kaikaiyao/sd-image-attribution

SD Image Attribution

A public benchmark for image attribution across Stable Diffusion families, built from six models and three prompt datasets.

Version v1 includes PartiPrompts (1,632 prompts), DrawBench (200 prompts), DiffusionDB sample_10k (10,000 prompts), 6 models, and 70,992 generated images.

Why this dataset

• Same prompt sources across multiple Stable Diffusion families make model-level comparisons cleaner and more controlled.
• Public metadata links each image to its prompt source, model identity, and generation settings.
• The current release is sized for practical experiments while staying simple to load from a single Parquet index.
• The dataset viewer uses a lightweight thumbnail preview subset so you can browse by model in the config dropdown and narrow to a prompt dataset with the prompt_dataset filterable column.

At a glance

Images	Models	Prompt datasets	Version
70,992	6	3	v1

Coverage

This release is designed for image provenance, model attribution, model fingerprinting, watermarking, and related generated-image analysis studies. It combines the same prompt sources with multiple Stable Diffusion variants so researchers can study model-specific artifacts under a shared prompt distribution.

Prompt dataset	Source	Release split	Prompts used
PartiPrompts	nateraw/parti-prompts	full benchmark split	1,632
DrawBench	shunk031/DrawBench	full benchmark split	200
DiffusionDB	poloclub/diffusiondb	sample_10k curated slice	10,000

PartiPrompts and DrawBench use their full benchmark splits in this release. DiffusionDB uses the curated sample_10k release slice.

Model	Checkpoint	Resolution	Inference steps	Guidance scale	Scheduler
Stable Diffusion 1.5 (sd15)	stable-diffusion-v1-5/stable-diffusion-v1-5	512x512	30	7.5	PNDMScheduler
Stable Diffusion 2.1 (sd21)	sd2-community/stable-diffusion-2-1	768x768	30	7.5	DDIMScheduler
Stable Diffusion XL (sdxl)	stabilityai/stable-diffusion-xl-base-1.0	1024x1024	30	5.0	EulerDiscreteScheduler
SDXL Turbo (sdxl-turbo)	stabilityai/sdxl-turbo	512x512	4	0.0	EulerAncestralDiscreteScheduler
Stable Diffusion 3.5 Medium (sd35-medium)	stabilityai/stable-diffusion-3.5-medium	1024x1024	28	4.5	FlowMatchEulerDiscreteScheduler
Stable Diffusion 3.5 Large Turbo (sd35-large-turbo)	stabilityai/stable-diffusion-3.5-large-turbo	1024x1024	4	0.0	FlowMatchEulerDiscreteScheduler

Gallery

PartiPrompts

DrawBench

DiffusionDB

How to use

Load the main metadata table with datasets:

from datasets import load_dataset

ds = load_dataset(
    "parquet",
    data_files={
        "train": "https://huggingface.co/datasets/kaikaiyao/sd-image-attribution/resolve/main/metadata/all.parquet"
    },
)["train"]
print(ds[0]["image_path"])

Or read it directly with pandas:

import pandas as pd

df = pd.read_parquet(
    "https://huggingface.co/datasets/kaikaiyao/sd-image-attribution/resolve/main/metadata/all.parquet"
)
print(df[["dataset", "model_key", "prompt", "image_path"]].head())

Metadata

metadata/all.parquet is the main table for the release.

• Identity and source: dataset, dataset_split, dataset_row_id, model_key, resolved_model_id
• Prompt and generation: prompt, seed, steps, guidance_scale, scheduler, width, height
• Image and integrity: image_path, sha256, status
• Original prompt metadata: source_record keeps dataset-specific fields from PartiPrompts, DrawBench, or DiffusionDB

Example row excerpt:

{
  "dataset": "parti-prompts",
  "dataset_split": "train",
  "dataset_row_id": 4,
  "model_key": "sdxl",
  "resolved_model_id": "stabilityai/stable-diffusion-xl-base-1.0",
  "prompt": "A watercolor fox reading a newspaper in a cafe",
  "seed": 3096288400,
  "steps": 30,
  "guidance_scale": 5.0,
  "width": 1024,
  "height": 1024,
  "image_path": "images/parti-prompts/sdxl/parti-prompts-000004-sdxl-b88e7b50.png",
  "sha256": "34f9f56d0f1bf3ef43a4a0ee6b4f614f6fc0a55e942d08dc4f28c5614fd4dfe7",
  "source_record": {
    "Prompt": "A watercolor fox reading a newspaper in a cafe",
    "Category": "Art"
  }
}

Uses and limitations

• Intended for research on image provenance, model attribution, model fingerprinting, watermarking, and related generated-image analysis tasks.
• Only succeeded generations are included in the public release.
• Future releases are intended to expand prompt coverage, model families, and overall scale.
• Use of this dataset remains subject to the licenses and terms of the upstream prompt datasets and model checkpoints.

Citation

If you use this dataset, please cite:

@dataset{yao2026sd_image_attribution,
  author = {{Kai Yao}},
  title = {{SD Image Attribution}},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/kaikaiyao/sd-image-attribution},
  note = {Version: v1}
}