asiimo/genai-bench-sdxl-latents

GenAI-Bench SDXL with Latent Trajectories

SDXL generations for the GenAI-Bench prompts, paired with the per-step decoded-latent trajectory of each generation. This dataset was created to evaluate NoisyCLIP. For every prompt, 10 images were generated (1600 prompts → 16,000 generations). Each generation provides:

• the full-resolution final image,
• the 50-step denoising trajectory (each step's latent decoded to a small preview image), and
• a CLIP-FlanT5-XXL VQAScore measuring prompt–image alignment.

The original GenAI-Bench basic / advanced prompt categories are merged into a single train split — the category is still recoverable from the Tags column.

Configs

The dataset ships as two configs in one repo so you can skip the heavy trajectories when you only need final images.

Config	Pick with	Contents	Approx size
default	load_dataset(repo)	metadata + final_image	~22 GB
latents	load_dataset(repo, "latents")	join keys + the 50 trajectory frames	~38 GB

default

Column	Type	Description
id	string	Per-generation id, "<prompt>_<imageIdx>" (e.g. "0_0"). Unique per row.
key	int64	Prompt id (0–1599). Shared by all 10 images of a prompt.
prompt	string	The text prompt.
Tags	{basic: [string], advanced: [string]}	GenAI-Bench skill tags. The non-empty side indicates the prompt's category (basic vs advanced).
score_clip-flant5-xxl	float	VQAScore (CLIP-FlanT5-XXL) prompt–image alignment for final_image.
final_image	Image	Full-res SDXL output (1024×1024).

latents

Column	Type	Description
id, key, prompt, Tags	—	Same as above (for joining/lookup).
frame_00 … frame_49	Image	The 50-step denoising trajectory. frame_kk is the decoded latent at step k (128×128). frame_00 is the first/noisiest step, frame_49 the last/cleanest.

Usage

from datasets import load_dataset

# final images + metadata
ds = load_dataset("<your-username>/genai-bench-sdxl-latents", split="train")
ex = ds[0]
ex["final_image"]          # PIL.Image, 1024x1024
ex["prompt"], ex["score_clip-flant5-xxl"]

# trajectories (heavier)
lat = load_dataset("<your-username>/genai-bench-sdxl-latents", "latents", split="train")
row = lat[0]
trajectory = [row[f"frame_{k:02d}"] for k in range(50)]   # 50 PIL images, step 0 -> 49
final_latent = row["frame_49"]

Join the two configs on id (same row order is not guaranteed across configs, so join on the key rather than by position):

final_by_id = {r["id"]: r["final_image"] for r in ds}
img = final_by_id[lat[0]["id"]]

Load just one denoising step cheaply via column selection:

last = load_dataset("<your-username>/genai-bench-sdxl-latents", "latents",
                    split="train", columns=["id", "frame_49"])

Notes

• final_image vs frame_49. final_image is the full-res VAE output (1024×1024, ~1.4 MB). frame_49 is the small decoded-latent preview of the last step (128×128). They are not the same image.
• key is not unique. It identifies the prompt; use id to identify a specific generation.

Source

Prompts come from GenAI-Bench (Li et al., 2024). Images are generated with Stable Diffusion XL; generated images are subject to the SDXL model license. The VQAScore follows the GenAI-Bench / VQAScore methodology (CLIP-FlanT5-XXL).

Citation

If you find this dataset useful for your research, please cite:

@misc{ramos2026earlyestimationlanguagelatent,
      title={Early Estimation of Language to Latent Alignment in Diffusion Models}, 
      author={Vasco Ramos and Regev Cohen and Idan Szpektor and Joao Magalhaes},
      year={2026},
      eprint={2512.08505},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.08505}, 
}