Red-Pathfinder/dcdd-wacv-2027

Disability-Conditioned Demographic Drift (DCDD) Benchmark Dataset

Official benchmark dataset for the paper:

Disability-Conditioned Demographic Drift: Intersectional Stereotyping and Algorithmic Erasure in Text-to-Image Generative Models
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2027)
Code & Reproduction: https://github.com/Red-Pathfinder/DCDD

---

Dataset Overview

The DCDD dataset is an empirical audit benchmark evaluating whether text-to-image generative models alter demographic distributions and contextual framing when disability conditioning is added to neutral occupational prompts.

This Hugging Face repository hosts the benchmark data assets (image binaries, metadata manifests, VLM annotations, human audit corrections, and blinded reliability study records). All reproduction code, statistical pipelines, and documentation are maintained on GitHub.

---

Dataset Summary & Accounting

Model Architecture	API Endpoint	Planned Tasks	Cataloged Attempts	Completed Images	Safety Refusals	Timeouts	Image Format
FLUX.1 [dev]	fal-ai/flux/dev	385	380	380	0	5	1024 × 768 JPEG
Stable Diffusion 3.5 Large	fal-ai/stable-diffusion-v35-large	385	385	385	0	0	1024 × 1024 JPEG
Nano Banana 2	Google Imagen 3 (gemini-3.1-flash endpoint)	1,925	1,925	1,707	218	0	1408 × 768 JPEG
Total	—	2,695	2,690	2,472	218	5	100% JPEG

• 2,690 Cataloged Attempts: Cataloged in metadata/dataset_manifest.csv (2,472 successful image generations and 218 safety filter refusals).
• 5 Infrastructure Timeouts: Unrecorded HTTP timeouts during initial FLUX batch generation on fal.ai. Accounting for these reconciles the cataloged dataset to the planned 2,695 tasks.

---

Repository Structure

dcdd-wacv-2027/
├── README.md                          # Dataset card (this file)
├── checksums.sha256                   # SHA-256 verification manifest
├── data/                              # 2,472 generated image binaries
│   ├── flux/                          # 380 JPEG images
│   ├── sd35/                          # 385 JPEG images
│   └── nanobanana/                    # 1,707 JPEG images
├── metadata/                          # Relational metadata tables
│   ├── dataset_manifest.csv           # Master catalogue of all 2,690 attempts
│   ├── generation_metadata.csv        # Seeds, endpoints, parameters
│   ├── annotations.csv                # VLM annotations across 12 visual attributes
│   ├── human_validation.csv           # 1,643 curated human corrections
│   ├── prompts.csv                    # 1,150 unique prompt texts
│   └── nanobanana_filename_mapping.csv# Original vs normalized filenames
└── human_reliability_study/           # 3-annotator measurement reliability study
    ├── README.md                      # Study protocol and documentation
    ├── image_manifest.csv             # Blinded image lookup table
    ├── annotator_1.csv                # Annotator 1 evaluations (300 images)
    ├── annotator_2.csv                # Annotator 2 evaluations (300 images)
    ├── annotator_3.csv                # Annotator 3 evaluations (300 images)
    └── blinded_images/                # 300 blinded stimulus images

---

Data Ingestion & Download

1. Download Metadata Manifests via Python (~1.5 MB)

from huggingface_hub import hf_hub_download
import pandas as pd

manifest_path = hf_hub_download(
    repo_id="Red-Pathfinder/dcdd-wacv-2027",
    filename="metadata/dataset_manifest.csv",
    repo_type="dataset"
)
df = pd.read_csv(manifest_path)
print(f"Loaded {len(df)} attempts across {df['model'].nunique()} models.")

2. Download Full Dataset via Python (~1.6 GB)

from huggingface_hub import snapshot_download

dataset_dir = snapshot_download(
    repo_id="Red-Pathfinder/dcdd-wacv-2027",
    repo_type="dataset"
)
print(f"Dataset downloaded to: {dataset_dir}")

---

Official Code & Reproduction

For the complete reproduction pipeline, environment setup, and verification of paper tables and figures, visit the companion GitHub repository:

👉 https://github.com/Red-Pathfinder/DCDD

git clone https://github.com/Red-Pathfinder/DCDD.git
cd DCDD
python reproduce.py

---

Citation

@inproceedings{dcdd2027,
  title={Disability-Conditioned Demographic Drift: Intersectional Stereotyping and Algorithmic Erasure in Text-to-Image Generative Models},
  author={[Author details omitted for double-blind review]},
  booktitle={IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year={2027}
}