img2threejs explained: turning one photo into editable Three.js code with 80k-180k tokens
maier_ak · x · 2026-09-04
img2threejs, a repo that went viral in summer 2026 with 13k+ stars in weeks, is not a neural image-to-3D pipeline at all. It lets a multimodal coding model (Claude Code, Codex, or OpenCode) read a photo, emit a JSON spec plus TypeScript, render the result in a browser, and iteratively refine it via side-by-side comparison until it matches.
- Each object costs only 80k-180k model tokens, far less than full image-to-mesh pipelines, and the output is human-readable, diffable, version-controllable, and animation-ready.
- The whole system lives as a skill in the model's skill directory; the Python orchestration scripts use only the standard library and never touch pixels beyond packaging screenshots — the real backbone is the host coding model.
- No quantitative fidelity claims: results rely on the model's prior. It succeeds on hard-surface objects the model knows (weapon skins, cartoon props) but fails on organic shapes, and swapping the host model changes the generated code.
More from coding & agent
- Reddit debate: How many AI-built apps do you actually run in production? — LocustKitten · 2026-09-04
- Chromium Dev Explains Why Repetitive Inline Styles Hurt Rendering vs Shared StyleSheets — mohamedmansour · 2026-09-04
- How Ollama Tames LLM Complexity With a Single LlamaServer Interface — Mahmoud_Zalt · 2026-09-04
- Codex compaction validates RLM principles: no destructive summaries, let models search context — lateinteraction · 2026-09-04
- Grok Bot tip: your PC becomes a remote host you can command from your phone — dean_rie · 2026-09-04
- VS Codium Continue agent tool errors vanish after upgrading Ubuntu user to Administrator — nikhilb_it · 2026-09-04