acceptodds
Under review as a conference paper at ICLR 2027

ScaffoldTree: Scaffold-to-Subtree Diffusion for 3D Tree Generation

Abstract

We present ScaffoldTree, a species-conditioned diffusion framework for structured 3D tree generation. ScaffoldTree decomposes a tree according to the structural roles of its branches: the trunk and major branches form a global scaffold, while fine-scale branches are represented as locally attached residual subtrees. Global Scaffold Diffusion first generates the scaffold by jointly modeling branch curves, attachment locations, and parent relationships. The Structure-Conditioned Residual Allocator then predicts the total number of residual branches from the generated scaffold and distributes them across scaffold anchors. Conditioned on the scaffold and allocation, Local Residual Diffusion generates the corresponding residual subtrees, which are hierarchically transformed according to their predicted parent–child relations and merged with the scaffold to form the complete tree. This decomposition separates global tree organization from local branching detail while maintaining their structural dependence. Experiments demonstrate improved structural and local-curve distribution matching and reference coverage compared with the evaluated baselines.

Primary area:
generative models
Keywords:
3D Tree Generation, Diffusion Models, Hierarchical Tree 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_BwVrOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work introduces a text-conditioned for a target tree species, 3D tree generation method using a diffusion method. Especially, the generation method consists of three multiple stages, from CLIP text Encoder, Scaffold Tree representation, and residual allocation to get the final output shape. It uses TreeNet3D as a main dataset with nine tree species. From this evalaution, this method shows the best scores in coverage, structural agreement, local-curve matching and branch collisions.

Critical strengths
  • Tree Species Condition 3D generation.
  • Well designed tree specific representation, especially branch geometry representation to get the global and local branch information instead of having one unified information.
  • Five metrics are used to evaluate the quality of generated trees.
Critical weaknesses
  • There is prior work that is very similar, including text-to-species condition with 3D Tree generation with Diffusion models, but this paper did not introduce this work and did not make any comparison. Tree-D Fusion (Lee et al, ECCV 2024).
  • The evaluation method is hard to understand. Authors mention “we generate 200 trees per species and evaluate them on the test split”, but if this is a generation task, what does it mean “test split”? Trees are complex, so I wonder if 200 test cases were enough. 200 might small and the evaluation should be larger.
  • There are no reports on how long it takes and what the peak VRAM is for generating one tree.
  • Since, this is a tree-specific generation method. It will be better to include tree-specific metrics such as DBH, branching angles, the number of twigs or the number of branches.
Questions for authors
  • Why Tree-D fusion is not discussed in the related work and why not compared with this paper since it is a direct comparable baseline.
  • How authors get L-strings from TreeNet3D dataset? Since Latent L-system method needs L-strings, I wonder about the quality of the training data in L-strings.
4 comments

Comment 1: There is prior work that is very similar, including text-to-species condition with 3D Tree generation with Diffusion models, but this paper did not introduce this work and did not make any comparison. Tree-D Fusion (Lee et al, ECCV 2024).

Answer: (Section 2) Thank you for the suggestion. We have expanded Section 2, “Related Work,” to discuss Tree-D Fusion (Lee et al., ECCV 2024) and its relationship to ScaffoldTree. Although both methods involve diffusion models and tree-category conditioning, they address different tasks. Tree-D Fusion requires a reference image and a genus label to reconstruct a 3D tree, whereas ScaffoldTree generates 3D branching structures directly from species conditions without requiring reference images. The former therefore addresses image-guided reconstruction, while the latter targets species-conditioned generation. The reference image is an essential input to Tree-D Fusion’s task, rather than an optional condition that can simply be removed. This difference in task formulation makes a direct quantitative comparison under identical input conditions and evaluation protocols difficult.

Comment 2: The evaluation method is hard to understand. Authors mention “we generate 200 trees per species and evaluate them on the test split”, but if this is a generation task, what does it mean “test split”? Trees are complex, I wonder if 200 test cases were enough. 200 might small and the evaluation should be larger.

Answer: (Section 4.1, Appendix A.12) The test split contains 270 held-out real trees (30 per species), used as a reference distribution, not paired reconstruction targets or generator inputs. We clarify this in Section 4.1, “Evaluation protocol.” We expanded ScaffoldTree and both external baselines from 200 to 400 generated trees per species (1,800 to 3,600 per method). Expanded sets retain the original samples; checkpoints, inference settings, and references are unchanged. The macro-averaged rankings remain unchanged across all six metrics, supporting stable conclusions when the generation budget is doubled. ScaffoldTree maintains higher Coverage and lower Struct. W1-ND, Curve W1, and collision rates, while DivineTree-9x maintains lower Fidelity and MMD. Appendix A.12 reports the full 200/400 comparison.

Comment 3: There are no reports on how long it takes and what the peak VRAM is for generating one tree.

Answer: (Section 4.1) On an NVIDIA RTX 5090, ScaffoldTree takes an average of 8.73 seconds to generate one tree, with observed peak GPU memory of 4.18 GiB. We clarify this in Section 4.1, “Dataset and Implementation.”

Comment 4: Since, this is a tree-specific generation method. It will be better to include tree-specific metrics such as DBH, branching angles, the number of twigs or the number of branches.

Answer: (Section 4.3) We added tree-specific metrics in Section 4.3, “Quantitative Analysis.” Branch count is the total number of branches, including the root; branching angle is measured between a child’s initial segment and the parent tangent at attachment; terminal count is the number of childless branches; terminal ratio divides this count by the total branch count. We compare generated and reference distributions using the same samples and normalized Wasserstein-1 distances, macro-averaged over nine species. ScaffoldTree best matches terminal-branch abundance and proportion, with the lowest terminal-count and terminal-ratio errors and second-best branch-count and angle errors. DBH would be a useful complementary measure. However, it requires physical-scale and radius information not shared by all methods; Latent L-Systems does not predict radii. We therefore evaluate centerline geometry and topology shared by all methods.

Comment 5: How authors get L-strings from TreeNet3D dataset? Since Latent L-system method needs L-strings, I wonder about the quality of the training data in L-strings.

Answer: (Section 4.1, Appendix A.6) Our adapter converts processed TreeNet3D centerlines and parent–child connections into L-strings. Each centerline is encoded segment by segment as yaw, pitch, and forward-step commands, with bracketed child-rule references at discretized attachment positions. We use the released Latent L-Systems architecture and fixed-bin quantization in our adapter. All 11,847 training trees were successfully converted. Encoding and reconstructing all 270 held-out trees preserves branch counts and parent–child connections for all 30 trees in each of the nine species (Appendix A.6, per-species table). Mean paired symmetric squared Chamfer distances are 0.007873 and 0.006006 at 128 and 512 points, respectively, measured in root-aligned, unit-height coordinates without additional rotation or scale fitting. These checks support the conversion as an effective way to construct structurally consistent L-string training targets, with geometric approximation explicitly quantified. Section 4.1 describes the adapter; Appendix A.6 details conversion and validation.

I appreciate the authors. But my concerns are not fully resolved because of missing related work that is very closely related and the specific scores for the tree.

Comment 1: Tree-d fusion needs the image and the tree species in the text, which is also the species condition this method use. I do not think missing the closely related work is a good thing.

Comment 4: Thank you for adding these tree-specific scores, but I do not understand how they are good or how they are aligned.

Comment 1: Tree-d fusion needs the image and the tree species in the text, which is also the species condition this method use. I do not think missing the closely related work is a good thing.

Answer: (Section 2) Thank you for taking the time to revisit our response and clarify this point. Tree-D Fusion is closely related to our work, particularly in combining diffusion models with text-based tree-category conditioning. Our earlier response emphasized the difference in input requirements without sufficiently explaining this important connection. We appreciate your clarification and have revised Section 2 accordingly. The discussion now explicitly recognizes the shared use of diffusion models and tree-category text conditioning, before explaining the methodological differences: Tree-D Fusion uses diffusion priors to reconstruct a tree envelope and then constructs branches through genus-conditioned space colonization, whereas ScaffoldTree directly learns branch geometry and parent topology through scaffold diffusion, branch allocation, and local residual diffusion. Our contribution lies in directly generating explicit branching structures, including branch curves and parent–child topology, through a learned scaffold-to-subtree factorization, rather than species conditioning alone. A direct quantitative comparison would need to account for the additional reference image required by Tree-D Fusion. We have made these connections and distinctions explicit and given the work a fuller discussion in the revised Related Work section.

Comment 4: Thank you for adding these tree-specific scores, but I do not understand how they are good or how they are aligned.

Answer: (Section 4.3, Appendix A.5) Thank you for asking us to clarify the interpretation. The Table 2 reports distribution-matching errors, not raw branch counts or angles. For each species, we compare statistics from 200 generated trees with the same 30 held-out reference trees using normalized Wasserstein-1 distance. Lower values indicate closer agreement with the reference distribution; zero denotes identical empirical distributions of the measured statistic. The normalization scale is computed only from the reference data and shared by all methods, and scores are averaged equally over nine species. No generated tree or branch is paired with a particular reference: comparison is between same-species distributions. These statistics also do not require spatial registration, since counts, ratios, and relative branching angles are invariant to translation, rotation, and uniform scaling. Branch count measures overall branching complexity; branching angle describes local parent-child geometry; terminal count and ratio measure the number and proportion of childless branches. Thus, a lower score means better matching of these properties, not that fewer branches or smaller angles are intrinsically better. ScaffoldTree has the lowest terminal-count and terminal-ratio errors (0.4573 and 0.3640) and ranks second on branch count and angle, supporting its stronger matching of terminal-branch structure alongside competitive global complexity and local geometry. We have clarified the interpretation in the “Tree-specific evaluation” paragraph in Section 4.3 and the Table 2 caption, and provided the calculation details in the Appendix A.5.

Dear Reviewer BwVr,

We are still awaiting your response to our revised manuscript and point-by-point rebuttal.

Following your additional comments, we expanded Section 2 to discuss the close relationship between Tree-D Fusion and ScaffoldTree, and clarified the interpretation and comparison procedure for the tree-specific metrics in Section 4.3, Table 2, and Appendix A.5. Our response identifies the corresponding changes.

We would appreciate your assessment of whether these revisions address your remaining concerns. If further issues remain, your feedback would give us an opportunity to clarify them during the discussion and help ensure that the final assessment considers the revised manuscript and additional evidence.

Thank you for your time.

Reviewer_sj3VOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ScaffoldTree generates 3D trees from a species label. It first generates the trunk and major branches, then decides how many smaller branches to add to each scaffold branch, and finally generates these smaller branches as local subtrees. Both diffusion stages generate branch shapes and decide which branch each one grows from.

The authors test the method on nine TreeNet3D species and compare it with Latent L-Systems and DivineTree. ScaffoldTree does better on coverage, branch-structure statistics, and local branch-curve statistics, while DivineTree has better point-set fidelity and MMD. The ablations test the global/local split, branch allocation, and scaffold context. Leaves are added for visualization, not generated by the model.

Critical strengths
  1. I like the idea of generating the big structure first and adding smaller branches afterwards. The trunk and major branches set the overall shape, while smaller branches fill in the details. Treating these two jobs separately seems reasonable. The method also keeps track of where each branch attaches. Table 2 shows that this split works better than generating the whole tree in one stage.

  2. The allocator is useful, and the ablations help explain why. For example, giving every scaffold branch the same number of small branches may miss the difference between a busy junction and a sparse one. The learned allocator improves the branch-structure statistics even when the scaffold and total number of added branches stay the same (lines 453–466). Tables 3 and 14 give further evidence that this is a real improvement.

Critical weaknesses
  1. I would like to see a comparison with Wang et al. (2025), “Autoregressive Generation of Static and Growing Trees.” It is cited in lines 92–94 but not tested. Since it also generates trees as branches using a multi-resolution Transformer, it seems closely related. The current results are not better on every metric: DivineTree has better MMD in Table 1 and better point-set fidelity in Table 13. This leaves me unsure how ScaffoldTree would compare with a recent autoregressive method.

  2. I am not quite sure how much training each method received. Lines 318–319 list the training epochs, but lines 1025 and 1033–1035 say these are only the final fine-tuning stages. How were the earlier checkpoints trained? This also matters for the ablation: Full trains a global model, a local model, and an allocator, while DirectTreeDiffusion trains one model (lines 1115–1119). Using the same type of backbone does not necessarily mean using the same total model size or training budget. I would like these totals to be made clear.

  3. I am not sure how many trees the fixed limits leave out. The method allows at most 64 scaffold branches, 32 residual branches per anchor, and 192 branches in total. Trees that do not fit these limits are removed during preprocessing (lines 887–891). For example, a tree with 200 branches is already outside the total branch limit. The paper acknowledges this limitation (lines 482–485), but does not say how much of the dataset is removed. I would especially like to know whether trees with many small branches are being left out.

Questions for authors
  1. Why is Wang et al. (2025), “Autoregressive Generation of Static and Growing Trees,” not included as a baseline?

  2. Could you provide the complete training schedule for the checkpoints preceding the final fine-tuning stages described in lines 1025–1035?

  3. Does the advantage over DirectTreeDiffusion remain under a comparable total parameter or training-compute budget?

  4. What fraction of trees is excluded by the capacity and decomposition constraints in lines 887–891, broken down by species?

2 comments

Comment 1: Why is Wang et al. (2025), “Autoregressive Generation of Static and Growing Trees,” not included as a baseline?

Answer: (Section 2) Thank you for this suggestion. We agree that a direct comparison with Wang et al. (2025) would be meaningful, as both works address structured 3D tree generation by modeling branch geometry and hierarchical relationships. Their autoregressive approach offers a relevant comparison to ScaffoldTree, which separates generation into global scaffold diffusion, structure-conditioned branch allocation, and local residual diffusion. Comparing these approaches would help clarify the benefits and trade-offs of autoregressive construction and hierarchical diffusion for generating coherent branching structures. Unfortunately, the authors have not provided publicly available official code with their paper, so we could not directly evaluate an official implementation in our experimental setting. We regret that this quantitative comparison could not be included. We have expanded Section 2, “Related Work,” to give this closely work a fuller discussion and explain how ScaffoldTree differs in branch representation, structural decomposition, and generation framework.

Comment 2: Could you provide the complete training schedule for the checkpoints preceding the final fine-tuning stages described in lines 1025–1035?

Answer: (Section 4.1, Appendix A.4) We have now added the complete training schedule in the Appendix A.4. The revised table summarizes all training stages of Global Scaffold Diffusion, the Allocator, and Local Residual Diffusion, including the purpose of each stage, batch size, learning rate, and epochs. We also clarify that each fine-tuning stage is initialized from the checkpoint obtained in the preceding stage. In addition, we have updated the Section 4.1 Experimental Setup section of the main manuscript accordingly.

Comment 3: Does the advantage over DirectTreeDiffusion remain under a comparable total parameter or training-compute budget?

Answer: (Section4.4, Appendix A.11) We added a total-parameter-matched comparison by widening DirectTreeDiffusion from 512 to 704 dimensions. The resulting model has 273.53M parameters, compared with 274.06M for the complete ScaffoldTree pipeline—a difference of only 0.19%, excluding the frozen CLIP encoder. We retain the original DirectTreeDiffusion training recipe and evaluation protocol. At this comparable capacity, ScaffoldTree reduces Fidelity, MMD, Struct. W1-ND, and Curve W1 by 32.6%, 19.4%, 37.0%, and 28.0%, respectively. The structural and curve-distribution improvements hold across all nine species. These results show that our geometric and structural advantages persist after matching total model capacity. Direct-704 achieves higher coverage and a slightly lower collision rate. We retain Direct-512 as the backbone-controlled ablation in the main text and report the additional comparison in the Appendix A.11 “Parameter-Matched Single-Stage Diffusion.”

Comment 4: What fraction of trees is excluded by the capacity and decomposition constraints in lines 887–891, broken down by species?

Answer: (Appendix A.2) Our previous wording was ambiguous about whether the capacity constraints exclude entire trees or truncate branches within retained trees. We have clarified this distinction in Appendix A.2, “Hierarchy filtering and capacity audit,” and provided a complete per-species breakdown. All 12,387 eligible trees across the nine evaluated species are retained, giving a tree-level exclusion rate of 0% for every included species. The 192-branch cap nevertheless removes branches from 1,986 trees (16.03%): 1,165/1,350 Fir trees (86.30%) and 821/1,396 Flooded gum trees (58.81%). The other seven species require no capacity truncation. Truncation is concentrated in Fir and Flooded gum because their dense branching yields averages of 218.2 and 197.3 branches per tree after initial filtering, respectively. Importantly, breadth-first truncation preserves the level-0 trunk and all level-1 branches in every affected tree; only some level-2 branches are removed. Fir and Flooded gum retain 87.08% and 96.06% of their respective post-filtering branch totals. Thus, the high proportion of affected trees does not imply an equally large loss of branches: their trunk–primary-branch framework remains intact, while secondary-branch detail is reduced. Overall, the cap removes 48,916 of 1,273,871 branches, retaining 96.16%. Appendix A.2 now provides a complete per-species capacity table (Table 7).

Dear Reviewer sj3V,

We would appreciate your feedback on our revised manuscript and point-by-point response.

Following your suggestions, we provided the complete training schedule in Appendix A.4, a total-parameter-matched comparison with DirectTreeDiffusion in Appendix A.11, and a per-species capacity audit in Appendix A.2. We also expanded the discussion of Wang et al. (2025) in Section 2 and explained why a quantitative comparison was not included. The additional experiment shows that ScaffoldTree’s geometric and structural advantages persist at comparable total parameter counts, while the audit clarifies the distinction between tree exclusion and branch truncation.

We would appreciate your assessment of whether these additions address your concerns. If any issues remain, your feedback would allow us to clarify them during the discussion and help ensure that the final assessment considers the revised manuscript and additional evidence.

Thank you for your time and consideration.

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

This paper proposes ScaffoldTree, a species-conditioned framework that decomposes tree generation into global scaffold diffusion, structure-conditioned residual branch allocation, and local residual subtree diffusion. The approach explicitly represents continuous branch geometry and discrete parent–child topology. Results on TreeNet3D show gains in structural matching, coverage, and collision rates over the evaluated baselines.

Critical strengths
  1. Separating major branches from fine-scale branching reflects the structural roles of different parts of a tree.
  2. Joint continuous–discrete generation is appropriate for recursively branching 3D structures.
  3. Learned allocation substantially improves count and anchor predictions compared with uniform assignment.
  4. Coverage, geometry, structural statistics, and collisions give a richer picture than point-set distance alone.
Critical weaknesses
  1. DivineTree-9x obtains better MMD (0.0073 versus 0.0101), while ScaffoldTree improves coverage and structural metrics. The geometric-fidelity versus structural-plausibility trade-off merits per-species analysis.
  2. Results cover nine predefined species with identity conditioning, leaving open whether the model can transfer to unseen species or flexible morphological conditions.
  3. The saliency score uses engineered branch statistics. Robustness to selection thresholds, score weights, and incorrect scaffold choices should be analyzed.
  4. Diversity calibration remains imperfect. The reported diversity ratio improves over ablations but remains substantially above its ideal value of one.
  5. Broader applicability is unclear. It would help to explain whether scaffold-to-subtree diffusion transfers to other recursively structured 3D objects.
Questions for authors

NA

3 comments

Comment 1 : DivineTree-9x has lower MMD, whereas ScaffoldTree improves coverage and structural metrics. The trade-off should be analyzed per species.

Answer: (Section 4.3, Appendix A.7) Thank you for this suggestion. We added a per-species comparison of all three methods using the original evaluation pools. Relative to DivineTree-9x, ScaffoldTree has lower Curve W1 and collision rates for all nine species, lower Struct. W1-ND for six, and higher coverage for four, with two ties. On Lombardy poplar, coverage rises from 46.67% to 93.33%, while collisions fall from 10.200% to 0.910%. DivineTree-9x retains lower MMD for all nine species. These metrics capture complementary properties: Chamfer measures centerline proximity without directly assessing parent connectivity; structural and collision metrics assess branching organization. Coverage counts distinct nearest-reference assignments, not matching distances. Thus, curve and collision gains are consistent across species, while structural and coverage gains vary. Section 4.3 summarizes these findings; Appendix A.7, Table 14 provides the full comparison and discussion.

Comment 2 : Results cover nine predefined species with identity conditioning, leaving open whether the model can transfer to unseen species or flexible morphological conditions.

Answer: (Section 4.2, Appendix A.9) Our goal is to generate multiple species with one shared framework, rather than train a separate pipeline for each species. The main experiments evaluate held-out trees from the nine training species, not unseen-species transfer. Appendix A.9, Figure 12 shows four examples each for “oak tree” and “willow tree,” generated with frozen models and the same four fixed seeds. Both prompts produce trees, but this alone does not establish reliable transfer to the corresponding species’ characteristic morphology. Extending the model to new species and controlling tree shape through richer descriptions remain directions for future work. Section 4.2 clarifies this scope, and Appendix A.9 presents all eight examples and their discussion.

Comment 3: The saliency score uses engineered branch statistics. Robustness to selection thresholds, score weights, and incorrect scaffold choices should be analyzed.

Answer: (Section 4.4, Appendix A.8) We tested saliency weights, scaffold selection quotas, and controlled scaffold membership perturbations on all 270 validation trees. To isolate selection effects, we use reference-derived scaffolds with the downstream allocator and local diffusion model fixed. The results show limited sensitivity to the tested weight changes, with Struct. W1-ND remaining close to 0.2222. In contrast, changing the scaffold selection ratio from 0.30 to 0.20 or 0.40 increases Struct. W1-ND to 0.7672 and 0.6318, respectively. Replacing approximately 20% of non-root scaffold members with lower-saliency alternatives also increases Struct. W1-ND to 0.3650 and the collision rate from 0.891% to 1.120%. These results indicate that the fixed downstream pipeline is relatively insensitive to the tested saliency weights but more sensitive to scaffold density and membership. Section 4.4 summarizes these findings; Appendix A.8 provides the full protocol and results.

Comment 4: Diversity calibration remains imperfect. The reported diversity ratio improves over ablations but remains substantially above its ideal value of one.

Answer: (Section 5, Appendix A.7) Although the diversity ratio remains above one, Full achieves the best diversity calibration among the ablation variants, reducing the ratio from 5.7382 for DirectTreeDiffusion to 2.7788 while matching the highest reference coverage. The complete framework therefore improves dispersion matching without reducing coverage. A separate evaluation of the same Full model with 400 trees per species yields a diversity ratio of 2.1779. Camphor and Ajing olive remain close to reference dispersion, whereas Flooded gum and Lombardy poplar have the largest ratios at 3.4002 and 3.2267. The remaining calibration gap therefore varies substantially by species. Appendix A.7 (Table 17) reports the per-species results for both 200 and 400 generated trees. The main ablation analysis summarizes these findings; Section 5 discusses the remaining calibration challenge.

Comment 5: Explain whether scaffold-to-subtree diffusion can transfer to other recursively structured 3D objects.

Answer: (Section 5, Appendix A.10) ScaffoldTree's core scaffold-to-subtree factorization models branch geometry and explicit parent–child topology rather than botanical growth processes. This formulation could extend to tree-like vascular structures or neuronal arbors, with domain-specific changes to decomposition criteria and geometric constraints. Section 5 and Appendix A.10 discuss this potential applicability and the rooted, acyclic representation.

Thank you for the response. The new per-species evaluation clarifies the trade-off between point-set MMD and structural plausibility, and the controlled scaffold-sensitivity experiments are particularly helpful. My remaining concerns are that the approach still relies on a hand-designed scaffold selection mechanism, diversity calibration remains imperfect, and broader transfer beyond the training species or tree domain has not been established. However, I tend to maintain my score.

Thank you for your reassessment and constructive feedback. We appreciate your recognition of our additional analyses and suggestions that have helped improve our work.

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

We are grateful for the thoughtful feedback throughout the discussion. It has helped improve ScaffoldTree’s evaluation, reproducibility, and clarity. The revised manuscript incorporates these changes, with their locations provided in our individual replies.

  1. Reviewer QDXn’s questions strengthened our analysis of the method’s strengths and limitations. We added per-species comparisons and controlled scaffold-selection sensitivity experiments, and clarified the remaining limitations in diversity calibration and generalization. We appreciate the reviewer’s recognition of these additions and their confirmation that they maintain their weak-acceptance rating.

  2. Reviewer sj3V’s questions strengthened reproducibility and the experimental comparisons. We provided the complete training schedule, a total-parameter-matched comparison with DirectTreeDiffusion, and a per-species capacity audit distinguishing tree retention from branch truncation. We also expanded the discussion of the closely related autoregressive method. The additional comparison shows that our geometric and structural advantages persist after matching total parameter counts. We would appreciate the reviewer’s assessment of these additions.

  3. Reviewer BwVr’s feedback helped clarify the relationship to prior work and the evaluation protocol. We expanded the discussion of Tree-D Fusion to acknowledge its close connection and explain the methodological differences. We also doubled the generation budget for ScaffoldTree and both external baselines, reported inference time and memory usage, validated the L-string conversion, and added tree-specific metrics. Our latest reply further clarifies how these metrics compare generated and reference distributions and how to interpret their scores. We would appreciate the reviewer’s assessment of whether these clarifications address their remaining concerns.

Our ScaffoldTree’s contribution is a learned scaffold-to-subtree factorization that generates explicit branch geometry and parent-child topology through global scaffold diffusion, structure-conditioned residual allocation, and local subtree diffusion. The additional experiments provide further support for this design.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.