oloruntobiolutola/facebench-pilot-r200

FaceBench: teacher selection and distillation data for binary face editing

FaceBench asks which instruction-based image editor makes the best teacher for identity-preserving face edits. It then distills that teacher into small one-step students. This repo holds the data for both stages:

1. Teacher selection. Three editors (FLUX.1 Kontext, Qwen-Image-Edit and LLaDA-Image) ran on three CelebA request sets: exact80, binary40 and b40u. Real target photos and blind human grades are included.
2. Distillation data. Qwen-Image-Edit, the chosen teacher, edited 1,216 CelebA photos. Each sample keeps the edited image and 8 of the 50 denoising steps, so pixel, latent and trajectory distillation are all possible.

The trained students are in oloruntobiolutola/facebench-students. That repo has a ready-to-run inference.py and every training checkpoint.

Result in one line. Qwen-Image-Edit is the best teacher (96% of edits done, 91% done while keeping the person). A 331M one-step SDXS student distilled from it is not significantly different from it in a blind human evaluation (96% / 88%). It is about 1,400× faster.

Contents

Path	What	Size
g2000_qwen_part{1,2,3,4a,4b,5a,5b}/	Distillation data: 1,216 Qwen-Image-Edit edits with teacher trajectories	~36 GB
selections/training/	Training selection: 2,000 planned requests, the parts, prompt tests, and the train / val split of the 1,216 samples that ran	small
pilot_exact80_*.zip	Teacher selection, round 1: 80 fine-grained level requests × 3 teachers	6 GB
binary40/*.zip	Teacher selection, round 2: 40 On/Off requests × 3 teachers (varied wordings)	3 GB
b40u_flux/, b40u_qwen/, b40u_llada/	Teacher selection, round 3: 40 On/Off requests on unseen people with the generic prompts × 3 teachers	130 MB
b40u_students/, b40u_sdxs_base/	The same 40 requests edited by the 2 students and by untrained SDXS-512	33 MB
b40u_eval/	b40u selection, prompts, human grades and comparison sheets	small
selections/teacher_selection/	The exact80, binary40 and b40u selections	small
annotations/	Blind human grades: binary40 (4 graders, 2 rounds) and b40u (2 graders)	small

The source photos are from flwrlabs/celeba (aligned, 178×218). The fine-grained attribute levels (0–5) used to pick requests come from the FaceBench request annotation file (request_annotated_without_request_mode_bangs.json in the code repository); each request's level is in samples.jsonl (attr_label for the source, target_score for the target).

The eight edits

Every request switches one attribute between two states. Off is level 0 (for age: 0–1, young), and On is level ≥ 3. The CelebA binary label must agree with the state.

Edit	Attribute	Direction	Generic prompt (distillation data and b40u)
Eyeglasses+	Eyeglasses	positive (Off→On)	"Add eyeglasses."
Eyeglasses-	Eyeglasses	negative (On→Off)	"Remove the eyeglasses."
No_Beard+	No_Beard (beard)	positive	"Add a beard."
No_Beard-	No_Beard (beard)	negative	"Remove the beard."
Smiling+	Smiling	positive	"Make the person smile."
Smiling-	Smiling	negative	"Change the facial expression to neutral, not smiling, with the lips closed."
Young+	Young (age)	positive: older	"Make the same person look older, keeping their face shape, gender, hairstyle and background unchanged apart from age."
Young-	Young (age)	negative: younger	"Make the same person look like a young adult, keeping their face shape, skin tone, ethnicity, gender, hairstyle, accessories and background unchanged apart from age."

The prompts were chosen with small prompt tests (selections/training/g_prompt_test*, g_age_test*). The plain wordings for removing a smile and for both age edits often failed with Qwen, hence the longer prompts.

1. Distillation data (g2000_qwen_part*)

Selection. selections/training/g2000_generic.csv lists 2,000 photos of 1,954 people (50% women), 250 per edit, with one fixed generic prompt per edit (g2000_generic.requests.json). It was split into 8 balanced parts so that each could run on a separate free Modal account. Seven parts ran, giving 1,216 samples: 152 per edit. Parts 5c and 6–8 did not run.

Teacher. Qwen/Qwen-Image-Edit: 50 steps, true CFG 4.0, negative prompt " ", output 928×1120, one fixed seed per sample, run on an H100 (about 52 s per sample).

Split. selections/training/g2000_trained_samples_splits.csv has columns sample_id, image_id, identity, attribute, direction, instruction, run, split.

• 1,098 train / 118 val, split by person (sha256(str(identity)) % 10 == 0 goes to val), so no person is in both.
• The b40u evaluation people are not in either split.

Layout of each part:

g2000_qwen_partN/
  run.json                 run settings (selection, seed, image source, model)
  samples.jsonl            one request per line: sample_id, image_id, identity, instruction, attribute, edit,
                           score_change_direction, target_state, seed, the 40 CelebA attributes, ...
  skipped.jsonl            requests that could not run (empty in every part)
  sources/<bucket>/<id>.jpg                              the source photo
  models/qwen_image_edit/images/<bucket>/<id>.png        the teacher's edit (928x1120)
  models/qwen_image_edit/teacher/<bucket>/<id>.safetensors   the teacher trajectory (28.5 MB, fp16)
  models/qwen_image_edit/records.shard*.jsonl            per-sample status, latency, paths (one file per worker)
  models/qwen_image_edit/model_config.json, model_info.json   the exact pipeline settings
  logs/

<bucket> is the first 3 digits of the 6-digit id.

Teacher file, 8 of the 50 steps, evenly spaced (latents are Qwen's packed 2×2 patches: 4,060 tokens × 64 channels = a 116×140×16 latent):

Key	Shape	Meaning
latents	[8, 4060, 64]	z_t at the 8 saved steps
velocity	[8, 4060, 64]	the guided velocity at those steps: z_next = z_t + (sigma_next − sigma) · velocity
sigma, sigma_next, timestep, step_index	[8]	the schedule at those steps
final_latents	[4060, 64]	z_0 given to the VAE decoder
cond_image_latents	[4060, 64]	the source photo in Qwen's VAE latent space
cond_text, cond_text_negative	[~1340, 3584]	the VLM (Qwen2.5-VL) condition for the prompt and the negative prompt

The file metadata stores the prompt, seed, size, scheduler and call arguments. Checks on every part:

• records are complete;
• image, teacher and source counts match;
• no NaN values;
• the last saved step reproduces final_latents (relative error about 1.7e-3).

from huggingface_hub import snapshot_download
from safetensors.torch import load_file
import json

root = snapshot_download("oloruntobiolutola/facebench-pilot-r200", repo_type="dataset",
                         allow_patterns=["g2000_qwen_part1/samples.jsonl", "g2000_qwen_part1/sources/000/*",
                                         "g2000_qwen_part1/models/qwen_image_edit/images/000/*"])   # add teacher/* for trajectories
for line in open(f"{root}/g2000_qwen_part1/samples.jsonl"):
    s = json.loads(line)          # s["edit"], s["instruction"], s["sample_id"], ...

To download everything (36 GB): hf download oloruntobiolutola/facebench-pilot-r200 --repo-type dataset --include "g2000_qwen_part*/*" --local-dir data/.

2. Teacher selection

Three teachers ran with the same seeds per sample:

Model	Settings	GPU	s / sample	Output
black-forest-labs/FLUX.1-Kontext-dev	28 steps, guidance 2.5	L40S	~36	944×1104
Qwen/Qwen-Image-Edit	50 steps, true CFG 4.0	H100	~53	928×1120
inclusionAI/LLaDA-Image	50 steps, guidance 5.0	H100	~60	896×1120

Each request has a real target photo: another CelebA photo of the same person in the requested state. It is for reference only and was never shown to the models.

Round	Files	Requests	Edit definition	Human grades
exact80	pilot_exact80_*.zip	80 (10 per edit, 40 people from the pilot batch)	change to an exact fine-grained level (0–5)	–
binary40	binary40/*.zip	40 (5 per edit, 40 people from all of CelebA)	On/Off, 40 different free wordings ("Remove the glasses", "The face can be older", ...)	4 graders, annotations/binary40_human_grades_round{1,2}.csv
b40u	b40u_*/	40 (5 per edit, 40 people not in the distillation data)	On/Off, the generic prompts above	2 graders, annotations/b40u_human_grades.csv

Run layout (the zips unpack to it, and the b40u_* folders use it directly): index.csv, samples.jsonl, targets.jsonl, sources/, targets/, models/<model>/images/<bucket>/<id>.png and models/<model>/records*.jsonl. The index.csv columns are sample_id, identity, instruction, attribute, direction, change, source_level, target_level, original, edited, target, target_image_id, target_photo_level, target_match, teacher, teacher_model, seed, latency_s.

The student folders (b40u_students/<model>/, b40u_sdxs_base/sdxs_base/) contain <bucket>/<id>.png plus index.csv (sample_id, edit, prompt, student, path, ms) and sources/.

Human evaluation

All grading was blind, on a web page that showed the source, the real target and every model's edit in a random order per sample (letters A–C, or A–F for b40u), at equal size, with model names hidden. Questions per edit:

• change: right (done), partial, none (not done) or wrong (the opposite edit, or another attribute changed);
• same person: 1–5;
• image quality: 1–5;
• per sample, the best edit or none.

CSV columns: grader, sample_id, model, shown_as, instruction, attribute, direction, source_level, target_level, change, same_person, quality, is_best, best_none, graded_at. Graders are pseudonymised as grader_1 to grader_4, with the same alias for the same person in every file. Round 1 of binary40 is an earlier pass with less consistent use of the guidelines; use round 2.

Agreement (Krippendorff's α, own implementation checked against Krippendorff's published example):

Set	Done (nominal)	Change, 4 classes	Same person (ordinal)	Quality (ordinal)
binary40 round 2, 4 graders, 120 edits	0.80	0.74	0.61	0.36
b40u, 2 graders, 240 edits	0.81	0.76	0.77	0.67

Results (mean over graders):

Model	binary40: done	b40u: done	b40u: done & same person ≥ 4	b40u: same person	b40u: quality	b40u: picked best
Qwen-Image-Edit	87%	96%	91%	4.47	4.69	30%
LLaDA-Image	85%	86%	51%	3.27	4.20	6%
FLUX.1 Kontext	57%	70%	70%	4.74	4.74	20%
SDXS student (331M, 1 step)	–	96%	88%	4.36	4.45	29%
Pixel student (29M)	–	56%	52%	3.44	3.09	10%
SDXS-512 untrained (zero-shot)	–	11%	2%	1.39	1.84	1%

• FLUX.1 Kontext keeps the person because it often does not edit. It leaves age unchanged (0–30% done).
• LLaDA-Image edits, but changes the person.
• Qwen-Image-Edit is the best teacher in both rounds.
• The SDXS student does not differ significantly from Qwen on any b40u measure (paired McNemar / Wilcoxon, Holm-corrected).

The full analyses are the notebooks 03_human_annotation_binary40.ipynb and 04_human_annotation_b40u.ipynb in the FaceBench code repository.

Cost

• Teacher data: about $105–115 for 1,216 samples, on the free $30 Modal credits of several team members.
• Teacher selection runs: about $10 per round of 40 × 3 teachers.
• Student training: about $1–2 per pixel / latent / flow run (30–90 min on an L4) and $4–5 per SDXS run (100 min on an L40S).
• Total for the project: about $150 across 6 Modal accounts.

Licence and responsible use

Research use only.

• The images derive from CelebA, which is for non-commercial research. Its terms apply to the source photos and to every edit of them.
• The edits were made with Qwen-Image-Edit (Apache 2.0), FLUX.1 Kontext [dev] (FLUX.1 [dev] non-commercial licence) and LLaDA-Image. Their licences apply to their outputs.
• The data shows real (celebrity) faces with synthetic edits. Do not use it to identify people, or to make or pass off misleading images of real people.

Citation

If you use this data, please cite CelebA (Liu et al., "Deep Learning Face Attributes in the Wild", ICCV 2015) and the teacher models (Qwen-Image, FLUX.1 Kontext, LLaDA-Image).