HF Staff Open-Sources Self-Iterating 3D Asset Prompt With Vision Critique Loops for Small Quantized Models
mervenoyann · x · 2026-10-21
Hugging Face staff member Merve Noyan shares a hard problem for small open models: acting as a self-iterator that generates three.js 3D assets and critiques its own output — especially brutal below 4-bit quantization.
She open-sourced the full prompt (pi-3d-prompt.txt) on HF. Key engineering constraints:
- Build a reusable Three.js asset factory, kept separate from scene, materials, cameras, capture and vision-critique modules
- Force five genuine improvement rounds (00–05), each based on the previous accepted snapshot, archiving HTML, source, screenshots, critic requests and changes.md — no overwriting accepted versions
- Ten named screenshots (six orthographic, two perspective, two detail), cameras calibrated then frozen, real browser PNGs at ≥1280x800
- Since Pi is text-only, images are inspected via a separate vision request to the same llama server, verified first with a random color-panel test
Her takeaway: small models write the app code fine, but asset generation and self-critique iteration is where they fall apart — this prompt scaffolding is how she makes it work on high-end consumer hardware.
More from coding & agent
- How do you evaluate GenAI output quality before shipping? Devs weigh in — ssnuopi · 2026-10-21
- RunInfra launches $10/mo coding plan: fastest open models across 19 agents incl. Claude Code — ycombinator · 2026-10-21
- Matt Shumer launches an 'AI employee' you assign work to over email — mattshumer_ · 2026-10-21
- uv-ty integration: how to enable PEP 723 script type checking — charliermarsh · 2026-10-21
- Astral ships experimental uv-ty integration: uv check type-checks your project — charliermarsh · 2026-10-21
- Dev's self-built tracker claims 7.1x speedup with Claude, 2.5 work-years banked — arthurcolle · 2026-10-21