute candidates at this step
    "all_candidate_scores": {
      "rectangle-shaped glasses": { "score": 0.95, "category": "accessories", "subcategory": "glasses" },
      "stubble":                  { "score": 0.90, "category": "hair",        "subcategory": "beard"   },
      ...
    },
    "base_categories_per_subaxis": {
      "hair/color":          "black hair",
      "accessories/glasses": "rectangle-shaped glasses",
      "accessories/hats":    null,
      ...
    }
  }
}

The state_dict captures the full attribute fingerprint of the image at each step. This is the basis for detecting unintended changes — if an attribute that wasn't targeted (e.g., hair color) shifts between steps, that counts as an off-diagonal edit (OER).

---

summary_statistics — session-level rollup

"summary_statistics": {
  "base_score":            0.956,   // face_id_score of the original image
  "final_score":           0.409,   // face_id_score after all 4 edits
  "total_degradation":     0.547,   // base_score − final_score
  "avg_edit_score":        0.436,   // mean face_id_score across edit steps 1–4
  "progressive_failure_step": 1     // first step where identity degradation is detected
}

---

state_tracking — attribute state at each step (compact view)

A cleaner summary of step_by_step_metrics, showing only the top-scoring candidate per sub-axis at each step. Used by the analysis scripts to trace how attributes evolve across the edit sequence.

"state_tracking": {
  "step_analysis": [
    {
      "step": 0,
      "step_name": "Base Image",
      "top_scores": {
        "accessories/glasses":  { "candidate": "round glasses", "score": 0.005 },
        "accessories/hats":     { "candidate": "None",          "score": 0.97  },
        "hair/color":           { "candidate": "black hair",    "score": 0.86  },
        ...
      }
    },
    {
      "step": 1,
      "step_name": "After Edit 1",
      "prompt": "Add square glasses",
      "top_scores": {
        "accessories/glasses":  { "candidate": "rectangle-shaped glasses", "score": 0.95 },
        ...
      }
    },
    ...
  ]
}

---

Image Data (face_edit_image_data)

The face_edit_image_data/ portion of the dataset contains the actual images for every editing session. It is split into one .tar.gz per model × dataset × axis, mirroring the metadata structure.

Each tar extracts to a single folder with the same naming convention as the metadata:

celeba_full_gemini-image_2.5-flash_test_set_accessories/
├── 000245_Accessories_seq0/
│   ├── input_image.png    ← base image from CelebA/CelebSet
│   ├── output_0.png       ← after edit 1: "Add square glasses"
│   ├── output_1.png       ← after edit 2: "Add a wool hat"
│   ├── output_2.png       ← after edit 3: "Replace their wool hat with a knit hat."
│   ├── output_3.png       ← after edit 4: "Change to rectangular sunglasses"
│   └── session_results.json
│   ...

• input_image.png — the unedited base image drawn from CelebA or CelebSet
• output_N.png — the model's output after the (N+1)-th sequential edit prompt (prompts are in sequence_info.prompts)
• session_results.json — same file as in the metadata archive, included here for convenience

Note: Downloading the image data is not required to reproduce the paper's quantitative results. All scores are pre-computed and stored in session_results.json files within face_edit_results_metadata.tar.gz.

---

Annotation Files

list_attr_celeba.txt

Standard CelebA attribute file. Required by both fss_and_sc_analysis.py and oer_analysis.py when running demographic bias analysis over the CelebA subset.

celebset_unique_images.csv

CelebSet identity-level annotations including gender and ethnicity labels. Required by both analysis scripts when running demographic bias analysis over the CelebSet subset.

face_edit_attributes.json

The edit candidate database that defines all attribute candidates, categories, and sub-axes used in the benchmark. Required by oer_analysis.py for all runs (not just demographic analysis).

---

License

This dataset is released under CC-BY-4.0.

The underlying images derive from CelebA and CelebSet, which carry their own respective licenses. Please ensure compliance with those licenses when using the image data.

---

Citation

If you use Face-Edit-Bench in your research, please cite our paper:

@inproceedings{nair2026faceeditbench,
  title     = {A Large-scale Evaluation of Text-guided Models for Facial Editing},
  booktitle = {Proceedings of the 34th ACM International Conference on Multimedia (ACM MM)},
  year      = {2026}
}