An evaluation benchmark released under CC BY-NC 4.0 for research evaluation only, not for training detection models or commercial use. Optional: tell me what you work on, and opt in below if you want a heads-up when datasets like this drop. I plan the next dataset around what people actually need.

Social Media Robustness Benchmark: SDXL+InstantID Synthetic Face Detection

Version: v1.0.0 · License: CC BY-NC 4.0 (research evaluation only)

Detector accuracy on clean lab test sets does not predict in-the-wild performance. Social platforms re-encode every uploaded image: platform-specific JPEG, resize, chroma subsampling, metadata stripped. This benchmark lets detector authors and procurers measure robustness under documented, paired, demographically-balanced conditions instead of blunt lab proxies.

A companion blog post and white paper cover the methodology, statistics, and findings in full. This card describes what the dataset is, how it is built at a high level, and how to use it.

1. Details

Field	Value
Name	Social Media Robustness Benchmark: SDXL+InstantID Synthetic Face Detection
Version	v1.0.0
Base corpus	2,400 images (1,200 real, 1,200 generated), sampled from danb21/synthetic-face-sdxl-instantid-bench at a deterministic per-cell quota
Perturbations	12 single-axis lab variants (JPEG, resize, noise, blur) + 4 platform-pipeline approximations (Instagram, Facebook, TikTok, X)
Total rows	40,574 image rows across 17 configurations (2,400 clean + 28,800 lab + 9,374 platform)
Cell axis	6 skin tones × 2 genders × 2 labels, 100 per cell at the clean baseline
Pairing	Every configuration evaluates the same images (paired by `media_id`)
Blog post	Read the write-up (companion post; published before the paper)
Paper	In preparation; methodology and results reported there
Maintainer	Daniel Babalola, danielbabalola@alumni.upenn.edu

2. What This Dataset Is For

Use it to:

Compute paired AUC deltas, AUC(clean) − AUC(perturbation), per detector per condition.
Measure per-cell robustness (skin tone × gender) under each perturbation.
Compare detector architectures under matched conditions.

It is not training data. It is small by design (2,400 base images), paired by construction (every condition evaluates the same images), and the platform pipelines are calibrated approximations of each platform's mean re-encode behaviour, not pixel-faithful platform reproductions (see Limitations).

3. Structure

Configurations

Config	Layer	Description
`clean`	clean	2,400 unperturbed base images; the reference for every paired delta
`layer1_jpeg_q{30,50,70,80,95}`	lab	JPEG re-encode at the named quality factor
`layer1_resize_{0.5,0.75}`	lab	Bicubic downsample then upsample back
`layer1_noise_{5,10}`	lab	Additive Gaussian noise (variance 5 / 10)
`layer1_blur_{1,2,4}`	lab	Gaussian blur (sigma 1 / 2 / 4)
`layer2_ig_pipeline`	platform	Instagram (JPEG ~92, max edge 1440, 4:2:0, EXIF stripped)
`layer2_fb_pipeline`	platform	Facebook (JPEG ~93, max edge 1920, 4:2:0, EXIF stripped)
`layer2_tt_pipeline`	platform	TikTok (JPEG ~80, max edge 1920, 4:2:0)
`layer2_x_pipeline`	platform	X (JPEG ~93, max edge 1920, 4:2:0, EXIF stripped)

All configurations share the same column schema. Key columns: image, label (real/generated), cell_skin_tone, cell_gender, media_id (stable across configurations for pairing), perturbation_slug, perturbation_layer, and the measured-encoding fields.

Balance

The clean baseline and all 12 lab configurations are uniform at 100 images per cell per label (6 skin tones × 2 genders × real/generated). The 4 platform configurations re-crop the laundered output back to a 256×256 face crop for comparability; all real-side cells retain 100/100, while a small fraction of synthetic-face crops do not survive re-detection (~2,342–2,344 rows per platform config). That concentration is itself a finding, analyzed in the white paper.

4. How It Is Built (high level)

Source. A deterministic per-cell subset of the v1 Synthetic Face Detection Benchmark: 1,200 real Pexels frames and 1,200 SDXL+InstantID outputs, uniform at 100 per cell per label.
Lab perturbations (Layer 1). Twelve single-axis variants applied independently to each base image, with deterministic seeds so the pipeline is reproducible. All outputs are 256×256.
Platform pipelines (Layer 2). For Instagram, Facebook, TikTok, and X, each platform's mean re-encode behaviour (resize to measured max edge, JPEG at measured quality, 4:2:0 chroma, EXIF handling) was measured and then applied to every base image. The laundered image is re-cropped to a 256×256 face crop so it is geometrically comparable to the clean and lab configurations.

The full calibration procedure, preregistered hypotheses, statistics, and caveats are documented in the companion white paper.

5. Results

Baseline accuracy, per-cell fairness, and per-platform robustness deltas are reported in the companion blog post and white paper, not in this card.

6. Comparison with Prior Robustness Datasets

Dataset	Year	Compression coverage	Paired pre/post	Platform calibration	Demographic balance	Scale (paired)
FaceForensics++	2019	Lab c0 / c23 / c40	Yes	No	None	~1,000,000
DFDC	2020	Mixed	Partial	No	Stated, not quantified	~500,000
GenImage	2023	Lab JPEG QF=96 only	Yes	No	None	~1,000,000
OpenFake	2025	None documented	No	No	Limited	Variable
This dataset	2026	12 lab axes + 4 calibrated platform pipelines	Yes (by `media_id`)	Yes (4 platforms)	6 skin tones × 2 genders, 100/cell/label	2,400 base (×17 configs)

Scale is smaller than prior work by design. This is an evaluation benchmark, not training data.

7. Bias, Risks, and Limitations

Architecture-level coverage only. The synthetic side is SDXL+InstantID (vanilla, no community fine-tunes, LoRA stacks, or frontier I2V models). Those are out of scope for v1.
Platform pipelines are calibrate-and-apply, not pixel-faithful. Each platform's measured mean encode behaviour is applied to every base image; the full upload chain and per-account or per-region variation are not reproduced.
TikTok calibration is a point estimate; treat TikTok results accordingly. Instagram, Facebook, and X measured qualities are close and should be read as applied parameters, not distinguishing fingerprints.
Synthetic-face re-crop drop. A small fraction of synthetic-face crops fail face re-detection after platform laundering, concentrated on certain cells. This is reported as a finding (analyzed in the white paper), not back-filled.
Real/generated track differences. The two tracks differ in crop and encoding history; a detector could in principle exploit that rather than genuine synthesis signal. The white paper quantifies a control for this.

8. License and Ethics

License: CC BY-NC 4.0 (research evaluation only). Commercial licensing inquiries: danielbabalola@alumni.upenn.edu.
Per-row attestation: LICENSES.csv lists each row's license. Real-side rows inherit Pexels licensing; generated-side rows carry the SDXL backbone + InstantID adapter terms.
Consent, age, identity: Inherited from the v1 base corpus. Real-side frames are sourced from Pexels; potentially under-18 frames were human-reviewed and only confirmed-adult rows are included; real-side identities are capped per cluster to prevent leakage.
Platform calibration disclosure: The platform parameters were measured by posting a small set of reference images and inspecting the platform-returned encoding. No image in this dataset is a platform post; the platform configurations are produced locally by applying the measured parameters to the base images.

9. Citation

@dataset{babalola_social_media_robustness_2026,
  title  = {Social Media Robustness Benchmark: SDXL+InstantID Synthetic Face Detection},
  author = {Babalola, Daniel},
  year   = {2026},
  url    = {https://huggingface.co/datasets/danb21/social-media-robustness-sdxl-instantid},
  note   = {Version v1.0.0}
}

A companion paper is in preparation; this citation will be updated with the DOI on publication.

Downloads last month: 7

Total file size:

702 MB