Turning mouse sketches into clean drawings via gestalt-like priors and staged interpretation
debreuil · x · 2026-10-25
debreuil demos a pipeline that converts mouse-tracked drawings into a Gojo-style 'pre-numeric' encoding, uses gestalt-like priors to build a conceptual representation, then 'hallucinates' a clean image from those evidence streams.
The screenshot walks through 7 stages of the same house doodle: fitting raw strokes (5 strokes, 118 samples, 2.80 RMS), carrying long lines and whole curves, reading supported walls/roofs/arches, balancing parts while retaining the drawing, merging nearby joints and trimming overshoots, reading complete arches and a supported side frame, and finally synthesizing the facade while preserving side-view depth (9.94 RMS) — with every intermediate step visible and toggleable.
More from Multimodal
- CapCut plugin launches in early access on ChatGPT desktop, mobile and web due mid-November — Aiden_Tech_Ai · 2026-10-25
- MMGR benchmark: top video models solve only ~3% of mazes despite stunning visuals — zhaofeng_wu · 2026-10-25
- Dressa voice model holds conversations without an LLM, runs locally on pure voice training — TomLikesRobots · 2026-10-25
- Fiery horse artwork shares two-color prompt palette — LudovicCreator · 2026-10-25
- Five days of AI-generated visuals blended with ASCII dither techniques — Teknium · 2026-10-25
- One frame, one prompt: Seedance 2.5 vs Kling 4 vs MiniMax H3 vs Wan 3.0 — LudovicCreator · 2026-10-25