datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Delighting-Benchmark-Diverse-v1
Paint3D vs Hunyuan3D-2.0 Diverse De-lighting Benchmark
This repository contains a category-balanced benchmark for comparing how
Paint3D and Hunyuan3D-2.0 handle image-conditioned 3D texturing when the input
image contains visible illumination and shading.
Contents
50 Google Scanned Objects cases from 10 categories (5 cases per category).
Strongly lit conditioning images and reference unlit albedo renders.
100 canonical method packages: 50 Paint3D and 50… See the full description on the dataset page: https://huggingface.co/datasets/ZeyuLing/Delighting-Benchmark-Diverse-v1.Delighting-Benchmark-v1-Canonical-Outputs
Delighting Benchmark v1 — Canonical Outputs
Canonical Paint3D and Hunyuan3D-2.0 texture-generation outputs for 30 test
cases containing visible illumination and specular effects:
10 ABO matte cases
10 ABO glossy cases
10 Poly Haven PBR cases
Layout
Results are grouped by method:
paint3d/<case_id>/
hunyuan3d/<case_id>/
Every case contains exactly seven files:
input_image.png
input_mesh.obj
textured.glb
albedo_render.mp4
albedo_texture.png
gt_lit_render.mp4… See the full description on the dataset page: https://huggingface.co/datasets/ZeyuLing/Delighting-Benchmark-v1-Canonical-Outputs.flux-delight-dataset
FLUX Delighting Dataset
A dataset for training delighting models - removing lighting effects and converting images to uniform ambient lighting while preserving subject identity.
The dataset was created using imgaes from https://unsplash.com.
Task
Input: Image with various lighting conditions (spotlights, colored lights, dramatic shadows, etc.)
Output: Same image with uniform, ambient lighting and no harsh shadows
Prompt: "Remove all lighting and shadows. Make the image… See the full description on the dataset page: https://huggingface.co/datasets/linoyts/flux-delight-dataset.Seoul_bikeh2f_delight_train
Description
This is the dataset from DELIGHT [1], a model that locates the host galaxy of a
transient from multi-resolution images. The difference with the original is that
the images are not the cutouts used in the paper: they were regenerated from
hips2fits
(PanSTARRS DR1 HiPS, served by CDS).
It contains 7,350 objects, meant for training and validation. The fold
column carries a ready-made 5-fold cross-validation split, so the same file
covers both roles. Held-out evaluation… See the full description on the dataset page: https://huggingface.co/datasets/PRISM-Astro/h2f_delight_train.h2f_delight_radius_norm
Description
This is the dataset from DELIGHT [1], a model that locates the host galaxy of a
transient from multi-resolution images. The difference with the original is that
the images are not the cutouts used in the paper: they were regenerated from
hips2fits
(PanSTARRS DR1 HiPS, served by CDS).
It contains 9,246 examples, built by autolabeling: transient positions are
simulated, not observed. For each of the 7,093 host galaxies in
h2f_delight_train,
positions are drawn… See the full description on the dataset page: https://huggingface.co/datasets/PRISM-Astro/h2f_delight_radius_norm.h2f_delight_autolabeling_1pct
Description
This is the dataset from DELIGHT [1], a model that locates the host galaxy of a
transient from multi-resolution images. The difference with the original is that
the images are not the cutouts used in the paper: they were regenerated from
hips2fits
(PanSTARRS DR1 HiPS, served by CDS).
It contains 535,743 examples, built by autolabeling: transient positions are
simulated, not observed. For each of the 7,350 host galaxies in
h2f_delight_train,
positions are drawn… See the full description on the dataset page: https://huggingface.co/datasets/PRISM-Astro/h2f_delight_autolabeling_1pct.h2f_delight_radius_norm_p4
Description
This is the dataset from DELIGHT [1], a model that locates the host galaxy of a
transient from multi-resolution images. The difference with the original is that
the images are not the cutouts used in the paper: they were regenerated from
hips2fits
(PanSTARRS DR1 HiPS, served by CDS).
It contains 12,610 examples, built by autolabeling: transient positions are
simulated, not observed. For each of the 7,093 host galaxies in
h2f_delight_train,
positions are drawn… See the full description on the dataset page: https://huggingface.co/datasets/PRISM-Astro/h2f_delight_radius_norm_p4.Seoul_bike2