acceptodds
Withdrawn from ICLR 2027

FreePart: Towards Generation-Friendly Decomposition for 3D Part Synthesis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Part-level 3D generation aims to synthesize complex 3D objects by decomposing them into semantically meaningful parts and generating them separately before assembly. However, we find that even semantically plausible decompositions can lead to substantially different generation difficulties and 3D reconstruction quality. When such decomposition choices are unfavorable, they can introduce errors that propagate through downstream generation, compromising the structural coherence and completeness of the resulting objects. To address this, we propose FreePart, a plug-and-play framework that searches for generation-friendly part decompositions and seamlessly integrates them into diverse part-level 3D generation pipelines. Specifically, we first introduce a Part-level Decomposition Assessment (PDA) to evaluate the generation difficulty induced by different decompositions. Guided by PDA, we then design a Multi-agent Collaborative Search (MCS) method, where three complementary agents iteratively diagnose, refine, and evaluate candidate decompositions, with valid decomposition states organized as a directed acyclic graph for efficient search. Extensive experiments across multiple generation pipelines demonstrate that FreePart consistently improves both part-level and overall-object 3D generation quality. Code and website are available in the supplementary material.

Primary area:
generative models
Keywords:
3D Generation, Part Decomposition, Multi-Agent Search

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_FdZzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FreePart, a framework that improves part-level 3D generation by refining how an object is divided into parts. It introduces Part-level Decomposition Assessment using surface complexity and occlusion-based risks, and trains a VLM through SFT and GDPO. The Diagnoser, Strategist, and Evaluator then guide iterative adjacent-part merges through a DAG-based search, accepting or rejecting proposals based on predicted SC/OCC risks and the preservation of part-level logic. It consistently improves part-level and overall 3D generation quality across diverse generation pipelines. The motivation is convincing, but some implementations are not well explained.

Critical strengths
  1. The paper treats decomposition as an optimizable component. It offers a well-motivated way to improve 3D generation by refining part decomposition without modifying the generator, making it applicable to existing pipelines.

  2. Reported improvements across three generators support the practical value of decomposition refinement.

Critical weaknesses
  1. The non-monotonic relationship between OCC and the generation quality, as well as the reference value of 0.6, directly determine the training labels and search decisions, but are insufficiently justified. The paper does not adequately explain how this relationship was established or what evidence supports it, and does not explain why lower occlusion can yield higher errors.

  2. The paper does not report end-to-end latency, average VLM call counts, or actual search overhead. FreePart is designed to avoid the cost of exhaustive generation and Sec. 1 emphasizes efficient exploration. Given the sequential execution of three roles using a 72B VLM, it remains unclear whether FreePart achieves better efficiency.

  3. The paper states that QSCQ_{\mathrm{SC}} maps continuous SC values to low, medium, and high risk (Sec. 3), but does not specify the intervals or selection procedure. Their justification and influence on the reported results remain insufficiently examined.

  4. The fixed-rule strategy ablation in Table 5 is insufficiently described. Table 5 replaces the adaptive Strategist with fixed rules, but does not specify how these rules select merge candidates.

Questions for authors
  1. The paper emphasizes DAG-based search and the complementary roles of the three agents. Table 5 shows that removing a certain capability reduces performance, but it does not isolate the benefits of separating decisions into three roles or adopting the proposed search structure. How would the method perform with more intuitive alternative strategies, such as greedy merging using the same SC/OCC signals, single-agent decision-making, or directly predicting the final grouping? These comparisons may help clarify the benefits of the proposed design.

  2. What is the average number of VLM calls per input, the computational cost of the search stage, and the end-to-end runtime?

4 comments
AuthorsOpenReview ↗

FdZz — Response 1/3

We thank the reviewer for recognizing the motivation, plug-and-play applicability, and effectiveness across different generation pipelines of our work. Below, we address your concerns point by point.

Q1. Non-monotonic OCC trend and empirical reference.

The non-monotonic relationship between OCC and the generation quality, as well as the reference value of 0.6, directly determine the training labels and search decisions, but are insufficiently justified. The paper does not adequately explain how this relationship was established or what evidence supports it, and does not explain why lower occlusion can yield higher errors.

We thank the reviewer for requesting a more explicit basis for the OCC reference. We divide 1,913 objects into five equal-frequency bins according to GT object-level OCC, yielding:

R1. Relationship Between OCC and Reconstruction Quality.

GT OCC rangeObjectsMean OCCOverall CD ↓F1-0.1 ↑F1-0.05 ↑
[0, 0.535]3830.3440.14000.79400.5994
(0.535, 0.633]3820.6000.12330.82890.6313
(0.633, 0.708]3830.6700.12630.82700.6243
(0.708, 0.787]3820.7470.13440.82180.6001
(0.787, 0.983]3830.8530.14370.78970.5624

The interval centered near 0.6 has the lowest mean error and highest mean F-scores, providing a data-driven reference for this evaluated setting. We carried out 2,000 bootstrap resamples per bin to examine uncertainty.

Based on our observations of the dataset, parts with low OCC often exhibit thin, sheet-like geometries. Although these parts are largely visible, their thickness is difficult to infer accurately from 2D observations, often leading to geometric distortions and higher reconstruction errors. This provides a plausible explanation for why lower OCC does not necessarily correspond to better generation quality.

Q2. Computational Efficiency and End-to-End Inference Cost.

Q2.1 Actual VLM Calls, Search Overhead, and End-to-End Time.

The paper does not report end-to-end latency, average VLM call counts, or actual search overhead.

Candidate previews do not repeatedly invoke the downstream full 3D generator. We benchmark all methods after warm-up, using identical initial masks and generation seed 42 on three A800 GPUs. All three search methods share a maximum budget of six VLM calls per object. The results are summarized below.

R2. Reconstruction Quality and End-to-End Inference Efficiency.

PipelineCalls / obj ↓Search time ↓End-to-end ↓CD↓F1-0.1 ↑F1-0.05 ↑
Original OmniPart0—76.1 s0.123100.840980.57624
Single-path greedy1.3143.2 s207.7 s0.122170.844930.58305
Risk-greedy5.7319.7 s386.2 s0.107770.880180.63897
FreePart3.9260.6 s328.8 s0.105500.885680.64877

As shown in Table R2, FreePart achieves the best reconstruction quality across all metrics, while requiring 31.6% fewer VLM calls and 18.5% shorter search time than risk-greedy. Although single-path greedy is faster, its reconstruction quality is substantially lower than FreePart's. Compared with search-free OmniPart, FreePart introduces additional computational overhead but significantly improves reconstruction quality. These results demonstrate that FreePart achieves a favorable balance between reconstruction quality and search efficiency.

Q2.2 Stage-Wise Inference Cost and Search Efficiency.

FreePart is designed to avoid the cost of exhaustive generation and Sec. 1 emphasizes efficient exploration. Given the sequential execution of three roles using a 72B VLM, it remains unclear whether FreePart achieves better efficiency.

FreePart costs 252.7 s more than search-free OmniPart but completes 57.4 s faster than risk-greedy and saves 31.6% actual VLM calls. Both search variants reduce mean part count from 7.8 to 6.9 on this subset. We therefore distinguish avoiding repeated full generation during candidate search from reducing end-to-end runtime relative to the original generator.

R3. Stage-Wise Breakdown of Inference Time.

PipelineInitial seg. ↓Search ↓Final gen. ↓Other ↓Total ↓
OmniPart22.7—45.97.576.1
FreePart + OmniPart23.0260.637.18.1328.8
Risk-greedy + OmniPart22.5319.736.67.5386.2
AuthorsOpenReview ↗

FdZz — Response 2/3

Q3. SC Discretization and Threshold Sensitivity.

The paper states that (Q_{\mathrm{SC}}) maps continuous SC values to low, medium, and high risk (Sec. 3), but does not specify the intervals or selection procedure. Their justification and influence on the reported results remain insufficiently examined.

SC risks are defined using training-set percentile thresholds: low ≤ Q30, medium (Q30, Q70], and high > Q70. We further evaluate three percentile configurations (25/75, 30/70, and 35/65), with 30/70 achieving the highest SC risk classification accuracy, as shown in Table R4.

R4. Sensitivity of SC Risk Prediction

SC Percentile CutoffsSC ThresholdsClassification Accuracy ↑
25% / 75%0.00901 / 0.0612482.40%
30% / 70% (Ours)0.01667 / 0.0558985.11%
35% / 65%0.02338 / 0.0511377.16%

Separately, risk-stratified reconstruction results show that the high-SC group is more challenging (baseline Part CD 34.51 versus 26.85/26.93 in the low/medium groups) and benefits from FreePart (Part F1-0.1: 36.45 → 44.85).

R5. Reconstruction Performance Across SC Risk Levels.

SC RiskMethodPart-level Bbox IoU ↑Part-level CD ↓Part-level F1-0.1 ↑Part-level F1-0.05 ↑Overall-level CD ↓Overall-level F1-0.1 ↑Overall-level F1-0.05 ↑
LowOmniPart23.8626.8549.0927.4114.8777.0849.90
LowOmniPart + FreePart26.9023.6355.1732.1913.9280.5154.92
MediumOmniPart23.3126.9349.2527.4314.6677.4052.29
MediumOmniPart + FreePart24.4024.4853.9531.2714.0079.8654.59
HighOmniPart20.9134.5136.4519.3216.5176.9560.84
HighOmniPart + FreePart24.7531.5944.8523.8416.1577.1461.29
AllOmniPart23.3127.3748.4426.9414.8077.3052.21
AllOmniPart + FreePart25.0724.6953.7231.0614.1379.8555.05

Q4. Fixed-rule Strategist baseline.

The fixed-rule strategy ablation in Table 5 is insufficiently described. Table 5 replaces the adaptive Strategist with fixed rules, but does not specify how these rules select merge candidates.

We apologize for the insufficient description of the fixed-rule Strategist ablation. The Diagnoser and Evaluator remain unchanged, while Strategist is replaced by deterministic contact-based selection. From the current 2D part masks, we enumerate directly contacting and allowed nearby pairs; candidates are ranked by direct contact first, then contact-pixel count, and then proximity-overlap score. The highest-ranked candidate not previously rejected is previewed and passed to the original Evaluator. This ablation changes candidate-selection policy while retaining the remaining diagnosis and verification mechanisms.

AuthorsOpenReview ↗

FdZz — Response 3/3

Q5. Alternative decision organizations and search.

The paper emphasizes DAG-based search and the complementary roles of the three agents. Table 5 shows that removing a certain capability reduces performance, but it does not isolate the benefits of separating decisions into three roles or adopting the proposed search structure. How would the method perform with more intuitive alternative strategies, such as greedy merging using the same SC/OCC signals, single-agent decision-making, or directly predicting the final grouping? These comparisons may help clarify the benefits of the proposed design.

We add risk-greedy, single-agent iterative, single-agent direct grouping, and three-role direct grouping baselines. The iterative single-agent combines proposal and verification; the two direct-grouping variants predict the final grouping without the same stepwise accepted-merge trajectory. Results from the exploratory comparison subset are:

R6. Comparison of Alternative Search Strategies.

MethodBbox IoU ↑Part CD ↓Part F1-.1 ↑Part F1-.05 ↑Overall CD ↓Overall F1-.1 ↑Overall F1-.05 ↑
FreePart0.2051750.2395380.5592850.3271150.1316500.8142590.574048
Risk-greedy0.2030030.2400710.5585390.3258330.1320640.8135400.572634
Single-agent iterative0.1961340.2445010.5452020.3176090.1326110.8142360.571939
Single-agent direct0.1936000.2448440.5405100.3067850.1323810.8088480.555346
Three-role direct0.1958700.2500970.5294760.2972120.1317740.8064700.570741

FreePart obtains the best mean values across the seven reported metrics. In a separate equal-maximum-call comparison, FreePart also improves Overall CD (0.10550 vs. 0.10777) with fewer actual VLM calls (3.9 vs. 5.7) than risk-greedy.

Finally, the DAG is the space of reachable decompositions; the implemented controller follows a constrained sequential path through accepted merges, rather than exhaustive DAG traversal. Rollback rejects an uncommitted candidate preview at the current state and does not undo a previously accepted merge.

Q6. Inference Cost and Runtime.

What is the average number of VLM calls per input, the computational cost of the search stage, and the end-to-end runtime?

As detailed in Q2.1–Q2.2 (Tables R2–R3), FreePart requires an average of 3.9 VLM calls per object, with 260.6 s search time and 328.8 s end-to-end runtime. These measurements provide a breakdown of FreePart's computational overhead and practical inference cost.

Thank you again for your insightful suggestions. We have provided additional experiments on SC/OCC calibration, inference efficiency, threshold sensitivity, and alternative search strategies. We sincerely appreciate your time and effort and would be grateful for any further feedback before the discussion period concludes.

Reviewer_2KH1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper takes the initial part decomposition in decompose-then-generate 3D pipelines and treats it as an optimizable quantity rather than a fixed pre-processing step. FreePart fine-tunes Qwen2.5-VL-72B to predict per-part generation risks using supervision from two mesh-derived proxies: surface complexity and occlusion rate (PDA). The model then operates under three role prompts (Diagnoser, Strategist, Evaluator) to guide a merge-only DAG search (MCS). A merge is accepted only if the predicted functional logic is preserved and no risk has increased; the Strategist is trained to propose a merge that reduces at least one risk. The downstream generation pipeline is then run on the final partition. Experiments are performed with FreePart integrated with OmniPart, Hunyuan3D-Part and MIDI-3D across PartObjaverse-Tiny, PartNeXt and 3D-Front-MIDI. The authors report improvements in Bbox IoU as well as in part-level and object-level CD and F-score, with support from a pairwise user study and ablations.

Critical strengths
  1. The problem is well-motivated and under-explored. The paper focuses on how the initial part decomposition affects downstream generation quality and makes a credible case for optimizing this stage rather than treating it as fixed preprocessing.
  2. The fundamental search formulation is straightforward and well-structured. The merge only DAG, lazy edge creation, adjacency constraints, per node rejected action masks and the finite termination argument all combine to make the search procedure very easy to follow.
  3. A single checkpoint is evaluated across three input modalities and three backends without any generator specific retraining. The authors report whole object metric improvements and user study preference across all settings, providing evidence beyond the part level metrics which are impacted by the matching issue.
Critical weaknesses
  1. The benefit over simple coarsening remains unclear. FreePart only does part merging, gives no initial/final part counts, and makes no comparison to simple merging at the same part count. The instruction to count the removal of an unnecessary internal boundary as an SC improvement further blurs the distinction between generation-aware guidance and segmentation repair. Although the SC/OCC ablations show incremental benefits, unmodified OmniPart is missing from that same ablation set.
  2. The part-level evaluation doesn’t penalise unmatched parts. CD, F-score and Bbox IoU are all averaged only over the matched pairs, so unmatched parts incur no penalty. This makes the comparison sensitive to the number of parts being reduced. Missing predictions are also omitted, which can alter the denominator of the evaluation between different methods. Results should be reported with unmatched-part penalties and with the same object set.
  3. The proxy is not sufficiently validated. The OCC calibration omits the generator, sample counts, uncertainty and size/complexity controls. Neither the accuracy of the risk prediction nor the relationship between the predicted risk reduction and the actual improvement in reconstruction is reported. The transfer to occlusion-robust models (e.g. SAM 3D) and the OCC observation protocols for mesh and scene inputs are not described.
  4. Practical cost and statistical reliability are both poorly assessed. Per-object search time, VLM call counts and the actual search budget are all unreported. Seeds are fixed, but no results from repeated seeds are given and no confidence intervals are provided, so small ablation differences are very hard to interpret.
Questions for authors

Refer to weaknesses.

5 comments
AuthorsOpenReview ↗

2KH1 — Response 1/4

We thank the reviewer for recognizing the motivation, well-structured search formulation, and cross-pipeline generalizability of our work. Below, we address your concerns point by point.

Q1. Generation-Aware Optimization Beyond Simple Coarsening and Segmentation Repair.

Q1.1 Generation-aware optimization versus simple coarsening.

The benefit over simple coarsening remains unclear. FreePart only does part merging, gives no initial/final part counts, and makes no comparison to simple merging at the same part count.

We apologize for not clearly distinguishing our method from simple part coarsening in the original submission. We have conducted additional experiments at matched part counts to clarify this distinction. Unlike simple coarsening, FreePart evaluates merge choices based on SC/OCC risks and functional constraints, rather than merely reducing the number of parts. In matched-final-count comparisons (K = 3, 4, 5; 10 eligible objects per K), FreePart consistently outperforms random adjacent merging, fixed contact-based merging, and SAM threshold adjustment in reconstruction quality. These results demonstrate that the improvements primarily arise from generation-aware merge decisions rather than part count reduction alone, as detailed in the table below.

R1.Reconstruction quality comparison at matched final part counts.

KMethodCD ↓F1-0.1 ↑F1-0.05 ↑
3Random adjacent11.2085.3863.85
3Fixed contact10.9487.2464.94
3SAM threshold12.4383.0259.67
3FreePart10.8387.6965.89
4Random adjacent11.8281.0562.95
4Fixed contact12.1980.5660.49
4SAM threshold14.8973.7547.62
4FreePart11.7782.0362.98
5Random adjacent14.2176.3353.41
5Fixed contact13.5577.9555.72
5SAM threshold15.7971.9846.39
5FreePart13.3678.8756.00

Q1.2 Generation-Aware Optimization Beyond Segmentation Repair.

The instruction to count the removal of an unnecessary internal boundary as an SC improvement further blurs the distinction between generation-aware guidance and segmentation repair.

We clarify that removing an unnecessary internal boundary does not necessarily improve 3D generation quality. To further illustrate this distinction, we have added the bread-basket and flower-pot examples in Fig. 2 of the revised paper. Both fine-grained and coarse-grained decompositions are semantically plausible, yet coarser decompositions can either improve or degrade reconstruction quality, demonstrating that semantic plausibility alone does not guarantee generation-friendliness. Therefore, rather than simply removing internal boundaries, FreePart evaluates candidate merges based on SC/OCC-related generation risks and functional consistency to identify generation-friendly decompositions.

To further distinguish generation-aware optimization from segmentation repair, we evaluate a subset of objects with negligible changes in Bbox IoU (|ΔBbox IoU| ≤ 0.001, normalized), examining whether reconstruction gains persist without substantial improvements in segmentation overlap.

R2.Reconstruction Performance with Nearly Unchanged Segmentation Overlap.

MethodBbox IoU ↑Part CD ↓Part F1-0.1 ↑Overall CD ↓Overall F1-0.1 ↑
Original OmniPart32.65625.86950.90611.97583.288
FreePart32.67023.92955.36311.76884.417

As shown in Table R2, our experiments provide complementary evidence for this distinction. SC/OCC-guided optimization outperforms logic-only merging, while reconstruction quality improves even when Bbox IoU remains nearly unchanged, suggesting that the gains extend beyond improved segmentation alignment.

Q1.3 SC/OCC Ablation with the Original Baseline.

Although the SC/OCC ablations show incremental benefits, unmodified OmniPart is missing from that same ablation set.

We apologize for omitting the unmodified OmniPart results from the original ablation study.We have now included this baseline for a more comprehensive comparison.Importantly, logic-only still performs constrained merging and is not equivalent to the original generator:

R3. SC/OCC Ablation with the Original OmniPart Baseline.

MethodSCOCCCD ↓F1-0.1 ↑F1-0.05 ↑
Original OmniPart——14.3480.0753.84
Logic only✗✗14.0880.9955.16
SC only✓✗13.7181.6857.01
OCC only✗✓13.7181.7256.89
SC + OCC✓✓13.4882.2657.86

Both risk indicators improve on logic-only merging, indicating an incremental benefit beyond coarsening.

AuthorsOpenReview ↗

2KH1 — Response 2/4

Q2. Clarification on Part Matching and Unmatched Predictions.

The part-level evaluation doesn’t penalise unmatched parts. CD, F-score and Bbox IoU are all averaged only over the matched pairs, so unmatched parts incur no penalty. This makes the comparison sensitive to the number of parts being reduced. Missing predictions are also omitted, which can alter the denominator of the evaluation between different methods. Results should be reported with unmatched-part penalties and with the same object set.

We thank the reviewer for raising this concern. We clarify that every generated part is optimally matched to a GT part, ensuring that no predicted parts are left unmatched or selectively excluded from evaluation. Specifically, our evaluation follows the established part-level evaluation protocol of OmniPart, using widely adopted CD, F-score, and Bbox IoU metrics for comparability with prior work. All methods are evaluated on the same test objects with identical, fixed GT part annotations, and generated parts are optimally matched to GT parts to accommodate differences in decomposition. Therefore, all components will be matched. Furthermore, we report overall-level CD and F-scores, which evaluate the complete reconstructed geometry independently of part matching and consistently demonstrate improvements. Together, these evaluations provide complementary evidence that FreePart's reconstruction gains are not merely a consequence of reduced part granularity.

Q3. Risk Assessment, OCC Calibration, and Cross-Modal Generalizability.

Q3.1 OCC Calibration Settings and Reliability.

The proxy is not sufficiently validated. The OCC calibration omits the generator, sample counts, uncertainty and size/complexity controls.

Our OCC analysis is based on Objaverse-XL, a large-scale dataset comprising over 10 million 3D assets collected from diverse sources . It provides a broad geometric foundation for OCC risk analysis, rather than relying on a single asset source or modeling pipeline. The detailed calibration settings are summarized below.

R4. Diversity and Characteristics of Objaverse-XL.

Generator / Asset CreationSample CountsUncertaintySize / Complexity
Manual modeling (e.g., Blender, Maya), photogrammetry (e.g., COLMAP, RealityCapture), and 3D scanning (e.g., Polycam)~10.2M 3D assets from GitHub, Thingiverse, Sketchfab, Polycam, and the Smithsonian InstitutionHeterogeneous sources and asset qualityDiverse object scales and geometric complexities; vertex, polygon, and edge counts available

Our OCC calibration further covers 1,913 objects, with the empirical OCC–reconstruction relationship reported in Q3.4 (Table R6).

Q3.2 Risk Prediction Accuracy.

Neither the accuracy of the risk prediction nor the relationship between the predicted risk reduction and the actual improvement in reconstruction is reported.

We apologize for not reporting the risk prediction accuracy in the original submission. To validate the effectiveness of our fine-tuned VLM, we compare its predictions against geometry-derived ground-truth labels from 5,000 parts across 2,049 test objects. The model achieves 92.08% OCC risk accuracy, 85.11% SC risk accuracy, and an AUC of 0.904 for OCC risk identification, demonstrating its effectiveness in identifying generation-related risks.

AuthorsOpenReview ↗

2KH1 — Response 3/4

Q3.3 Relationship Between Predicted Risk and Reconstruction Improvement.

Predicted generation risk exhibits a strong association with reconstruction difficulty, while risk-guided optimization consistently improves reconstruction quality.We conduct a risk-stratified reconstruction analysis to examine the relationship between SC risk and downstream generation quality. As shown in the tables, high-SC parts exhibit substantially higher baseline reconstruction errors than low- and medium-SC parts , supporting SC as an indicator of generation difficulty. FreePart consistently improves reconstruction quality across all three risk groups. Notably, the high-SC group achieves the largest improvement in Part F1-0.1 (36.45 → 44.85, +8.40 percentage points). These results demonstrate that SC-based risk assessment can identify challenging parts and provide effective guidance for generation-friendly decomposition optimization.

R5. Reconstruction Performance Across SC Risk Levels.

SC RiskMethodPart-level Bbox IoU ↑Part-level CD ↓Part-level [email protected] ↑Part-level [email protected] ↑Overall-level CD ↓Overall-level [email protected] ↑Overall-level [email protected] ↑
LowOmniPart23.8626.8549.0927.4114.8777.0849.90
LowOmniPart + FreePart26.9023.6355.1732.1913.9280.5154.92
MediumOmniPart23.3126.9349.2527.4314.6677.4052.29
MediumOmniPart + FreePart24.4024.4853.9531.2714.0079.8654.59
HighOmniPart20.9134.5136.4519.3216.5176.9560.84
HighOmniPart + FreePart24.7531.5944.8523.8416.1577.1461.29
AllOmniPart23.3127.3748.4426.9414.8077.3052.21
AllOmniPart + FreePart25.0724.6953.7231.0614.1379.8555.05

On the full reported OmniPart benchmark, Part CD changes 27.37 → 24.69, Overall CD 14.80 → 14.13, and Overall F1-0.1 77.30 → 79.85. These results support improvements across risk categories.

Q3.4 OCC Observation Protocols and Model Transferability.

The transfer to occlusion-robust models (e.g. SAM 3D) and the OCC observation protocols for mesh and scene inputs are not described.

For OCC calibration, we analyze five equal-frequency bins of GT object-level OCC, with 382–383 objects per bin, and compute 2,000 bootstrap resamples per bin. The bin (0.535, 0.633] (mean OCC 0.600) obtains the lowest mean overall CD (0.1233), motivating an empirical, setting-dependent reference.

R6. Relationship Between OCC and Reconstruction Quality.

GT OCC rangeObjectsMean OCCOverall CD ↓F1-0.1 ↑F1-0.05 ↑
[0, 0.535]3830.3440.14000.79400.5994
(0.535, 0.633]3820.6000.12330.82890.6313
(0.633, 0.708]3830.6700.12630.82700.6243
(0.708, 0.787]3820.7470.13440.82180.6001
(0.787, 0.983]3830.8530.14370.78970.5624

Regarding different input modalities, FreePart estimates risks from RGB images and part masks using the fine-tuned VLM for image inputs, while SC/OCC can be directly computed from available 3D geometry for mesh inputs under specified observation views. This allows the framework to accommodate different input modalities without modifying downstream generators. We also appreciate the reviewer's suggestion regarding occlusion-robust models such as SAM 3D and will consider extending and evaluating FreePart on such models in future work.

Reviewer_47v4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FreePart optimizes the input decomposition of existing part-based 3D generators through adjacent-part merges. A shared VLM, trained using surface-complexity and occlusion-risk labels, acts as Diagnoser, Strategist, and Evaluator. Experiments show improvements across image-, mesh-, and scene-conditioned generation without changing the downstream generators.

Critical strengths
  • Optimizing decomposition for a fixed generator is a useful intervention: semantic plausibility alone does not establish that parts can be generated reliably in isolation.
  • The same trained framework improves three different generation pipelines and both part- and whole-object metrics, supporting practical transfer beyond one backbone.
Critical weaknesses
  • Quality–granularity trade-off. Every accepted merge removes a separately controllable part. Functional-instance supervision and logic constraints are useful safeguards, and Bbox IoU improves, but the paper does not establish which editing capabilities remain. Final part counts and comparison at matched granularity would clarify what the reconstruction gains deliver.
  • Surface-complexity supervision depends on tessellation. Equation 4 normalizes weighted dihedral angles by total internal-edge length. Subdividing planar faces introduces zero-angle edges and can lower SC without changing the surface. This does not negate the reported gains, but leaves unclear how consistently the central risk label measures geometry across meshes and transferred generators.
  • The search architecture's added value is not fully isolated. Appendix C follows one accepted merge trajectory; rollback rejects a preview rather than revisiting an earlier accepted state. Table 5 supports warnings, adaptive proposals, and verification, including a fixed-rule baseline. It does not separate the three-role design from a learned greedy merger using comparable risk information and model-call budget.
Questions for authors
  • Clarification: What is the final part-count distribution, and do gains persist against a simpler merger at matched part count and inference budget?
  • Clarification: Is mesh preprocessing standardized enough to stabilize SC under retriangulation, and how is label robustness assessed?
4 comments
AuthorsOpenReview ↗

47v4 — Response 1/3

We thank the reviewer for recognizing the motivation, effectiveness, and generalizability of our work. Below, we address your concerns point by point.

Q1. Quality–granularity trade-off and preserved editability.

Q1.1 Preserved editability.

Quality–granularity trade-off. Every accepted merge removes a separately controllable part. Functional-instance supervision and logic constraints are useful safeguards, and Bbox IoU improves, but the paper does not establish which editing capabilities remain.

We agree that semantic or functional validity alone does not guarantee independent part editability. FreePart supports user-specified part protection: candidate merges involving a designated target are filtered before the Strategist acts, while other parts remain optimizable. The protected target stays independent throughout search. In our experiment, target preservation improves from 40% (FreePart W/O protection ) to 100% (FreePart).

Although restricting merges may slightly compromise reconstruction quality compared with unconstrained FreePart, the protected variant still improves over the initial decomposition.The target-preservation results are summarized below:

R1. Trade-off Between Part Preservation and Reconstruction Quality.

MethodIndependent targets ↑Average final partsCD ↓F1-0.1 ↑F1-0.05 ↑
Initial decomposition100%5.812.2284.1153.75
FreePart W/O protection40%3.210.8387.0764.20
FreePart100%4.011.3286.3459.59

Q1.2 Part-count distribution.

Final part counts and comparison at matched granularity would clarify what the reconstruction gains deliver.

We apologize for not clearly reporting the final part counts in the original submission. We clarify this below. Across PartObjaverse-Tiny objects using OmniPart, FreePart reduces the average number of parts from 7.48 to 4.69 (median: 7 → 4). The part-count distribution shifts toward coarser decompositions, with the number of objects containing 1–3 parts increasing from 12 to 93, while those containing 7 or more parts decrease from 124 to 56. These results characterize the granularity changes introduced by FreePart.

Q1.3 Matched-granularity comparison.

Our reconstruction improvements primarily stem from the effectiveness of our merge strategy, rather than merely reducing the number of parts. To disentangle the benefits of generation-aware decomposition optimization from simple part coarsening, we conduct controlled comparisons with random adjacent merging, fixed contact-based merging, and SAM threshold adjustment at matched final part counts K ∈ {3, 4, 5}, using 10 eligible objects per K. All methods are evaluated on the same objects within each K group and produce decompositions with identical part counts, thereby isolating the influence of decomposition choices rather than granularity alone.

R2.Reconstruction quality comparison at matched final part counts.

KMethodCD ↓F1-0.1 ↑F1-0.05 ↑
3Random adjacent11.2085.3863.85
3Fixed contact10.9487.2464.94
3SAM threshold12.4383.0259.67
3FreePart10.8387.6965.89
4Random adjacent11.8281.0562.95
4Fixed contact12.1980.5660.49
4SAM threshold14.8973.7547.62
4FreePart11.7782.0362.98
5Random adjacent14.2176.3353.41
5Fixed contact13.5577.9555.72
5SAM threshold15.7971.9846.39
5FreePart13.3678.8756.00

As shown in the results, FreePart consistently achieves the best CD and F-scores across all three granularity settings, demonstrating that its improvements cannot be explained solely by reducing the number of parts. These findings highlight the importance of generation-aware merge selection in identifying decompositions better suited to downstream 3D generation. The part-count distribution and matched-granularity experiments are conducted on different evaluation subsets.

AuthorsOpenReview ↗

47v4 — Response 2/3

Q2. SC tessellation sensitivity.

Surface-complexity supervision depends on tessellation. Equation 4 normalizes weighted dihedral angles by total internal-edge length. Subdividing planar faces introduces zero-angle edges and can lower SC without changing the surface. This does not negate the reported gains, but leaves unclear how consistently the central risk label measures geometry across meshes and transferred generators.

We acknowledge that Eq. (4) is not strictly invariant to mesh subdivision. However, SC serves as a relative risk proxy under consistent preprocessing, rather than an intrinsic mesh-independent invariant. As shown in Table R3, SC risk labels exhibit reasonable consistency under subdivision, while the fine-tuned VLM accurately predicts geometry-derived SC risks.

R3.SC risk-label consistency under subdivision and VLM prediction accuracy.

SC Risk-Label Agreement (1 Round)SC Risk-Label Agreement (2 Rounds)VLM SC Prediction Accuracy
82.8%75.0%85.11%

Importantly, this sensitivity does not invalidate the relative comparisons required by FreePart. Our empirical analysis establishes a generally monotonic relationship between SC and reconstruction difficulty within a fixed generation setting. During decomposition optimization, candidate merges are assessed for the same object under a consistent geometric representation and downstream generator, where the underlying mesh tessellation remains unchanged. Thus, FreePart relies on relative changes in SC-related risk before and after merging, rather than comparing absolute SC values across differently tessellated meshes. This substantially reduces the influence of tessellation differences on the decisions made within a given search process.

Q3. Three-role search versus risk-greedy merging.

The search architecture's added value is not fully isolated. Appendix C follows one accepted merge trajectory; rollback rejects a preview rather than revisiting an earlier accepted state. Table 5 supports warnings, adaptive proposals, and verification, including a fixed-rule baseline. It does not separate the three-role design from a learned greedy merger using comparable risk information and model-call budget.

We agree that the component ablations alone do not isolate the benefit of the decision organization. We therefore introduce single-path greedy and risk-greedy baselines, using identical initial decompositions, the same VLM checkpoint and SC/OCC signals, and a shared maximum budget of six VLM calls per object.

R4. Reconstruction Quality and Search Efficiency Compared with Greedy Baselines.

MethodCalls / obj ↓Search time ↓Overall CD ↓F1-0.1 ↑F1-0.05 ↑
Single-path greedy1.3143.2 s0.122170.844930.58305
Risk-greedy5.7319.7 s0.107770.880180.63897
FreePart3.9260.6 s0.105500.885680.64877

As shown in Table R4, single-path greedy achieves lower computational cost but substantially compromises reconstruction quality. In contrast, FreePart achieves the best reconstruction performance across all metrics, while requiring 31.6% fewer VLM calls and 18.5% shorter search time than risk-greedy. These results demonstrate that FreePart achieves a better balance between reconstruction quality and search efficiency.

We additionally evaluate single-agent iterative and direct-grouping alternatives to distinguish iterative decisions from immediate final-grouping predictions. On the separate exploratory architecture subset, their overall-level results are:

R5. Comparison of Alternative Decision Organizations.

StrategyOverall CD ↓F1-0.1 ↑F1-0.05 ↑
FreePart0.1316500.8142590.574048
Risk-greedy0.1320640.8135400.572634
Single-agent iterative0.1326110.8142360.571939
Single-agent direct0.1323810.8088480.555346
Three-role direct0.1317740.8064700.570741

These comparisons complement, rather than replace, the Table 5 capability ablations. Finally, DAG search describes the reachable state space, while the implemented controller follows a sequential trajectory of accepted merges. Rollback rejects an uncommitted preview at the current state; it does not reverse an earlier accepted merge.

AuthorsOpenReview ↗

47v4 — Response 3/3

Q4. Part-Count Distribution and Matched-Budget Comparison.

Clarification: What is the final part-count distribution, and do gains persist against a simpler merger at matched part count and inference budget?

As detailed in Q1.2–Q1.3 and Q3, FreePart improves reconstruction quality over simple merging at matched final part counts (Table R2). Under the same maximum VLM-call budget, comparisons with single-path greedy and risk-greedy (Table R4) further demonstrate that FreePart achieves the best reconstruction quality while maintaining a favorable balance between performance and search efficiency.

Q5. SC Preprocessing and Label Robustness.

Clarification: Is mesh preprocessing standardized enough to stabilize SC under retriangulation, and how is label robustness assessed?

As clarified in Q2, SC is evaluated under consistent mesh preprocessing rather than assumed to be strictly invariant to tessellation. Our additional subdivision experiments (Table R3) demonstrate reasonable SC risk-label consistency, supporting its use as a relative risk indicator within a fixed preprocessing pipeline.

We sincerely appreciate your constructive feedback. We have provided additional experiments and clarifications on the quality–granularity trade-off, SC robustness, and search strategy. We greatly value your time and look forward to any further feedback before the discussion period ends, so that we can address remaining concerns and further improve our work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers

We thank all reviewers for their constructive feedback and for recognizing FreePart's motivation, effectiveness, and generalizability. In response to your comments, we have provided the following additional evidence and clarifications:

  1. Generation-aware optimization and editability. Matched-granularity comparisons demonstrate improvements beyond simple coarsening, while additional evaluations validate FreePart's existing target-part protection mechanism.

  2. Risk assessment and calibration. We provide further validation of SC/OCC risk prediction, SC/OCC robustness, and the relationship between generation risks and reconstruction quality.

  3. Search strategy and decision organization. Comparisons with greedy, single-agent, and direct-grouping alternatives further support the effectiveness of FreePart's multi-agent search framework.

  4. Computational efficiency and robustness. We report inference costs, VLM call counts, search efficiency, and multi-seed evaluations to assess practical performance and stability.

Our individual responses provide detailed results and address the remaining methodological and evaluation concerns.

The corresponding revisions in the main text and appendices are highlighted in blue.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.