acceptodds
Under review as a conference paper at ICLR 2027

Partial3D: Controllable 3D Asset Generation from Multiple Partial Images

Abstract

While 3D generative models have achieved unprecedented visual quality, controlling these systems remains an underexplored frontier compared to 2D generation, primarily due to the significant data bottleneck associated with high-quality 3D assets. Consequently, existing approaches for 3D control often rely on indirect pipelines that leverage 2D control methods and lift the results to 3D, a process that frequently leads to error accumulation and lacks an efficient feed-forward mechanism. In this paper, we propose Partial3D, a native 3D generative model that utilizes partial images, optional 3D layouts, and text for precise and controllable 3D synthesis, enabling native 3D control without relying on intermediate 2D pipelines. Our main insight is to treat partial images as separable visual conditions, allowing for compositional control over the generated 3D content, analogous to how text prompts are used in 2D generation. To address the inherent ambiguity of composing partial images, Partial3D incorporates optional 3D layout guidance through a Spatial Prompt Modulator (SPM) and text guidance through a training-free velocity-fusion strategy. By treating partial images as separable conditioning signals, while incorporating 3D layouts and text prompts, Partial3D enables native 3D editing through image replacement and supports unified control across image, text, and 3D layout modalities. Extensive experiments demonstrate that Partial3D generates 3D content under precise multimodal control and successfully enables downstream applications such as high-fidelity 3D editing.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_buFoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses the controllable 3D asset generation problem. It proposes a new framework Partial3D, which composes 3D assets from multiple partial images by adapting TRELLIS2 with spatial prompt modulation, cross-attention masking, and optional text-guided velocity fusion. The framework also supports component editing. Experiments and ablation studies demonstrate advantages over native 3D and indirect 2D-to-3D baselines.

Critical strengths
  1. Composing a complete 3D asset directly from partial images without a holistic reference is a useful and well-motivated formulation.

  2. The lightweight adaptation is supported by extensive comparisons and ablations covering spatial control, editing, and preservation of the pretrained generation prior.

  3. The writing of this paper is good and easy to follow.

Critical weaknesses
  1. Uncharacterized behavior under conflicting controls. Overlapping boxes allow the same voxels to attend to multiple image conditions, but yet the paper does not characterize how incompatible conditions in these regions are resolved. Similar uncertainty arises when text contradicts with the requested image/layout composition.

  2. Incomplete backbone experiment. The paper only compares against TRELLIS.2 velocity averaging result, omitting the commonly used TRELLIS multi-image strategy of increasing the sampling steps and using a different conditioning image at each denoising step.

  3. Limited spatial isolation. Cross-attention masking is only applied during sparse-structure generation, leaving final geometry and texture susceptible to cross-part interference, as suggested by the flowerpot in Fig. 4, whose texture does not faithfully match with the input image.

Questions for authors
  1. How transferable is the proposed framework to other large pretrained 3D generative models, such as Hunyuan3D and Direct3D, given its reliance on TRELLIS2's sparse-structure pipeline?

  2. Given training on complementary parts of conventional 3D assets, will generation quality degrade substantially when the input images depict semantically distant components at inference time?

3 comments
AuthorsOpenReview ↗

Response to Reviewer buFo

We thank the reviewers for recognizing our use of partial images as "composable visual conditions" (uGq2), our "intuitive and useful control interface" (xgzP), and our "useful and well-motivated formulation" (buFo). We will carefully revise the manuscript in response to their valuable feedback.

  1. Conflicting controls.

Overlapping boxes. Our 3D boxes act as coarse spatial prompts that often intersect or nest (e.g., Fig. 5a–b). Since cross-attention masking only blocks voxels outside each box (Sec. 3.2), overlapping voxels attend jointly to multiple conditions via softmax, while SPM modulation and 3D self-attention maintain smooth transitions. Modest box errors are well tolerated on the 36 layout cases (Table 7):

SettingMasked DINO ↑DINO drop ↓BBox IoU ↑
Original boxes (Table 7)0.4480.0000.682
Position jitter (±5%)0.4390.0090.653
Scale jitter (±5%)0.4410.0070.657

If incompatible parts are assigned nearly identical boxes, shared voxels blend their features or favor the more contextually consistent part.

Text vs. image/layout. Text fusion acts only in Stage 1 with cosine decay (1 → 0), whereas Stages 2–3 condition solely on images (Fig. 3, Sec. 3.3). Thus, images strictly govern fine geometry and appearance, while severe text-layout contradiction yields a compromised coarse topology (mitigable by lowering w_start).

  1. Step-wise multi-image baseline.

Our TRELLIS2 baseline averages per-image velocity predictions at every sampling step (Supp. B.4). We will add a step-wise variant that conditions each step on one image, cycling through the references.

MethodCLIP ↑Masked CLIP ↑Masked LPIPS ↓Masked DINO ↑
TRELLIS2 (averaging, Table 6)0.1940.5030.6910.383
TRELLIS2 (step-wise)0.1890.4920.6980.376
Ours (Table 6)0.2380.5700.6400.443
  1. Spatial isolation in later stages.

Our layouts provide coarse spatial guidance for the overall shape, which is established during sparse-structure generation. We therefore apply spatial masking at this stage without imposing coarse box boundaries on subsequent shape and texture refinement. Following Fuse3D, we visualize cross-attention in the shape and texture DiTs (Fig. 12; Supp. C.4). In the shown examples, each partial image's attention largely concentrates on its corresponding region, with limited visible cross-part interference. We acknowledge the fine-detail fidelity remains limited by the representational and generative capacity of the TRELLIS2 backbone.

  1. Transferability to other backbones.

Our partial-image conditions are injected through cross-attention, and SPM modulates image features; neither operation intrinsically requires TRELLIS2's voxel representation. The same conditioning idea can therefore be adapted to vector-set models such as Hunyuan3D and triplane-based Direct3D through their image-conditioning attention interfaces. In cross-attention, 3D latent tokens act as queries and the modulated partial-image tokens as keys and values; Hunyuan3D 2.0 instead uses joint attention between latent and condition tokens. This requires adapting the conditioning interface and fine-tuning, while spatial masks need a representation-specific correspondence to 3D regions.

  1. Semantically distant compositions.

As described in Supp. B.2, 70 of our 100 evaluation cases combine independently generated parts from different source objects, rather than complementary parts of one object as in training. Fig. 7 includes a toast combined with cartoon facial features and a chest assembled from separate component references; Fig. 1 shows a violin-beetle hybrid. In particular, the toast-face and violin-beetle examples demonstrate coherent compositions from semantically distant inputs, beyond the same-object part completion used during training.

AuthorsOpenReview ↗

Summary of Manuscript Revisions (for Reviewer buFo)

We have uploaded the revised manuscript (with all updates highlighted in blue) incorporating all experiments, visual examples, and analyses addressing Reviewer buFo's questions: (1) on Pages 26–27 (Appendix C.13, Table 14), we added the ±5% bounding-box position/scale perturbation study on the 36 layout cases and analyzed model behavior under overlapping 3D boxes and conflicting text vs. image/layout controls; (2) on Page 22 (Appendix C.3) and Page 26 (Appendix C.12, Table 13), we added the step-wise reference-cycling TRELLIS2 multi-image baseline on the 30 image-only cases; (3) on Page 19 (Appendix B.7, Fig. 12) and Page 22 (Appendix C.4), we clarified that Fig. 12 visualizes cross-attention responses in the shape and texture DiTs, explained why spatial cross-attention masking is applied only during Stage 1 sparse-structure generation, and noted TRELLIS2's backbone capacity limits on fine surface detail; (4) on Pages 27–28 (Appendix D), we added an architectural transferability discussion covering cross-attention (Direct3D) and joint-attention (Hunyuan3D 2.0) 3D backbones and representation-specific spatial masking; and (5) on Page 15 (Appendix A, Fig. 7), Page 21 (Appendix C.1, Fig. 14), and Page 27 (Appendix C.13), we expanded our qualitative results with additional semantically distant, cross-style, and controllable editing cases (including a classical marble bust with modern headphones, a realistic pineapple with mechanical spider legs, and a cartoon dog with a realistic hotdog).

Reviewer_buFoOpenReview ↗

Your rebuttal and revisions

Thank you for providing your responses to my comments and questions. I am happy with your responses and do not have further questions. I will keep my original rating of Weak Acceptance.

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

Partial3D proposes a framework for controllable 3D asset generation from multiple independently cropped partial images, with optional text and 3D bounding-box conditions. Built upon TRELLIS2, the method combines concatenated image features, a Spatial Prompt Modulator, spatial cross-attention masking, and scheduled fusion of image- and text-conditioned velocity fields. It also supports component-level editing through partial-image replacement and source-latent preservation. The paper reports improvements in visual consistency, spatial controllability, and compositional coherence.

Critical strengths
  1. An intuitive and useful control interface. Composing a 3D asset from separate visual references offers a practical way to specify component appearance more precisely than text alone. Supporting component replacement, addition, and removal makes this formulation relevant to interactive asset creation.
  2. A well-motivated spatial conditioning design. The combination of layout-dependent feature modulation and spatial cross-attention masking provides a reasonable mechanism for associating partial-image conditions with intended 3D regions. The ablations support the effectiveness of both components.
  3. Encouraging empirical results. The paper includes quantitative comparisons, user studies, matched-condition evaluations, geometric evaluation on held-out assets, and editing-preservation experiments. These provide useful evidence beyond selected qualitative examples.
Critical weaknesses

W1. The acquisition of 3D layouts and its implications for practical scalability are insufficiently explored. I find the most convincing technical contribution to be the effective use of supplied 3D layouts for component-conditioned generation. However, obtaining these layouts remains an external prerequisite whose practical cost is not sufficiently characterized. During training, boxes are derived from segmented source assets, whereas inference assumes user-specified spatial regions. The construction and annotation procedure for layout conditions in the synthetic evaluation cases also needs clarification. The paper does support layout-free generation and provides image-only comparisons. Nevertheless, these experiments do not establish how accurately the model infers component placement without explicit spatial guidance. In particular, the geometric results in Table 8 are not separated by conditioning setting, making it difficult to assess layout-free structural accuracy. The authors should distinguish performance under user-provided layouts from performance when spatial relationships must be inferred. Reporting geometry and component-placement metrics separately for these settings would be informative. An automatic or assisted layout-estimation baseline would further clarify the usability of the pipeline, together with an analysis of sensitivity to inaccurate boxes. Automatic layout estimation need not become the paper’s main contribution, but its absence and the associated user burden should be discussed more thoroughly.

W2. Stronger task-matched baselines are needed to establish the advantage of native partial-image composition. TRELLIS2 is a useful backbone reference, but directly averaging predictions conditioned on separate partial images is a limited composition baseline. It does not necessarily represent the strongest approach available without task-specific 3D fine-tuning. I acknowledge that the paper already includes MUSE/XVerse followed by TRELLIS2, as well as qualitative comparisons with a commercial image-generation model. My concern is therefore about the strength and coverage of these comparisons. A particularly relevant baseline would use a strong image-composition or editing model to integrate the same partial-image references into a coherent image, followed by TRELLIS2 reconstruction, without additional task-specific training. Quantitative evaluation under matched reference images, text, layout information where applicable, and generation budgets would better establish the benefit of the proposed native pipeline. Joint text-image-conditioned 3D generation methods, such as TIGON, should also be considered when evaluating the text-guidance component. Since such methods may not directly support multiple partial images, any adaptation should be documented explicitly. This would help distinguish improvements from partial-image conditioning from those attributable to multimodal guidance more generally.

W3. The relationship to pixel-aligned 3D generation and the portability of the proposed controls remain unclear. Preserving the visual identity of input components is a central motivation of Partial3D. Recent approaches such as Pixal3D explicitly establish pixel-to-3D correspondences through image-feature back-projection, making them relevant to the paper’s fidelity claims. The authors should discuss whether Partial3D’s compositional conditioning and spatial modulation can complement such aligned representations, and which aspects depend specifically on TRELLIS2’s attention-based conditioning. This is not a straightforward substitution: independently cropped component images do not necessarily share a consistent camera or global coordinate system. Understanding how the method could accommodate explicit alignment would therefore provide useful insight into its broader applicability. A comparison using a stronger aligned image-to-3D model within the composition-first baseline would be informative. A complete port of Partial3D to another backbone could remain future work, but the current manuscript should more clearly explain the compatibility challenges and the expected scope of its approach.

Questions for authors

How sensitive is generation to errors in box position, scale, and overlap?

How are the target 3D boxes produced for the synthetic evaluation cases, and how much manual intervention is required?

What prevents the proposed conditioning strategy from being combined with pixel-aligned 3D generators?

3 comments
AuthorsOpenReview ↗

Response to Reviewer xgzP

We thank the reviewers for recognizing our use of partial images as "composable visual conditions" (uGq2), our "intuitive and useful control interface" (xgzP), and our "useful and well-motivated formulation" (buFo). We will carefully revise the manuscript in response to their valuable feedback.

  1. Layout acquisition and user effort.

In our 100-case benchmark (Supp. B.2), GPT-4o generates both partial images and target 3D boxes for the layout-conditioned cases among the 70 synthetic examples (to be clarified near line 844), while the 30 held-out cases derive boxes from segmented source assets as in training (Supp. B.1).

We agree manual 3D box specification adds user effort; thus, boxes remain optional, complemented by text guidance for layout-free global control (Sec. 3.3) and VLM agents (e.g., GPT-4o) for automated layout assistance.

  1. Evaluation under different conditions.

While Tables 6–7 (Supp. C.3) use matched conditions, Table 8 mixes conditioning settings. To isolate layout-free structural accuracy, we re-evaluate the 30 held-out cases using partial images only (no text/boxes) under the Supp. C.5 geometric protocol:

MetricTRELLIS2Fuse3DMUSEXVerseOurs
CD (×10⁶) ↓239.84181.6795.94153.2719.39
F1 (τ = 10⁻⁶) ↑0.2120.3280.5750.1130.641
  1. Box sensitivity.

During training, we drop box conditions with 50% probability and use a learned null prompt (Sec. 4; Supp. A.1). Although boxes affect cross-attention, self-attention still exchanges information among 3D tokens. We argue that this interaction can help maintain coherent generation under modest box errors. To examine this, we separately perturb box centers and sizes by up to 5% on the 36 image+text+layout cases from Table 7, keeping images, text, and sampling seeds fixed. We report Masked DINO and its drop from the unperturbed setting, together with BBox IoU measured against the actual input boxes.

SettingMasked DINO ↑DINO drop ↓BBox IoU ↑
Original boxes (Table 7)0.4480.0000.682
Position jitter (±5%)0.4390.0090.653
Scale jitter (±5%)0.4410.0070.657

Note that part AABBs naturally overlap or nest in training and inference (Fig. 5a–b), and position/scale jitter also alters inter-box overlap with minimal degradation (see also R3-buFo.1 on conflicting overlaps).

  1. Additional baseline comparisons.

Under comparable open-source budgets, Partial3D already outperforms strong composition baselines—XVerse (FLUX.1-dev) and MUSE (SDXL)—lifted via TRELLIS2 under matched conditions (Tables 1, 6–7). While commercial models like GPT-Image-2 (Supp. C.6) achieve strong 2D visual quality via vastly larger proprietary scale and compute, we target the intrinsic bottlenecks of 2D-to-3D lifting rather than raw 2D capacity: (i) occluded parts remain underconstrained in single 2D views, causing incomplete/hollow 3D geometry (Fig. 10); (ii) multi-view synthesis suffers from view inconsistency (Figs. 2, 11); and (iii) indirect 2D composition lacks precise 3D box control—on the 36 layout cases (Table 7), GPT-Image-2 + TRELLIS2 reaches 0.633 BBox IoU vs. 0.682 for Ours.

We also evaluate TIGON on the 34 image+text cases (Table 7) with identical inputs. Since TIGON assumes a single image like TRELLIS, we average its joint text-image velocity predictions across references at each step, which struggles to compose parts:

MethodCLIP ↑Masked CLIP ↑Masked LPIPS ↓Masked DINO ↑
TIGON (multi-image averaging)0.1960.4930.6830.391
Ours (Table 7, image+text)0.2510.5730.6370.441
  1. Portability and pixel alignment.

Our partial-image conditioning and SPM do not intrinsically depend on TRELLIS2's voxel representation. The modulated partial-image tokens are injected through cross-attention, and the same conditioning idea can be adapted to other attention-conditioned backbones, such as Hunyuan3D and Direct3D.

Direct transfer to pixel-aligned generators is more challenging: independently cropped parts lack accurate camera parameters in a shared target frame. Our boxes provide coarse spatial constraints but do not fully determine this mapping. A promising direction is to use them for approximate feature projection and train the model to accommodate alignment errors. This could use separate references for parts occluded in a composed image (Fig. 10), at the cost of harder learning from uncertain correspondences.

Nevertheless, this added learning difficulty need not be prohibitive. TRELLIS's VAE is trained on voxel features formed by projecting and averaging DINO features across views without visibility filtering, and its authors report that this aggregation works well despite including occluded views (TRELLIS GitHub issue #93). This provides a precedent for learning with imperfect projected features and motivates exploring bbox-guided projection.

AuthorsOpenReview ↗

Summary of Manuscript Revisions (for Reviewer xgzP)

We have uploaded the revised manuscript (with all updates highlighted in blue) incorporating the clarifications, quantitative evaluations, and discussions addressing Reviewer xgzP's comments: (1) on Page 16 (Appendix B.2), we clarified target 3D bounding-box construction for both the 30 held-out source-segmented cases and the layout-conditioned subset of the 70 GPT-4o-synthesized cases; (2) on Page 22 (Appendix C.5) and Pages 25–26 (Appendix C.11, Table 12), we clarified the mixed conditioning in Table 8 and added the pure partial-image-only (layout- and text-free) 3D geometric evaluation (CD and F1) on the same 30 held-out assets; (3) on Pages 26–27 (Appendix C.13, Table 14), we added the ±5% bounding-box position and scale perturbation study on the 36 layout-conditioned cases together with an analysis of inter-box overlap; (4) on Page 23 (Appendix C.6) and Page 22 / Page 26 (Appendix C.3; Appendix C.12, Table 13), we added the 3D BBox IoU comparison against GPT-Image-2 + TRELLIS2 (0.633 vs. 0.682) and the matched-condition evaluation of multi-image velocity-averaged TIGON on the 34 image+text cases; and (5) on Pages 27–28 (Appendix D), we added a dedicated discussion on portability to attention-conditioned backbones (Hunyuan3D 2.0, Direct3D) and compatibility challenges and coarse-box projection directions for pixel-aligned 3D generators (Pixal3D).

Thank you to the authors for their responses. Most of my questions have been thoroughly addressed.

I am inclined to keep my current score. That said, I believe that approaches based on Pixal3D and SAM 3D for ensuring consistent orientations could see wider use in the future, and I encourage the authors to give this direction further consideration.

Reviewer_uGq2OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Partial3D, a feed-forward native 3D generative framework that composes a coherent 3D asset from multiple independently cropped partial images. It concatenates DINO features from partial images into a unified visual-token sequence, uses a Spatial Prompt Modulator and spatial cross-attention masking for optional 3D layout control, and applies training-free velocity fusion for optional text guidance. Built on TRELLIS2 with parameter-efficient fine-tuning, it supports native 3D editing by adding, removing, or replacing partial-image conditions. Experiments show better part-level consistency and spatial controllability than 2D-to-3D and native 3D baselines.

Critical strengths

Partial3D treat independently cropped partial images as composable visual conditions for native 3D generation, rather than relying on a holistic image or an indirect 2D-to-3D lifting pipeline. This design preserves the pretrained image-to-3D prior while enabling component-level composition and native editing through adding, removing, or replacing partial-image conditions.

Critical weaknesses

W1:Most key modules are direct adaptations of existing techniques.

Token concatenation and 3D layout encoding are not new methods; the paper mainly adapts TRELLIS2 to multi-image input and layout control, which makes the contribution largely engineering-oriented. The approach is also very similar in spirit to Fuse3D.

W2:There is a gap between training and inference.

During training, each sample mainly consists of two complementary partial images from the same object and the same view, whereas inference demonstrates three- or four-part inputs and cross-object compositions. The method lacks analysis of joint global inference from partial images with independent sources, different styles, and different scales. Moreover, the qualitative results are mostly cartoon-like, raising concerns about generalization.

Questions for authors

see weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer uGq2

We thank the reviewers for recognizing our use of partial images as "composable visual conditions" (uGq2), our "intuitive and useful control interface" (xgzP), and our "useful and well-motivated formulation" (buFo). We will carefully revise the manuscript in response to their valuable feedback.

  1. Motivation and novelty.

As stated in the Introduction (Fig. 2), 2D editing offers intuitive control through text replacement. We seek a similar interface for native 3D editing, where users can replace individual visual conditions to change object components. Many 3D editors still perform edits in 2D before lifting them to 3D, while most image-to-3D methods use a holistic reference. Although Fuse3D supports multiple regional conditions, it encodes each full image and uses this context to align selected image regions with voxels initialized from a global reference. Independently cropped parts remove the surrounding context that supports this 2D-3D correspondence, limiting its reliability.

Motivated by this, we learn to synthesize a complete 3D asset from independently cropped partial images without a holistic reference, enabling native editing via image replacement. With cross-attention tuning and full-image regularization, single-image results remain close to TRELLIS2 (Fig. 6). On image-only cases, Masked DINO is 0.443 vs. 0.394/0.383 for Fuse3D/TRELLIS2 (Table 6; Fig. 4 shows qualitative comparisons). Also, this partial-image conditioning can extend in principle to vector-set backbones with cross-attention.

  1. Generalization.

App. C.10 already evaluates 2/3/4 inputs despite two-part training (Fig. 18; Table 11: Masked DINO 0.460/0.443/0.432). Of 100 test cases, 70 combine independently generated parts from different source objects (App. B.2). Independent crop normalization removes original positions and relative scales during training (App. B.1). Table 9/Fig. 15 also test occlusion and synthetic viewpoint changes. Existing cross-style examples combine a realistic sunflower with a 2D cartoon vase (Fig. 4), and a realistic toast slice with a cartoon face (Fig. 7). Our model inherits TRELLIS2's generative prior and its limitations. We will include more cross-style compositions in the revised paper.

AuthorsOpenReview ↗

Summary of Manuscript Revisions (for Reviewer uGq2)

We have uploaded the revised manuscript (with all updates highlighted in blue) to incorporate the new examples, experiments, and discussions addressing Reviewer uGq2's feedback: (1) on Page 15 (Appendix A, Fig. 7) and Page 21 (Appendix C.1, Fig. 14), we expanded our qualitative galleries with additional cross-style and cross-object 3D compositions and controllable editing cases—such as combining a classical marble bust with modern headphones, a realistic pineapple with mechanical spider legs, and a 2D cartoon dog with a realistic hotdog—and summarized cross-style and semantically distant synthesis on Page 27 (Appendix C.13); (2) on Pages 25–26 (Appendix C.11, Table 12; Appendix C.12, Table 13), we added layout-free partial-image-only 3D geometric comparisons (CD / F1) and a step-wise TRELLIS2 multi-image baseline, further validating our learned partial-image composition over holistic and training-free multi-image baselines; and (3) on Pages 27–28 (Appendix D), we added a discussion on adapting our partial-image conditioning and SPM beyond TRELLIS2 to vector-set (Hunyuan3D 2.0) and triplane (Direct3D) attention-conditioned backbones.

Thank you for the clarification. I have decided to maintain my score for two reasons:

  1. Motivation and Novelty The authors distinguish their method from Fuse3D by its ability to encode and compose part images independently. However, it takes pre-separated part images and additional 3D layout guidance. In my view, these inputs provide more information and make the task easier than extracting parts from an image, as Fuse3D does. CAST3D can compose parts from text without a supplied layout. Although the Introduction claims to eliminate a separate integration step, Partial3D still encodes parts independently and combines them under layout guidance. The only difference is that the composition takes place in latent space. This difference alone does not establish which method is more effective. In my view, the main differences between Partial3D and CAST3D are the use of TRELLIS.2 and 3D layout conditioning. The two-stage design of CAST3D that the authors highlight merely reflects the difference between the TRELLIS.1 and TRELLIS.2 backbones.

  2. Generalization Despite the authors’ explanation, the examples still consist mainly of high-resolution cartoon-style assets. The lower-fidelity, relatively real examples (bread and sculpture) in Figure 7 even suggest limited generalization.

Thank you for your follow-up and for articulating your remaining concerns. We respectfully clarify three factual points for the record:

  1. Optional layout, task difficulty & editability vs. Fuse3D: 3D layout is strictly optional—Tables 6 & 12, Figs. 1 & 4, and Fig. 7 (rows 1–3) use no layout at all. Unlike Fuse3D, which initializes global structure from a holistic reference, Partial3D removes the holistic image and normalizes away original scales/positions (App. B.1), making global 3D inference strictly harder; on the exact same layout-free parts, Fuse3D degrades sharply (Table 12: CD 181.67 vs. 19.39, F1 0.328 vs. 0.641; Table 6: Masked DINO 0.394 vs. 0.443). Crucially, this separable-part formulation directly enables flexible native 3D editing via part replacement, addition, or removal (Figs. 1, 14; Table 5), whereas Fuse3D's reliance on a holistic image couples components to the global reference and hinders component-level editing.

  2. Distinction from CAST3D: Respectfully, the two-stage pipeline of CAST3D does not merely reflect a difference between TRELLIS.1 and TRELLIS.2 (both share a native feed-forward SLAT paradigm). Rather, CAST3D is a training-free pipeline that explicitly infers a segmented 3D layout and then generates and integrates components separately with intermediate 3D segmentation/refinement. In contrast, Partial3D learns joint multi-part conditioning in a single feed-forward pass—working without any supplied layout (or optionally with text/boxes)—which is an algorithmic paradigm difference rather than a backbone upgrade.

  3. Generalization & visual fidelity: Because we freeze 81.7% of the pretrained TRELLIS2 backbone and its VAE (Table 4; App. D) to preserve its generative prior (Fig. 6), surface rendering style naturally reflects TRELLIS2/Objaverse-XL rather than a limitation of our compositional conditioning.

We appreciate your time and feedback throughout the review process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.