Qualitative results

Every qualitative comparison of the paper, at full resolution. Click a figure to open it.

Same protocol as the paper — the decoder, renderer and Stage 2 fitting are fixed and only the Stage 1 training data changes. Every recovered material is relit in a common scene: the same cloth mesh draped over a bar, with identical UV tiling, lighting and camera for every method, so any visible difference is due solely to the learned material. Columns are the decoder trained on RoboCloth, UBOFAB19 or MERL, the analytic PBR fit, and the ground truth where one exists.

UBO2014 — all 12 held-out materials

Materials 1–6
UBOFAB19MERLPBRRoboClothGT
carpet02carpet07carpet09carpet12fabric02fabric04
Extends the close-up to every held-out UBO2014 material. Across all 12, the RoboCloth decoder most consistently matches the GT reference in both colour and weave texture.
Materials 7–12
UBOFAB19MERLPBRRoboClothGT
fabric09fabric11felt01felt03felt05felt10
Materials 7–12; same columns and scene as above.

RoboCloth and UBOFAB19 test sets

We also reconstruct held-out materials from our own test set (in-domain) and from the UBOFAB19 test set (cross-dataset): for each material the decoder is frozen and a dense latent grid is optimized on a subset of the captured views, then relit in the same scene.

RoboCloth test set — five held-out materials
RoboClothUBOFAB19MERLPBR
226314370145452
No ground-truth render exists for a captured material. The RoboCloth-trained decoder recovers the sharpest and most faithful texture and colour; all four reconstructions render at high visual quality on the mesh, and the image-space comparison below shows that ours is nonetheless the most faithful to the photographs.

Renderingpath traced at 2048×2048, 512 spp, same scene and environment map.

UBOFAB19 test set — five held-out materials
RoboClothUBOFAB19MERLPBR
3183773222637
On UBOFAB19’s own test distribution the UBOFAB19-trained decoder is strongest, with the RoboCloth decoder a close second; both clearly outperform the MERL-trained decoder and the PBR baseline.

Renderingpath traced at 2048×2048, 512 spp, same scene and environment map.

Image-space comparison

Image-space comparison. For the materials of our own test set we also compare each reconstruction directly with the captured photographs: the reconstruction is rendered into four of the material’s captured views with that view’s calibrated camera and light, and shown next to the photograph (GT); cells are cropped to the sample.

Image-space comparison — material 226
GTRoboClothUBOFAB19MERLPBR
RoboCloth reproduces the ground-truth colour most faithfully; the UBOFAB19 and MERL reconstructions drift toward yellow and magenta.

Renderingeach view with the capture’s calibrated camera and light, 16 spp (the paper’s test setting), at capture resolution; the same for the four materials below.

Image-space comparison — material 452
GTRoboClothUBOFAB19MERLPBR
RoboCloth recovers more of the specular highlight.
Image-space comparison — material 370
GTRoboClothUBOFAB19MERLPBR
RoboCloth stays sharper at the grazing views, where the other methods blur the weave.
Image-space comparison — material 314
GTRoboClothUBOFAB19MERLPBR
Material 314, four captured views.
Image-space comparison — material 145
GTRoboClothUBOFAB19MERLPBR
Material 145, four captured views.

Back to the main page