3D Generation from STL Files — Working Notes31.10.26

# 3D Generation from STL Files — Working Notes

## 1. LLM vs 3D generative model

- An LLM trains on **text tokens**. STL files are geometry — not usable for an LLM.
- What you can build from 300k STLs is a **3D generative model**.
- If you want an "LLM" angle: caption each STL, then fine-tune a language model on the captions — a text model *about* the models, not a model of geometry.

## 2. STL basics

- **Point / vertex** — one `(x, y, z)` coordinate.
- **Triangle / face** — three vertices connected; the surface tile.
- **Facet** — STL's word for one triangle (normal + 3 vertices). Facet ≡ triangle in STL.
- **Mesh** — thousands of triangles stitched edge-to-edge.
- STL stores geometry only: no colour, no units, no construction history.

```
facet normal 0 0 1
  outer loop
    vertex 0 0 0
    vertex 1 0 0
    vertex 0 1 0
  endloop
endfacet
```

## 3. Why triangles

- Always planar (3 points define a plane; a quad can bend).
- Can be split/joined without changing the surface.
- Simple math for normals, lighting, hit-testing.
- GPUs draw triangles natively.

## 4. Triangle count

- Binary STL: `(file size − 84) / 50` = triangles (84-byte header + 50 bytes per triangle).
- ASCII STL: `grep -c "facet normal" file.stl`.
- `len(trimesh.load("x.stl").faces)`.
- `admesh` prints `Number of facets`.
- MeshLab / Blender / f3d show it in the UI.

## 5. More triangles?

- Tradeoff, not "more is better".
- Too few → visible faceting (bad for jewelry).
- Enough → smooth at printer resolution.
- Too many → bigger files, slower slicing, no visible gain, messier meshes.
- Ceiling = output resolution: FDM ~0.4 mm nozzle / 0.1–0.2 mm layers; resin ~25–50 µm pixels.
- Prefer **adaptive** tessellation and clean topology.

## 6. Repair: admesh vs trimesh

- **admesh** — C CLI, STL-focused: fixes normals, fills holes, drops unconnected/degenerate facets, prints edge stats. GPL-2.0, in Debian.
- **trimesh** — Python, broader: same repairs (`fix_normals`, `fill_holes`, `remove_degenerate_faces`, `process`) + formats, transforms, sampling, booleans, scenes. MIT.
- For 300k files, drive with trimesh; admesh as pre-clean/fallback.

## 7. Pipeline

1. **Audit** — parse check (binary/ASCII, header, triangle count), dedupe (vertex hash + render hash), watertightness (each edge shared by 2 triangles), scale outliers (bbox histogram), degenerates (zero-area, duplicates, flipped normals).
2. **Normalize** — load with trimesh, merge vertices, fix winding, centre at bbox centre, scale max extent = 1, canonical orientation (PCA or fixed up-axis), export one format.
3. **Representation** — choose point cloud / voxel-SDF / multi-view / mesh tokens.
4. **Train / fine-tune** — fine-tune an open base model, don't train from zero.
5. **Evaluate** — sample generations, check domain fit.

## 8. Representations

- **Point cloud** (easiest): `mesh.sample(20000)` → `(N, 3)` tensor (+ normals optional) → `.npy`/`.npz`/`.ply`. Loses connectivity; re-mesh with Poisson.
- **Voxel / SDF grid**: 64³ or 128³ occupancy or signed distance. Heavier, smoother.
- **Multi-view renders**: 12–24 turntable views per STL → image→3D models. Reuses pretrained models.
- **Mesh tokens**: serialise triangles into tokens (MeshGPT). Needs a custom tokenizer. Hardest.

## 9. Free / open base models

- Point-E (OpenAI) — MIT — point clouds
- Shap-E (OpenAI) — MIT — implicit → mesh
- TRELLIS (Microsoft) — MIT — mesh/3D assets
- TripoSR (Stability AI + Tripo) — MIT — image → 3D
- InstantMesh — Apache-2.0 — image → mesh
- One-2-3-45 — MIT — image → 3D
- Zero123 / Zero123-XL — MIT — image → novel views
- LGM — open — gaussian splats
- Hunyuan3D (Tencent) — free weights, community license
- Stable Fast 3D — free weights, Stability community license

Commercial, not open: Meshy, Tripo, Rodin.

## 10. Practical notes

- You don't need to be a 3D/jewelry designer: audit/normalize is a script over all 300k, not manual work.
- Preview by sampling: headless thumbnails (trimesh + pyrender, `f3d --output`) into a grid; outliers pop out.
- Fix by rule, log failures — don't hand-fix.
- 300k STLs is a real dataset; quality (watertight, manifold, consistent orientation) matters more than count.
- GPU cost: point clouds ≪ voxels ≪ mesh tokens.

## 11. Open-source stack

- OpenSCAD (GPL) — parametric CAD
- admesh (GPL) — STL repair
- trimesh (MIT) — batch audit/normalize/sampling
- pymeshlab / MeshLab (GPL) — repair/inspection
- Open3D (MIT) — programmatic viewer/processing
- f3d — fast STL viewer
- PrusaSlicer / Cura / Bambu Studio — print preview
- Blender — visual/rendered look

## Bottom line

Start with **point clouds**, fine-tune **Point-E / Shap-E / TRELLIS**, and put the effort into **audit + normalize** — that is where the dataset's value lives.

Anonymous

3D Generation from STL Files — Working Notes31.10.26

Information Epoch 1793785834