acceptodds
Under review as a conference paper at ICLR 2027

UniCAD: Unified CAD Generation with Continuous Conditional Guidance

Abstract

Generating editable CAD models with precise geometry from single-view images is crucial for downstream design and manufacturing applications. However, existing methods are either limited to unconditional generation or rely on restrictive sketch-and-extrude pipelines, limiting user control and generative diversity. We propose UniCAD, a unified generative model that bridges 2D images and structured 3D CAD representations. It models topology with a connectivity-guided autoregressive Transformer decoder and geometry with a Transformer-based diffusion model. In addition, UniCAD supports controllable CAD generation from a single image, enabling smooth interpolation and precise control over conditioning strength. Extensive experiments show that our model supports conditional and unconditional generation within a unified framework, delivering high-quality results.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
B-Rep, Diffusion, CAD Modeling, CAD Generative Modeling, 3D Generation, Generative Models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_5Zy1OpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents UniCAD, a unified framework for single-image-conditioned and unconditional BRep CAD generation. UniCAD generates topology autoregressively using connectivity-based face ordering and a pointer network, then generates bounding boxes, vertices, edges, and surfaces through multi-stage geometry diffusion. It integrates image guidance across discrete topology and continuous geometry generation, and models the complex joint CAD distribution with a two-stage decoupled architecture.

Critical strengths
  1. The unified framework offers practical value. It supports both image-conditioned and unconditional BRep generation, with both modes demonstrated experimentally.

  2. The generation quality improvements are supported by quantitative evidence. UniCAD outperforms the listed baselines in conditional CD on DeepCAD and unconditional MMD and JSD on both datasets.

Critical weaknesses
  1. The abstract states that existing methods are either unconditional or restricted to sketch-and-extrude pipelines, while Sec. 2 explicitly identifies HoLa and CMT as image-conditioned methods for direct BRep generation. This weakens the justification for the research gap and the distinction from prior work.

  2. The meaning of “image fidelity” is unclear (Line 237). The model outputs BRep, but Sec. 4.1 states that the guidance scale controls the trade-off between image fidelity and generative diversity.

  3. The FeF ordering rules are incompletely specified. Sec. 4.2 does not define whether “geometric scale” means surface area, bounding-box size, or another quantity.

  4. The independent effect of the random offset zz is not validated. Sec. 4.1 attributes increased variation to zz, but corresponding ablations are not included. Existing ablations do not establish whether zz contributes additional meaningful diversity.

  5. Fig. 2 shows baseline results and failure cases on ABC, and Table 1 reports only UniCAD’s results. This may be because the baselines did not succeed on all samples, but discussing why the baselines fail and explaining why their quantitative results are omitted would help clarify UniCAD’s advantages.

  6. The independent novelty and benefits of the unified conditioning mechanism are insufficiently established. Sec. 2 acknowledges that condition dropout and CFG jointly model conditional and unconditional distributions and control conditioning strength, while Eqs.5–6 largely follow this established mechanism. The ablations do not isolate gains over standard CFG-based conditioning, leaving the “unified probabilistic generative space” insufficiently justified as a major methodological contribution.

Questions for authors
  1. What are the inference speed and model parameter counts of ICG-DiT? Per-sample inference time, total parameters, or trainable parameters are not reported.

  2. Is ICG-DiT trained from scratch or initialized from pretrained weights?

1 comment

Thank you for your responses. The additional clarifications address some of my concerns. I maintain my score and increase the confidence.

Reviewer_7bqrOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This work proposes a method for image-conditioned generation of CAD models using the BRep representation. The model builds on the standard encoder-decoder Transformer architecture, Diffusion Transformers (DiT), and classifier-free-guidance (CFG). The BRep data representation is based on prior work DTGBrepGen (Li et al, 2025a), and this work largely follows their model architecture with some notable changes including an autoregressive decoder and importance ordering. The conditional method is evaluated on two datasets: DeepCAD and ABC (Tab.1). On DeepCAD, it compares with GenCAD and CADDreamer where it improves on one metric (CD) but not on another (Validity). On ABC, the reported scores are not compared with any baselines. Scores for Img2CAD are reported on another dataset, but the proposed method does not report scores for that dataset (why?). While this work focuses on conditioned generation, the paper also reports scores on unconditional generation, where it improves on some metrics but not on others. Given this, the experiments are not too convincing (more below under weaknesses).

Critical strengths

None

Critical weaknesses

1) Limited novelty

  • The paper could be described as DTGBrepGen with additional cross-attention to an encoded image for conditioning. This is the standard and very-well established approach to include conditioning to a generative model. The only other 2 differences are the replacement of the VAE with an auto-regressive transformer to obtain the face adjacency matrix, and a different ordering of the faces. The ablation study (Tab. 3) seems to suggest that these are helpful in the conditional setting, but less so in the un-conditional one.

  • A substantial part of the paper (Fig1 top left, Sec 4.1, Sec. 5.2, Fig 3.) describes and shows the effect of classifier-free-guidance (CFG). This feels strange, since it’s used by basically every modern generative approach since 2022.

2) Experiments

  • Limited evaluation. The key contribution of the paper is image-conditioned generation, yet a large number of experiments focuses on unconditional generation. The only comparison with baselines in the conditional-generation setting is Table 1 which only compares with two baselines on DeepCAD, for ABC no comparisons are provided, same for the Img2CAD dataset. All three datasets are publicly available, code is available for these baselines -- why are they missing? Just reporting scores of various methods on different datasets are of course not comparable so only the first 3 lines in Tab 1 are helpful while the last 2 rows are just decoration.
  • I find the evaluation setting strange, the input images look like renderings of existing CAD models, suggesting that the CAD models are already available. A more practical setting would be where the input images are photorealistic images of real objects. These could for example be generated with an image generation model conditioned on the current CAD renderings. Alternatively a 3D scan of a real object could be the conditioning. Of course this would change the conditioning modality, which is described in the limitations Sec.5.3 as future work, but really it is simply replacing the ViT image encoder with a standard pre-trained point cloud encoder like for example PTv3 -- this is very little effort. I also understand that prior work might follow a similar setup, but that does not make it better.
  • The qualitative results in the paper look very good, but it is unclear how well the model generalises to out-of-distribution examples or how well it performs on real images. Failure cases to better understand the limitations would be helpful.
Questions for authors

Additional experiments (affecting assessment)

  • Tab. 1: Missing scores of baselines (GenCAD, CADDreamer, Img2CAD) on ABC and ShapeNet dataset from Img2CAD.
  • Img2CAD released their dataset on HuggingFace, this method should be evaluated on the same setting.
  • Eq. 4 - how is this coarse-to-fine generation approach implemented, and how is it evaluated?
  • L.222-225 mention learnable memory tokens - an analysis of their effect would be helpful to appreciate their use.
  • How does the proposed method compare to DTGBrepGen with an added cross-attention to a VIT-encoded image?

Clarification Questions (likely not affecting assessment)

  • Sec. 4.3 - Eq 11 - how are these stages implemented? It seems the DiT predicts them in parallel while a sequential prediction would be more aligned with the text. Could this be clarified?
  • How is the CD computed, given that one cannot recover metric scale from a monocular image?
  • How are the output parameters normalized / scaled?

Notation remarks: (not affecting assessment)

  • The notation for both FeF and EV seems confusing. At first, they look like matrix/vector multiplications FeF and E*V -- maybe a single letter matrix name would be more clear?
  • Eq. 5: \phi is not defined, Eq. 2: N_f is not defined, Eq. 1: i is not defined. etc
  • An illustration would be helpful to better understand the topology representation.
  • Missing explanation of metrics MMD, COV, JSD etc (at least put a reference).
Reviewer_kGxtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

UniCAD generates BRep CAD models from a single image by first predicting topology and then generating geometry through a cascade of diffusion models. It builds on DTGBrepGen and adds image conditioning, connectivity-based face ordering, and a stochastic unconditional representation. Classifier-free guidance allows the same model to support conditional and unconditional generation. Experiments on DeepCAD and ABC show improved reconstruction accuracy and competitive unconditional generation.

Critical strengths
  1. The DeepCAD reconstruction results are promising: CD improves substantially over the evaluated baselines.
  2. The ablations provide useful evidence for the face-ordering strategy and show an informative trade-off between conditional and unconditional performance.
Critical weaknesses
  1. The contribution over DTGBrepGen is not sufficiently isolated. The representation, topology–geometry decomposition, and staged geometry generation largely follow that framework. A matched image-conditioned DTGBrepGen baseline would help establish the benefit of the proposed changes. Conditional comparisons are also limited: ABC has no quantitative baseline, and HoLa is compared only qualitatively.
  2. The evidence for controllable diversity is incomplete. Appendix A shows that weaker conditioning increases CD, but this alone does not establish greater diversity or smooth shape variation. Validity across guidance strengths is missing. It is also unclear how construction failures are handled in CD, which matters given the 50.8% conditional validity on ABC.
Questions for authors
  1. How does UniCAD compare with DTGBrepGen using the same image conditioning and training setup?
  2. For a fixed input image, how do diversity and validity change with guidance strength?
  3. How are failed BRep constructions handled when computing CD?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.