- Rating
- 3(+1)
- Confidence
- 3(+1)
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper addresses the generation of executable parametric CAD sequences from a single raster sketch. CAD Construction Process Recovery (CCPR) recursively predicts preceding construction-state features from the final image until reaching an empty-state reference, then decodes adjacent-state transitions into sketch-extrude operations in the forward direction. Error-Aware Redrawing (EAR) detects erroneous operations in the initial sequence, regenerates the affected operations, and refines line endpoints using GCN-predicted geometric relations and SLSQP. The model is jointly trained with five supervised losses while freezing the DINOv2 backbone. Ground-truth intermediate states are used for training supervision, not as inference inputs. Evaluation covers DeepCAD, zero-shot transfer to Fusion 360, hand-drawn sketches, component ablations, construction-prefix accuracy, computational cost, and failure cases.
Critical strengths
-
Explicit state recovery has substantive, partially controlled experimental support. The method connects the ambiguity of composite raster sketches to operation-level state transitions and supervises recovery using features of actual construction states. In Table 3, the invalid ratios for Forward + state supervision, Backward without state supervision, and CCPR are 21.0%, 19.5%, and 4.81%, respectively. Table 5 reports the same parameter count and computation for these three recovery variants: 49.22M parameters and 33.61 GFLOPs. These comparisons support the combined use of backward recovery and state supervision under the reported setting; differences in parameter count or GFLOPs among these variants do not explain the improvement.
-
The system evaluation covers reconstruction quality, transfer, refinement, and computational trade-offs. The paper reports in-domain evaluation, direct cross-dataset transfer, and hand-drawn inputs, with all task-specific training conducted on DeepCAD. CADFlow achieves 82.4% parameter accuracy and 3.9% invalid ratio on DeepCAD, and 77.8% parameter accuracy and 9.7% invalid ratio on Fusion 360. Compared with redrawing every operation, selective redrawing improves command accuracy from 78.6% to 82.3% while reducing median latency from 183.70 to 132.44 ms. The hand-drawn evaluation and explicit discussion of failures on invisible back-side details further clarify the practical scope of the system.
Critical weaknesses
-
The evaluation protocol and baseline adaptations are not specified well enough to verify the central comparisons. Section 4.1 gives only high-level definitions of command accuracy, parameter accuracy, invalid ratio, and Chamfer distance. It does not fully specify parameter tolerances and denominators, CD point sampling and distance convention, normalization and aggregation, or how invalid outputs are handled when computing CD. The textual definition of invalid ratio refers to CAD sequences, whereas Tables 1 and 2 also compare B-rep and CADQuery outputs; a common geometric-validity criterion across these representations needs clarification. Appendix A.2 describes shared source data, retraining, and input-module adaptations, which are positive steps. However, the actual evaluation counts and conversion/filtering rules, baseline versions, feature connections and freezing choices, and training/tuning budgets remain insufficiently documented. Free2CAD also receives ordered stroke information unavailable to raster-only methods, so this difference in input conditions should be explicit. These omissions limit confidence that the reported ranking isolates the contribution of CADFlow. They establish a need for clarification, rather than demonstrating that the comparisons are unfair.
-
The evidence does not fully validate the claimed mechanisms or their reliability boundaries. The matched recovery variants provide useful evidence, but direct generation and CCPR require 16.47 and 33.61 GFLOPs, respectively. This comparison does not fully separate the effect of construction-process modeling from differences in architecture, supervision, and computation. The EAR ablations support an overall benefit, but the paper does not report error-detection precision/recall, the frequency of changing correct operations into incorrect ones, or selection comparisons using the same redrawing model and initial predictions. Geometric constraint refinement also lacks direct measurements of relation-prediction quality, angular residuals, constraint satisfaction, and solver success or fallback rates. Adding geometric constraint refinement changes parameter accuracy from 82.3% to 82.4%, invalid ratio from 3.96% to 3.88%, and the scaled CD reported in Table 3 from 14.3 to 13.7, while increasing median latency from 132.44 to 173.15 ms. Without repeated-run variability, the stability of these gains is unclear. Appendix A.2.3 describes replacement of existing operations, but does not report insertion or deletion to correct an erroneous operation count; termination and length errors therefore remain relevant reliability questions. These gaps limit the evidence for meaningful error localization and improved geometric consistency, even though the reported system-level improvements are positive.
Questions for authors
-
Please specify the evaluation counts and conversion/filtering rules, parameter-accuracy tolerances and denominators, treatment of sequence-length mismatches, the common validity criterion across output representations, and CD sampling, normalization, aggregation, and invalid-output handling. If standard implementations are used, please identify their versions and modifications. These details could materially affect confidence in Tables 1 and 2.
-
Please document the adapted baselines' DINOv2 feature connections, freezing choices, versions, and training/tuning budgets. For Table 3, are the recovery variants identical in architecture and training settings apart from recovery direction and state supervision? How are differences in the image encoder, supervision, and inference computation controlled in the direct-generation comparison? These answers would affect the assessment of both comparative performance and the contribution of CCPR.
-
Please define the exact criterion used to construct operation-error labels. Are existing measurements available for incorrect-to-correct and correct-to-incorrect changes after redrawing, termination and operation-count failures, geometric-relation residuals, solver fallback rates, and variability in GCR gains? If these have not been measured, please clarify which mechanism claims can be supported by the current evidence.
5 comments
Q2-1. Please document the adapted baselines’ DINOv2 feature connections, freezing choices, and versions.
A: For Drawing2CAD, CADrawer, and BrepDiff, which have no native raster or image interface, we use a frozen DINOv2 ViT-S/14 with registers. We replace only the input module and feed the features into the original downstream networks. For BrepDiff, the features are injected into its original denoising network as a conditioning signal.
Methods that natively accept images (GenCAD, Img2CAD, Img2CADSeq, CADRILLE, and CADReasoner) keep their original encoders and do not use DINOv2. HiFi-BRep already conditions on images and wireframe sketches via a pretrained DINOv2 model; we follow this native interface and retrain it on our DeepCAD raster sketches. Please see Appendix A.2.2.
Q2-2. Please document the training/tuning budgets.
A: Each compared method retains its own original training and tuning settings, which are recorded in Appendix A.2.4 and Table 5. Multi-stage methods retain their stage-wise schedules; we did not impose a shared epoch or search budget. Epochs, steps, and batch sizes are method-specific.
Q2-3. Free2CAD also receives ordered stroke information unavailable to raster-only methods. Please make this difference in input conditions explicit.
A: We have reported the difference in Appendix A.2.2 in the original submission, and added further discussions.
Q2-4. For Table 3, are the recovery variants identical in architecture and training settings apart from recovery direction and state supervision?
A: Yes. The three state-recovery variants in Table 3 use the same architecture, image encoder, and training settings. They differ only in recovery direction or intermediate-state supervision, and have identical parameter counts and computation (49.22M parameters; 33.61 GFLOPs).
Q2-5. How are differences in the image encoder, supervision, and inference computation controlled in the direct-generation comparison?
A: The image encoder is unified by using the same frozen DINOv2. Supervision differs in that the direct-generation variants have no intermediate-state supervision, while CCPR uses backward recovery with state supervision. Inference computation is not equalized; we report parameters, GFLOPs, and median latency.
Compared with CCPR, the encoding-stage parameters in direct generation are fully frozen, so we introduce two trainable Transformers for this comparison. We also report a direct-decoding experiment without these Transformers:
| Variant | ACC_cmd ↑ | ACC_param ↑ | IR ↓ | CD ↓ | Parameters | GFLOPs | Median latency |
|---|---|---|---|---|---|---|---|
| Direct decoding | 79.2% | 75.8% | 27.29% | 26.2 | 41.20M | 14.62 | 13.87 ms |
| Direct generation | 80.2% | 76.6% | 23.21% | 23.1 | 55.54M | 16.47 | 14.16 ms |
| CCPR | 81.3% | 81.7% | 4.81% | 18.7 | 49.22M | 33.61 | 130.42 ms |
CCPR improves reconstruction, particularly geometric validity, at higher inference cost. Please see Appendix A.3.3 and Table 6.
Q3-1. Please define the exact criterion used to construct operation-error labels.
A: An entire sketch–extrude operation is labeled erroneous if any command mismatches the reference, or if any valid parameter violates the evaluation criterion (continuous quantized error <3, categorical exact match). Please see Appendix A.2.3 for the operation-error label, and Appendix A.2.5 for the quantized-error tolerance and exact-match rule.
Q3-2. Are existing measurements available for error-detection precision and recall?
A: Yes. Error-detection precision is 98.89% and recall is 71.53%. Please see Appendix A.3.7 and Table 10.
Q3-3. Are existing measurements available for incorrect-to-correct and correct-to-incorrect changes after redrawing?
A: Yes. After redrawing, 41.42% of initially incorrect operations become correct, and 0.65% of initially correct operations become incorrect:
| Measure | Result |
|---|---|
| Incorrect → correct after redrawing | 41.42% |
| Correct → incorrect after redrawing | 0.65% |
Please see Appendix A.3.7 and Table 10.
Q3-4. Are existing measurements available for selection comparisons using the same redrawing model and initial predictions?
A: Yes. All-operation and selective redrawing use the same redrawing model and the same initial predictions; only operation selection differs.
Q3-5. Are existing measurements available for termination and operation-count failures? Can EAR insert or delete operations to correct an erroneous operation count?
A: Yes. Operation-count agreement with the reference is 84.11% for CCPR and 86.02% after EAR. EAR can terminate early and truncate subsequent operations, but cannot insert operations or independently delete an intermediate operation. Please see Appendix A.3.7.
Q3-6. Are existing measurements available for geometric-relation prediction quality, residuals, and constraint satisfaction?
A: Yes. Relation prediction on actual EAR outputs achieves 97.61% accuracy and 97.10% macro-F1. Precision and recall are 92.53% / 99.92% for parallel relations and 94.33% / 99.98% for perpendicular relations.
| Angular residual | Before GCR: mean | After GCR: mean |
|---|---|---|
| Parallel | 0.1490° | 0.1063° |
| Perpendicular | 0.1794° | 0.1303° |
| Overall | 0.1640° | 0.1182° |
GCR reduces the overall mean angular residual by 27.9%, which is our direct measure of improved constraint satisfaction. Please see Appendix A.3.7 and Table 11.
Q3-7. Are existing measurements available for solver fallback rates and variability in GCR gains?
A: Yes. Of 5,681 solver calls, 54 did not converge and reverted to the initial parameters, giving a 0.95% fallback rate. Across three independently trained random seeds, we obtain:
| Metric | Without GCR | With GCR |
|---|---|---|
| ACC_param (%) ↑ | 82.2 ± 0.0319 | 82.3 ± 0.0342 |
| CD ↓ | 14.3 ± 0.0070 | 13.7 ± 0.0067 |
| IR (%) ↓ | 3.96 ± 0.0887 | 3.88 ± 0.0730 |
Please see Appendix A.3.7 and Table 12.
Thank you for your insightful comments. We answer your concerns and questions in the following.
Q1-1. Please specify the evaluation counts and conversion/filtering rules. If standard implementations are used, please identify their versions and modifications.
A: We use the original DeepCAD evaluation implementation without modifying its metric scripts. The DeepCAD and Fusion 360 test splits contain 7,881 and 1,346 CAD models, respectively. All test models were successfully converted for evaluation, with no exclusions during data preparation. These evaluation details are provided in Appendix A.2.5.
Q1-2. Please specify the parameter-accuracy tolerances and denominators.
A: Command accuracy uses position-wise matching. Parameter accuracy considers valid parameters only at correctly predicted commands: continuous parameters require a quantized error <3, and categorical parameters require exact matching. The denominator of parameter accuracy is therefore the number of valid parameters at correctly predicted commands. Both accuracies are computed per sample and then averaged. Please see Appendix A.2.5.
Q1-3. Please specify the treatment of sequence-length mismatches.
A: If a prediction is shorter than the reference, the missing command positions are counted as mismatches. If it is longer, commands beyond the reference length are excluded from sequence accuracy. Concretely, we insert EOS at each prediction’s stopping position and compare predictions up to the reference EOS. Please see Appendix A.2.5.
Q1-4. Please specify the common validity criterion across output representations, as well as CD sampling, normalization, aggregation, and invalid-output handling.
A: The common validity criterion is successful conversion to a solid. CD sampling is 2,000 points per shape. Normalization is the same procedure across representations. Aggregation is the arithmetic mean of bidirectional squared nearest-neighbor distances over valid outputs. Invalid outputs (solid-construction or point-sampling failures) count toward IR and are excluded from CD; the IR denominator is all evaluated test models. Please see Appendix A.2.5.