- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This method generates material properties like material class, density, Young's modulus and Poisson ratio on the Trellis generated results. The method use a local/global latent predictor to suppply base filed, projected CLIP features for global residual corrections. The method also used view alignments, visiblity checks and boundary filtering to transfer annotations from reference to genearated voxels. The evaluation shows reasonable improvments against existing methods.
Critical strengths
The basic solution is reasonable and practical. The method is also supported by evluations like the macro-F1 accuracy improvements.
The ablation study also shows the benifit of each design, altough the numbers are all closed to each other.
Critical weaknesses
I have some concern on the main objective of the paper, is the phsical properties really need to be 3D native, more importantly, need to be 3D input only? Do we need a 3D based model for this, or we better rely on VL-models?
The evaluation is on CM-25 rather than the full test set. The overall material taxonomy is somewhat coarse.
The improvements are varies on different sources, maybe the authors can have more input on this.
Also, I really don't know what is the proper way to evaluate the accuracy in practice since different outputs are in very different non-linear domain, is MAE really meaningful here?
Questions for authors
I would like the authors to discuss their motivation on the oveall design. Why train a model without leverging the knowledge of LLM-VLs. They can provide much better annotation for the input image or a rendred image, than a train-from-stratch 3D-native model.
I think the authors should at least include a basic VL benchmark here, the method can be simply render the image on multi-views and ask VL model to annotate, then back-project to voxels.
How the method works for occluded parts? Especially on alignment quality?
Having a more reasonable way for the reader understand the accuracy number means? How does this compares with a homogeneous estimation?
What is the relationship to part-segmentation? If the object is well segmented will this material assignment getting better?
11 comments
2. Evaluation beyond CM-25 and material taxonomy
CM-25 is the cross-pipeline comparison, not the entire test evaluation. It contains 13 PhysX-Mobility and 12 PixieVerse objects. The representation and supervision-transfer ablations use the larger object-disjoint test split: 225 of the 262 held-out objects have accepted physical targets.
On that larger evaluation, Table 2 reports material F1 of 0.585 for LP²SLat, 0.429 for the latent-only model, and 0.216 for the SLAT-Phys-style reimplementation.
The paper distinguishes two material metrics:
- Fine Material F1: our 16-class material vocabulary, evaluated where fine-material labels are available.
- Coarse Constitutive F1: a three-group mapping used for cross-pipeline comparison when the original label spaces differ.
3. Differences between data sources
The source-wise differences likely reflect both object composition and how closely each source matches a method’s training data and representation. PhysX-Mobility mainly contains multipart household objects, where physical-property prediction requires distinguishing materials across components. LP²SLat combines spatially organized 3D features with local image cues, which is well suited to this setting. On the PhysX-Mobility portion of CM-25, its Young’s modulus log-MAE is 0.713, compared with 6.480 for PIXIE.
PixieVerse contains a different mix of objects, including vegetation, toys, balls, and granular-appearance assets. PIXIE was developed using PixieVerse, and our comparison uses its released weights and multi-view visual features. This closer match to its training domain may help explain its lower Young’s modulus error on the PixieVerse portion: 1.192 versus 1.590 for LP²SLat.
1. Why use a lightweight predictor rather than directly query a VLM?
General-purpose VLM outputs cannot be assumed to provide reliable physical parameters without validation. PhysX-Anything explicitly notes that GPT-5 is not fully reliable for fine-grained physical reasoning when discussing its evaluator. PhysX-Omni uses GPT to produce preliminary physical annotations and then has human annotators verify and refine them. These practices show why direct VLM outputs require additional checking before being used as physical supervision.
Task-specific VLM training also requires substantial computational resources. PhysX-Anything fine-tunes Qwen2.5-VL-7B on 32 A800 GPUs for five days, while PhysX-Omni trains its VLM on 64 A100 GPUs for approximately fourteen days. These models address broader asset-generation tasks, including geometry and physical attributes. Our goal is narrower: predict spatial physical properties for an asset already produced by a single-image generator.
LP²SLat therefore reuses frozen TRELLIS and CLIP representations and trains only a 2.36M-parameter predictor. It is neither a 3D-input-only model nor a model that learns visual knowledge from scratch. TRELLIS provides spatial structure and appearance features, while the source image supplies local material cues. The submitted ablation shows that adding the image branch increases material F1 from 0.429 to 0.585. The prediction head takes 0.008 seconds per asset on an NVIDIA A40, excluding generation, feature extraction, and export.
This design provides a locally deployable physical-property predictor without task-specific VLM fine-tuning or external API calls during prediction.