Supplementary resultsAnonymous submission
Reviewer 2Supplementary experiments & responses

Interaction tasks & cross-dataset examples

Interaction-focused benchmarking and three additional composition cases from 3DCoMPaT200.

Route · Diagnose · RefineEvidence accompanying the author response
01

Dataset choice · task suitability and evaluation scale

Why the benchmark needs interaction tasks

“The evaluation uses 194 cases, and one source dataset is relatively old. Could a newer dataset such as 3DCoMPaT200 be used?”

Our task composes existing assets while preserving their identity and appearance outside the required changes. A useful test case therefore needs more than two shapes placed next to one another: the instruction should require nontrivial contact, attachment, deformation, or adaptation around an interaction region.

For example, a character holding a cup requires the hands and cup to form a plausible contact; combining a balloon and basket into a hot-air balloon requires a coherent attachment structure. A collection of individually annotated shapes does not automatically provide these paired assets, interaction instructions, and evaluation references.

We agree that broader and more recent source assets are valuable. Dataset age alone does not determine task suitability, and the number of single shapes is different from the number of constructed interaction tasks. Nevertheless, 194 compositions remain a limited evaluation set, and broader coverage is needed.

Within the rebuttal period, we have not completed a full evaluation on 3DCoMPaT200. We instead build a small subset from this dataset, conducted experiemnts on that, and provide three selected composition cases selected from our runs, with TRELLIS and VineDresser3D comparisons, as an additional qualitative check on a different asset source.

02

Cross-dataset evidence · edit instructions and results

Additional 3DCoMPaT200 compositions

“Does your method improve editing in other aspects such as robustness?”

The cases below examine whether the pipeline can be applied to assets from another source while retaining the same interaction-composition objective. The comparison focuses on contact quality, preservation of source appearance, and consistency across the three rendered views of the result.

This is a qualitative cross-dataset check. It does not measure robustness to random seeds, prompt paraphrases, adversarial targets, or the full range of shapes and materials in 3DCoMPaT200. We distinguish this limited evidence from a general robustness claim.

3DCOMPAT200 / 01

Balloon and basket · suspended gondola

The instruction, source assets, and three-view method results are shown below. Inspect attachment or contact quality together with preservation of the supplied asset appearance.

Edit instruction

Turn the supplied basket into the balloon's new gondola. Hang it level below the burner. Bend and extend the existing suspension lines so they loop around four separated points on the basket's upper frame, and pinch the frame slightly at the tie points. Keep the balloon envelope and the basket weave recognizable.

Input assets

Three rendered views per asset · click any image to enlarge

Input asset ATHREE VIEWS
Input asset A · view 1
Input asset A · view 2
Input asset A · view 3
Input asset BTHREE VIEWS
Input asset B · view 1
Input asset B · view 2
Input asset B · view 3

Composition results

Three rendered views per asset · click any image to enlarge

OursTHREE VIEWS
Ours · view 1
Ours · view 2
Ours · view 3
TRELLISTHREE VIEWS
TRELLIS · view 1
TRELLIS · view 2
TRELLIS · view 3
VineDresser3DTHREE VIEWS
VineDresser3D · view 1
VineDresser3D · view 2
VineDresser3D · view 3
3DCOMPAT200 / 02

Teddy bear and mug · two-paw grasp

The instruction, source assets, and three-view method results are shown below. Inspect attachment or contact quality together with preservation of the supplied asset appearance.

Edit instruction

Pose the teddy bear holding the mug in front of its chest. Curl its right plush paw around the mug handle and bend its left forearm so the other paw cups the underside. Flatten the plush slightly at the contact points. Keep the mug upright and preserve the bear's face, ears, and body silhouette.

Input assets

Three rendered views per asset · click any image to enlarge

Input asset ATHREE VIEWS
Input asset A · view 1
Input asset A · view 2
Input asset A · view 3
Input asset BTHREE VIEWS
Input asset B · view 1
Input asset B · view 2
Input asset B · view 3

Composition results

Three rendered views per asset · click any image to enlarge

OursTHREE VIEWS
Ours · view 1
Ours · view 2
Ours · view 3
TRELLISTHREE VIEWS
TRELLIS · view 1
TRELLIS · view 2
TRELLIS · view 3
VineDresser3DTHREE VIEWS
VineDresser3D · view 1
VineDresser3D · view 2
VineDresser3D · view 3
3DCOMPAT200 / 03

Fox and pixel sunglasses · fitted eyewear

The instruction, source assets, and three-view method results are shown below. Inspect attachment or contact quality together with preservation of the supplied asset appearance.

Edit instruction

Fit the pixel sunglasses onto the front of the fox's face at a realistic scale. Rest the bridge over the snout and angle the temple arms back along the sides of the head below the ears, slightly indenting the cheek fur at the contact points. Preserve the blocky pixel shape of the glasses and the fox's sitting pose, ears, and tail.

Input assets

Three rendered views per asset · click any image to enlarge

Input asset ATHREE VIEWS
Input asset A · view 1
Input asset A · view 2
Input asset A · view 3
Input asset BTHREE VIEWS
Input asset B · view 1
Input asset B · view 2
Input asset B · view 3

Composition results

Three rendered views per asset · click any image to enlarge

OursTHREE VIEWS
Ours · view 1
Ours · view 2
Ours · view 3
TRELLISTHREE VIEWS
TRELLIS · view 1
TRELLIS · view 2
TRELLIS · view 3
VineDresser3DTHREE VIEWS
VineDresser3D · view 1
VineDresser3D · view 2
VineDresser3D · view 3
Figure