Research map / 3D vision / 3D generation
Slice-100K: A Multimodal Dataset for Extrusion-based 3D Printing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Slice-100K, a multimodal dataset of over 100,000 STL/G-code pairs with renderings, LVIS categories and geometric properties, and demonstrates it by fine-tuning GPT-2 for Sailfish-to-Marlin G-code translation.
Problem
There is no large curated repository pairing CAD models with their G-code files for extrusion-based additive manufacturing, and LLMs struggle with low-level G-code manipulation.
Approach
Build a large STL-to-G-code dataset by slicing Objaverse-XL Thingiverse and Thingi10K models with PrusaSlicer in two G-code flavors, and use contour flipping and pair creation to make Sailfish/Marlin chunk pairs for LLM translation.
Key results
- GPT-2 finetuned on 25 shapes reaches IOU@0.9 of 98% and IOU@0.95 of 94% for Sailfish-to-Marlin translation, versus 67% and 61% for base GPT-2.
- Finetuning on as few as 1 shape (49 layers) raises IOU@0.99 from 4% to 27%.
- GPT-3.5, GPT-4, Claude-2, Llama-2-70b and Starcoder struggle with G-code scaling; GPT-3.5 and GPT-4 fail on the S-shape.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Finetune GPT-2 on 1 Slice-100K shape rather than using the base model | GPT-2 · G-code translation | IOU@0.9: 67 -> 95 | Helped |
| Finetune GPT-2 on 1 Slice-100K shape rather than using the base model | GPT-2 · G-code translation | IOU@0.95: 61 -> 88 | Helped |
| Finetune GPT-2 on 1 Slice-100K shape rather than using the base model | GPT-2 · G-code translation | IOU@0.98: 17 -> 71 | Helped |
| Finetune GPT-2 on 1 Slice-100K shape rather than using the base model | GPT-2 · G-code translation | IOU@0.99: 4 -> 27 | Helped |
| Increase finetuning data from 1 to 5 Slice-100K shapes | GPT-2 translation model · G-code translation | IOU@0.95: 88 -> 91 | Helped |
| Increase finetuning data from 5 to 25 Slice-100K shapes | GPT-2 translation model · G-code translation | IOU@0.98: 74 -> 71 | Hurt |
| Increase finetuning data from 5 to 25 Slice-100K shapes | GPT-2 translation model · G-code translation | IOU@0.99: 30 -> 30 | No clear effect |
Surprising findings
- Increasing training data beyond five shapes stopped improving translation quality, which the authors attribute to the chunking reducing the task to a local mapping.
- The finetuned translation model often outputs its input unchanged, ignoring Marlin-specific commands.
Limitations
- Difficult to verify the LVIS category labels assigned by the CLIP-based pipeline.
- All models were sliced only along the default Z-direction, limiting multi-directional slicing research.
- Focuses only on extrusion-based printing rather than other additive manufacturing methods.
Experimental setup
Datasets: Objaverse-XL (Thingiverse subset) Thingi10K LVIS
Benchmarks: Slice-100K
Metrics: image-space IoU IOU@k
Baselines: GPT-2 Base GPT-3.5 GPT-4 Claude-2 Llama-2-70b Starcoder
Builds on: Objaverse-XL Thingi10K CLIP LVIS
Closest papers by meaning
- CraftsMan3D: High-fidelity Mesh Generation with 3D Native Diffusion and Interactive Geometry Refiner (CVPR 2025 · similarity 0.83)
- Generating CAD Code with Vision-Language Models for 3D Designs (ICLR 2025 · similarity 0.83)
- Position: You Can't Manufacture a NeRF (ICML 2025 · similarity 0.83)
- MV2Cyl: Reconstructing 3D Extrusion Cylinders from Multi-View Images (NeurIPS 2024 · similarity 0.82)
- Synthesizing Multimodal Geometry Datasets from Scratch and Enabling Visual Alignment via Plotting Code (ICML 2026 · similarity 0.82)
- SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.