samfrench9/image-ocean2-sdxl-82k

Image Ocean 2: 82,481 SDXL images, 225 prompts, one fixed recipe

82,481 images generated with Stable Diffusion XL 1.0 base + refiner on one RTX 4090 between 2026-04-20 and 2026-05-13. 225 prompts, a median of 338 seeds per prompt, and one fixed sampler recipe for 82,473 of them. Every row carries the full recipe (prompt, negative prompt, seed, steps, CFG, scheduler, VAE, FreeU, refiner settings) plus a LAION aesthetic score and an NSFW probability.

• Seed sweeps at scale. Hundreds of seeds per prompt with everything else held constant. Useful for measuring seed variance, best-of-N selection, reward-model and preference work, and diversity studies.
• Lossless. Images are stored as lossless WebP, pixel-identical to the original PNG output (no JPEG artifacts — matters for synthetic-image detection research).
• Reproducible. Seeds and settings are exact; the generator code is public at samuelfrench/image-ocean2.
• Public domain (CC0). Use it for anything, no attribution required.

Same prompt, eight seeds

"an abandoned art deco theater being reclaimed by flowering vines" — 217 seeds of this prompt are in the dataset. These eight span its score range, lowest to highest (label: seed · LAION aesthetic score):

Seed choice alone moves the aesthetic score a lot. Across the 202 prompts with 20+ seeds, the median gap between a prompt's 10th- and 90th-percentile seed is 0.38 points:

Quick start

from datasets import load_dataset

# stream without downloading all ~76 GB
ds = load_dataset("samfrench9/image-ocean2-sdxl-82k", split="train", streaming=True)
row = next(iter(ds))
row["image"]          # PIL.Image, lossless
row["prompt"], row["seed"], row["laion_aesthetic"]

# metadata only (no image bytes, a few MB): filter first, then fetch what you need
meta = load_dataset("samfrench9/image-ocean2-sdxl-82k", "metadata", split="train").to_pandas()
best = meta[meta.laion_aesthetic >= 6.5]

Reproduce any image with diffusers: SDXL base for steps steps up to high_noise_frac, then the refiner for the rest, with the listed scheduler, vae, guidance_scale, FreeU (freeu_b1, freeu_b2, freeu_s1, freeu_s2) and refiner aesthetic conditioning (refiner_aesthetic_score, refiner_negative_aesthetic_score). The exact code is generate.py.

What is in it

Category	Images	Prompts	Median LAION aesthetic
meme	18,571	43	6.05
whimsical	16,472	47	6.12
animals	7,668	14	6.34
fantasy	7,213	14	6.41
nature	6,584	24	6.51
sci_fi	5,531	15	6.50
architecture	4,957	20	6.28
portraits	4,398	16	6.27
abstract	3,869	9	6.21
food	3,863	14	5.99
vehicles	3,355	9	6.33

Resolution	Images
1024 × 1024	49,992
1344 × 768	20,423
1152 × 896	7,668
832 × 1216	4,398

Files: data/ holds 165 Parquet shards (~76 GB total, ~500 rows each) in generation order; metadata/metadata.parquet holds the same rows without image bytes.

Columns

Column	Type	Meaning
image	image	lossless WebP, pixel-identical to the generated PNG
id	string	original filename stem: <local timestamp>_<model>_<category>_<prompt slug>_s<seed>
created_at	string	generation time, US Central local time
category	string	prompt category (11 values)
prompt_id	int32	stable integer id of prompt_subject (0–224)
prompt_subject	string	the scene description
prompt	string	full positive prompt sent to the model (subject + quality suffix)
negative_prompt	string	full negative prompt
seed	int64	torch generator seed
width, height	int32	output size (SDXL training buckets chosen per category)
steps, refiner_steps	int32	base and refiner inference steps
high_noise_frac	float64	fraction of denoising done by the base model before the refiner
guidance_scale	float64	CFG scale
scheduler, vae	string	sampler and VAE used
freeu_b1 … freeu_s2	float64	FreeU parameters (null for pipeline v1)
refiner_aesthetic_score, refiner_negative_aesthetic_score	float64	refiner aesthetic conditioning inputs (null for v1) — not a measured score
model_repo	string	checkpoints used
pipeline_version	int32	1 = first 8 images (Euler, default VAE, CFG 7.5, no FreeU); 2 = everything else
elapsed_seconds	float64	wall time to generate the image on an RTX 4090
caption_top, caption_bottom, caption_format	string	meme caption burned into the image with Pillow (115 images; null otherwise)
laion_aesthetic	float32	LAION improved aesthetic predictor v2 (sac+logos+ava1-l14-linearMSE, CLIP ViT-L/14), measured after generation
nsfw_prob	float32	probability of the nsfw label from Falconsai/nsfw_image_detection

How it was made

Pipeline v2 (all but 8 images): SDXL 1.0 base → refiner handoff at 0.8, 40 + 15 steps, DPM++ 2M Karras, CFG 7, madebyollin/sdxl-vae-fp16-fix, FreeU (b1 1.3, b2 1.4, s1 0.9, s2 0.2), refiner aesthetic conditioning 6 / 2.5. The generator ran unattended, picking a random category and prompt for each image with a fresh random seed; nothing was cherry-picked or removed for quality.

Removed before publishing: 7 images with nsfw_prob ≥ 0.5, 2 removed after manual review, 2 images whose metadata sidecar was missing.

Limitations

• Narrow prompt set. 225 prompts written for a "random pretty picture" generator: animals, fantasy, landscapes, food, vehicles, office-humor scenes. Not a general-purpose text-to-image training set.
• Uneven coverage. Seeds per prompt range from 1 to 956; the meme and whimsical categories dominate.
• Model biases. Everything SDXL 1.0 gets wrong is here too: hands, text, anatomy, Western-default people and places. The quality suffix (masterpiece, best quality, … 8k) pushes toward a glossy, high-contrast look.
• Scores are model outputs. laion_aesthetic and nsfw_prob are classifier predictions, not human ratings.

License

CC0 1.0 — public domain dedication. No attribution required. The images were generated with Stable Diffusion XL 1.0 (CreativeML Open RAIL++-M), whose license places no restrictions on outputs beyond its use-based restrictions; in the US, purely AI-generated images are not eligible for copyright. Please do not use it to impersonate real people or to pass AI-generated images off as photographs.

Citation

@misc{image_ocean2_sdxl_82k,
  title  = {Image Ocean 2: 82,481 SDXL images, 225 prompts, one fixed recipe},
  author = {French, Sam},
  year   = {2026},
  url    = {https://huggingface.co/datasets/samfrench9/image-ocean2-sdxl-82k}
}