GIFT: Bootstrapping Image-to-CAD Program Synthesis via Geometric Feedback
TLDR
GIFT uses geometric feedback to augment training data for image-to-CAD program synthesis, improving IoU by 12% and reducing inference compute by 80%.
Reasoning
The paper presents a novel data augmentation framework that leverages geometric feedback to generate high-quality training samples from test-time compute, addressing the scarcity of diverse CAD program-image pairs. Strengths include a clear bottleneck identification and two effective mechanisms (GIFT-REJECT and GIFT-FAIL), while weaknesses are the lack of explicit real-world dataset details and potential reliance on synthetic benchmarks.
Read-first score
Read-first score 57, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 69.
Field roles
Rank sensitivity
Stability: volatile; rank range: 4.
Keyword Scores
Deep Analysis
Innovations
- Geometric Inference Feedback Tuning (GIFT) framework that turns test-time compute into bootstrapped training samples via geometric feedback
- Soft-Rejection Sampling (GIFT-REJECT) that retains diverse high-fidelity programs beyond exact ground-truth matches
- Failure-Driven Augmentation (GIFT-FAIL) that converts near-miss predictions into synthetic training examples to improve robustness
- Amortization of inference-time search into model parameters, reducing inference compute by 80%
Methodology
GIFT is a data augmentation framework that uses geometric feedback to generate high-quality training samples from test-time compute. It combines Soft-Rejection Sampling to keep diverse programs with high geometric fidelity and Failure-Driven Augmentation to turn near-miss predictions into synthetic examples, bootstrapping training data without additional human annotation.
Key Results
GIFT improves mean IoU by 12% over a strong supervised baseline, reduces inference compute by 80%, and remains competitive with more complex multimodal systems.