Pretrained BRDF decoders, ready-to-render neural cloth materials, the scenes and media from the RoboCloth paper, and the low-resolution proxies behind the project-page viewer. Every RoboCloth stage-2 checkpoint is a complete material — a dense latent texture plus its frozen decoder — so you can render it without downloading any capture data (the Bonn and UBO stage-2 checkpoints store a per-point latent bank instead; see Details).

Capture dataset: theialab/RoboCloth
Code: https://github.com/theialab/robocloth (Apache-2.0)
Project page: https://theialab.github.io/robocloth/

Quick start

pip install -U huggingface_hub
hf download theialab/RoboCloth-assets --repo-type dataset \
    --include "checkpoints/stage2/RoboCloth/145/*" --local-dir ./robocloth-assets   # one material, all 4 variants, 4.4 GB (1.2 GB each)

Render it with rendering/render.py from the code repository: a scene folder holds a Mitsuba 3 scene.xml, its meshes and a materials.json that assigns checkpoints to shape ids (see Details → Using the checkpoints). Quote every --include pattern.

Layout

checkpoints/
  stage1/{Ours,Bonn,MERL}.ckpt        # shared BRDF decoders                         ~4.2 GB
  stage2/RoboCloth/<13 materials>/    # RoboCloth test + teaser materials            ~31 GB
  stage2/Bonn/<5 materials>/          # held-out UBOFAB19 materials                  ~2.1 GB
  stage2/UBO/<12 materials>/          # held-out UBO2014 materials                   ~2.5 GB
render_assets/
  cloth_on_bar/                       # onBars01_st_hp.ply, pole_spheres_v4.obj      ~22 MB
  teaser_room/                        # room.xml, meshes/ (35 .ply), textures/ (1)   ~123 MB
renders/
  cloth_on_bar/cloth_on_bar_2048_spp512.{png,exr}   # reference render, 2048², 512 spp
  teaser/teaser_1080p_spp8.{png,exr}                # draft teaser render, 1920×1080, 8 spp
media/
  teaser.png                          # paper Fig. 1
  capture_314_5x.mp4                  # one material being captured, 5x speed
  web/cloth_with_sphere_final.mp4     # project-page video
web/
  proxies/<id>/{proxy.tar,index.json,meta.json}   # 640 px WebP scrubbing proxies for the
                                      # project-page viewer, one folder per material  ~8.85 GB

Total ≈ 49 GB: the stage-2 RoboCloth checkpoints (31 GB) and the 500 proxy folders (1 500 files, 8.85 GB) dominate.

Details

Checkpoint naming and contents

checkpoints/<stage>/<test set>/<material>/<decoder source>_epoch<N>.ckpt

Stage 1 — the shared MLP BRDF prior, one per training corpus: Ours.ckpt (trained on RoboCloth, 2.0 GB), Bonn.ckpt (UBOFAB19, 2.2 GB), MERL.ckpt (MERL measured BRDFs, 1.5 MB — its latent bank has 120 rows instead of millions of surface points, hence the size). Stage 2 loads only the material.decoder.* tensors from these.
Stage 2 — one fitted material. The prefix names the frozen stage-1 decoder the fit started from: Ours (RoboCloth prior), Bonn, MERL, or PBR for the analytic Disney baseline fitted in place of a neural decoder.
epoch<N> is the checkpoint's stored epoch plus one, i.e. epoch80 was written after 80 completed epochs (stored epoch index 79). It is part of the filename, not a knob — match it with a glob rather than typing it.

Each neural stage2/RoboCloth file holds material.latent_texture.params (1,46,2048,2048) (C = 24 latent + 16 geometry + 6 frame), a frozen copy of material.decoder.*, the per-channel scale material.factor (3,), and the parallax query material.neural_geometry.*; the PBR variant stores 34 parameter channels and a scalar material.factor. The stage2/Bonn and stage2/UBO files instead store the material as a per-point latent bank material.point_latent_bank.weight (V,30) (UBO: a 400×400 texel grid) plus decoder and factor — no dense texture and no neural_geometry; the renderer's UBO/Bonn wrapper classes load them (Bonn additionally needs the UBOFAB19 point-grid metadata).

What is released — materials and variants

Set	Materials	Variants per material
`stage2/RoboCloth`	5 paper test materials — 145, 226, 314, 370, 452	`Ours`, `Bonn`, `MERL`, `PBR`
`stage2/RoboCloth`	8 teaser materials — 8, 9, 26, 50, 190, 311, 367, 453	`Ours` only
`stage2/Bonn`	32, 37, 226, 318, 377	`Ours`, `Bonn`, `MERL`, `PBR`
`stage2/UBO`	carpet02, carpet07, carpet09, carpet12, fabric02, fabric04, fabric09, fabric11, felt01, felt03, felt05, felt10	`Ours`, `Bonn`, `MERL`, `PBR`

Material ids under stage2/RoboCloth/ are the same ids as in the capture dataset; the Bonn and UBO sets are the held-out materials of the two external benchmarks used in the paper (UBOFAB19 and UBO2014).

Third-party datasets are not redistributed here. The MERL BRDF database and the UBO2014 BTFs have their own licenses and must be obtained from their owners; the code repository documents how to point the comparison scripts at your own copies. Only checkpoints trained on them are released, under this repository's license.

Download recipes

The hf command ships with huggingface_hub (pip install -U huggingface_hub). Quote every --include pattern — an unquoted * is expanded by the shell before it reaches the include filter.

The five paper materials plus the eight teaser materials (~31 GB):

hf download theialab/RoboCloth-assets --repo-type dataset \
    --include "checkpoints/stage2/RoboCloth/*" --local-dir ./robocloth-assets

The pretrained decoder, if you want to fit your own materials (~2 GB):

hf download theialab/RoboCloth-assets --repo-type dataset \
    --include "checkpoints/stage1/Ours.ckpt" --local-dir ./robocloth-assets

The bundled example scenes:

hf download theialab/RoboCloth-assets --repo-type dataset \
    --include "render_assets/*" --local-dir ./robocloth-assets

The viewer proxies for one material (7–31 MB; 145 is 31 MB):

hf download theialab/RoboCloth-assets --repo-type dataset \
    --include "web/proxies/145/*" --local-dir ./robocloth-assets

Using the checkpoints — materials.json and the example scenes

Set up the rendering environment and run rendering/render.py from the code repository — its README walks through the whole path. A scene is a folder with a Mitsuba 3 scene.xml, its meshes, and a materials.json that assigns checkpoints to shape ids:

{
  "checkpoint_root": "/absolute/path/to/robocloth-assets/checkpoints/stage2/RoboCloth",
  "assignments": {
    "sofa":    {"material": "370"},
    "curtain": {"material": "145", "uv_tiling": 12.0}
  }
}

"material": "<id>" resolves to <checkpoint_root>/<id>/Ours_epoch*.ckpt (the glob must match exactly one file), so you never hard-code the epoch; "ckpt" takes an explicit path instead. Shapes not listed keep the BSDF from the XML. Meshes need UVs — the material is a latent texture — and uv_tiling sets the weave repeat density. The repository's rendering/examples/cloth_on_bar and rendering/examples/teaser scenes consume render_assets/cloth_on_bar/ and render_assets/teaser_room/, relocated through the MESH_DIR, TEASER_SCENE_ROOT and BRDF_CKPT_ROOT environment variables. The stage-1 decoder is the starting point for fitting a new material (scripts/train_stage2.sh).

Renders and media

renders/cloth_on_bar/ is the reference output of the bundled cloth-on-bar example at 2048×2048 and 512 spp, as a tone-mapped PNG and the linear EXR — useful for checking that your setup reproduces the released result; renders/teaser/ is a draft-quality (8 spp) render of the bundled teaser scene for the same purpose. media/teaser.png is paper Fig. 1, media/capture_314_5x.mp4 shows one material being captured at 5× speed, and media/web/ holds the project-page video.

Project-page proxies — web/proxies/

The material viewer on the project page scrubs through every captured frame of a material without touching the multi-gigabyte HDR release. For each of the 500 materials, web/proxies/<id>/ holds:

File	What it is
`proxy.tar`	one ustar of all ≈580 frames as 640 px (long side) WebP, quality 90, storing `sqrt(linear)` in 8 bits; file names `scan-<id>_light-<l>_camera-<c>.webp`
`index.json`	frame name → `[byte offset, length]` inside `proxy.tar` (plus one `__meta` entry), so a single frame can be fetched by HTTP range request
`meta.json`	`count`, `long` (640), `quality` (90), `encoding` ("sqrt"), `source_hw` (2048×3072), `source`, `bytes`

The viewer undoes the square-root encoding with its tone LUT before display. The proxies are derived from the same frames as materials/<id>/hdr.tar in the capture dataset and are not meant for quantitative use.

Provenance of the teaser room scene — third-party assets, license carve-out

render_assets/teaser_room/ is a third-party interior scene built from freely available assets — the meshes and the textures/ image were not created by the RoboCloth authors. It is included so the paper's teaser figure can be reproduced; only the cloth surfaces in it are RoboCloth materials. Those assets keep their original terms and are not covered by this repository's CC BY 4.0 grant; check them before redistributing the folder on its own.

Citation

@misc{li2026robocloth,
  title  = {RoboCloth: A Large-Scale Real Cloth Material Dataset for Neural Reflectance Reconstruction},
  author = {Li, Zhen and Qin, Haoran and Di Sario, Francesco and Rebain, Daniel and Tagliasacchi, Andrea},
  year   = {2026}
}

License and contact

CC BY 4.0 (LICENSE), except the third-party teaser-room scene assets (free assets, original terms apply — see Details). The code repository is Apache-2.0. Please open a discussion here or an issue on https://github.com/theialab/robocloth.

Downloads last month: 185

Total file size:

49.3 GB