YiboZhang2001/G-buffer-TexVerse

G-buffer TexVerse

       
Companion dataset of UltraTex (SIGGRAPH Asia 2026)
Yibo Zhang^1,2, Ze Yuan³, Nan Cao^4,2, Li Zhang^5,2, Yan-Pei Cao⁶, Yuan-Chen Guo⁶, Rui Ma^{1 †
1}Jilin University   ²Shanghai Innovation Institute   ³The University of Hong Kong
⁴Tongji University   ⁵Fudan University   ⁶VAST
^† Corresponding author

G-buffer TexVerse is a large-scale, ultra-high-resolution multi-view rendering dataset built on TexVerse. This release contains 351,847 BSDF assets (the pool after visual-quality and non-BSDF filtering), rendered at 2048×2048 or 4096×4096. UltraTex training applies two further filters (albedo entropy and AI-content removal) and uses a 268,365-asset subset of this release.

It is the training data behind UltraTex: Unleashing 2K Multi-View Diffusion for 3D Texturing.

Feel free to contact me (ybzhang23@mails.jlu.edu.cn) if you have any questions or suggestions.

News

• [2026-09-22] Dataset card released.
• [2026-09-22] UltraTex paper available on arXiv.

Asset curation

Raw TexVerse (858K) is filtered in four stages in the paper. This Hugging Face release is the BSDF rendering pool: 351,847 assets (after visual-quality screening and dropping non-BSDF shaders). UltraTex training then applies albedo-entropy and AI-content filters, leaving 268,365 assets. We release the larger BSDF pool, not only the training subset.

Stage	Remaining
Raw TexVerse	858K
Visual quality assessment (GPT-5 on low-resolution four-view renders)	402K
Non-BSDF material filtering	348K
Albedo entropy filtering (drop buckets 0–3)	297K
AI-generated content removal (meshy, tripo, createdwithai)	268,365

The Hugging Face release is larger than the training subset: 351,847 BSDF assets (canonical and sphere splits, same ids). UltraTex training uses the 268,365-asset subset after the last two filters.

Camera configurations

Every retained asset is rendered under both configurations. They share camera intrinsics, object normalization, aspect-ratio-adaptive distance, and the three sampled HDR environment maps; they differ only in viewpoints.

Canonical — 6 views (used by UltraTex)

Four side views at azimuths 0°, 90°, 180°, 270° with zero elevation, plus a top and a bottom view (elevations ±90°). This is the six-view layout UltraTex is trained and evaluated on.

For canonical we additionally render training reference images under the same three HDR maps. For each lighting condition, four random front-facing reference views are sampled: azimuth from [-45°, -8°] ∪ [8°, 45°] and elevation from [-5°, 0°] ∪ [5°, 20°].

Sphere — 36 views (community release)

Twelve azimuths spaced every 30° at each of three elevations {−40°, −20°, 30°}. The azimuth grid subsumes the four canonical side-view azimuths, so the two configurations stay registered. Not used by UltraTex; released for novel-view synthesis, sparse/dense-view reconstruction, PBR estimation, and texture baking.

Per-view outputs

Renderer: Blender Cycles. Every image keeps an alpha channel for the visible foreground.

Group	Contents
Geometry	Shading normals in camera space (bump_normal_camera) and world space (bump_normal_world); canonical coordinate maps (position)
Material	albedo; roughness_metallic when the asset provides them
Lighting	Shaded images render_0 / render_1 / render_2 under three HDR maps sampled from a pool of 862 Poly Haven environments
Extra (canonical only)	Training reference images render_ref_{0,1,2} (4 views each) and their poses
Cameras	pose/*.npy, intrinsics.npy, env_indices.txt

Resolution is set by the asset's native texture resolution: 2048² if the source texture is 1024, 4096² if it is ≥2048. {res} in the filename (1024 / 2048 / 4096 / 8192) records that source texture size, not the rendered pixel size. Across the 351,847 released assets the source-resolution split is 1024: 102,254; 2048: 147,261; 4096: 79,198; 8192: 23,134. UltraTex resizes all views to 2048² at training time and uses only the normal maps as geometric conditions.

Each asset's env_indices.txt stores three integer ids into a pool of 862 Poly Haven HDRIs. The index → filename / Poly Haven page mapping is env_maps.json (ids are the os.listdir order used at render time, not alphabetical). For the example asset below, [728, 130, 559] is zwartkops_curve_sunset, music_hall_01, abandoned_hopper_terminal_02.

Repository layout

env_maps.json                   # 862 Poly Haven HDRIs; env_indices.txt indexes this list
canonical/bsdf/{00..ff}.zip     # 256 buckets, 351,847 assets, ~4.25 TB
sphere/bsdf/{00..ff}.zip        # 256 buckets, same 351,847 assets, ~21.09 TB

~25.3 TB in total. Both splits cover the same 351,847 assets, sharded by the first two hex characters of the TexVerse id. Each {prefix}.zip is a bucket of per-asset zips. Example below is a PBR asset: 0000ecca9a234cae994be239f6fec552_1024.

Canonical example (canonical/bsdf/00.zip)

Verified on the live Hugging Face file: 1,388 item zips + 00.manifest.jsonl. Six views (000–005) plus three groups of four reference views. This example includes roughness_metallic/.

canonical/bsdf/00.zip
 ├── 00.manifest.jsonl
 └── 0000ecca9a234cae994be239f6fec552_1024.zip
      └── 0000ecca9a234cae994be239f6fec552_1024/
           ├── albedo/000.webp … 005.webp
           ├── bump_normal_camera/000.webp … 005.webp
           ├── bump_normal_world/000.webp … 005.webp
           ├── position/000.webp … 005.webp
           ├── roughness_metallic/000.webp … 005.webp
           ├── render_0/  render_1/  render_2/     # 000.webp … 005.webp each
           ├── pose/000.npy … 005.npy
           ├── render_ref_0/000.webp … 003.webp, env_id.txt
           ├── render_ref_1/000.webp … 003.webp, env_id.txt
           ├── render_ref_2/000.webp … 003.webp, env_id.txt
           ├── render_ref_0_poses/000.npy … 003.npy
           ├── render_ref_1_poses/000.npy … 003.npy
           ├── render_ref_2_poses/000.npy … 003.npy
           ├── env_indices.txt                     # [728, 130, 559]
           └── intrinsics.npy

Sphere example (sphere/bsdf/00.zip)

Same asset id, 36 views (000–035), no reference images. 00.zip is a flat list of item zips (no manifest).

sphere/bsdf/00.zip
 └── 0000ecca9a234cae994be239f6fec552_1024.zip
      └── 0000ecca9a234cae994be239f6fec552_1024/
           ├── albedo/000.webp … 035.webp
           ├── bump_normal_camera/000.webp … 035.webp
           ├── bump_normal_world/000.webp … 035.webp
           ├── position/000.webp … 035.webp
           ├── roughness_metallic/000.webp … 035.webp
           ├── render_0/  render_1/  render_2/     # 000.webp … 035.webp each
           ├── pose/000.npy … 035.npy
           ├── env_indices.txt                     # same three HDR ids as canonical
           └── intrinsics.npy

Download

Install huggingface_hub:

pip install -U huggingface_hub
huggingface-cli login

Download one bucket (canonical 6-view is enough for UltraTex-style training):

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="YiboZhang2001/G-buffer-TexVerse",
    filename="canonical/bsdf/00.zip",
    repo_type="dataset",
)

Or snapshot selected folders:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="YiboZhang2001/G-buffer-TexVerse",
    repo_type="dataset",
    allow_patterns=["canonical/bsdf/00.zip"],
)

Unpack a canonical bucket, then the asset:

mkdir -p canonical/00 && unzip canonical/bsdf/00.zip -d canonical/00
unzip canonical/00/0000ecca9a234cae994be239f6fec552_1024.zip -d canonical/00

Sphere buckets are larger (~80 GB). Same unpack pattern:

mkdir -p sphere/00 && unzip sphere/bsdf/00.zip -d sphere/00
unzip sphere/00/0000ecca9a234cae994be239f6fec552_1024.zip -d sphere/00

License

This dataset is released under CC0 1.0 Universal. You may copy, modify, distribute, and use it, including for commercial purposes, without asking permission.

HDR environment maps used for lighting are from Poly Haven (also CC0).

Citation

If you use G-buffer TexVerse, please cite UltraTex and TexVerse:

@inproceedings{zhang2026ultratex,
  author    = {Zhang, Yibo and Yuan, Ze and Cao, Nan and Zhang, Li and
               Cao, Yan-Pei and Guo, Yuan-Chen and Ma, Rui},
  title     = {UltraTex: Unleashing 2K Multi-View Diffusion for 3D Texturing},
  year      = {2026},
  isbn      = {9798400728426},
  publisher = {Association for Computing Machinery},
  address   = {Kuala Lumpur, Malaysia},
  url       = {https://doi.org/10.1145/3829340.3842299},
  doi       = {10.1145/3829340.3842299},
  booktitle = {Proceedings of the SIGGRAPH Asia 2026 Conference Papers},
  series    = {SA Conference Papers '26},
}

@article{zhang2025texverse,
  title   = {TexVerse: A Universe of 3D Objects with High-Resolution Textures},
  author  = {Zhang, Yibo and Zhang, Li and Ma, Rui and Cao, Nan},
  journal = {arXiv preprint arXiv:2508.10868},
  year    = {2025}
}