Why the ICCV 2019 'draw via actions + renderer' idea explains visual agents

ziqiao_ma · x · 2026-09-07

Prompted by GPT-6 Astra's near-production Blender work, the author revisits a classic ICCV 2019 paper whose idea still holds: create through sequential actions plus a renderer rather than predicting pixels directly. Back then: parameterized Bézier strokes with a learned neural renderer; today: Blender code controlling geometry, materials, lighting and cameras. Key points: pixels are only final observations and collapse structure; graphics programs keep structure explicit while rendering gives a forward model for evaluation and refinement; the hard problem may be search and control before rasterization, not image synthesis itself.

Original post →

More from coding & agent

coding & agent channel →