Tesla FSD visualizations revealed as generative model outputting final 2D frames directly
gandamu_ml · x · 2026-11-22
- Nikharris explains Tesla FSD replay visualizations: camera frames in, pixels out — the model handles reconstruction, view orientation, fusing the car in, and rendering chevrons, which is why it's cheaper computationally than the old visualization stack.
- Gandamuml refines the claim: the model directly outputs final 2D frames and never outputs 3D geometry; 3D understanding lives only in intermediate activations.
- He speculates the annotated elements were trained in directly, and that navigation/driving intelligence may contribute extra inputs beyond raw camera views.
More from Embodied
- Tesla FSD XV: wrong-asset training renders reportedly fixed — yunta_tsai · 2026-11-22
- Tesla camera shows surprisingly strong low-light signal, developer says — yunta_tsai · 2026-11-22
- Winning fight robot gets kissed and its pilot signs autographs: "that's a real sport" — cixliv · 2026-11-22
- AMD owns the SC26 floor: Azure HDv2 on 500-core Venice EPYC, Helios open fabric — AccBalanced · 2026-11-22
- Tesla's real-time viz fulfills 4-year-old promise: 'like looking outside' — pduan · 2026-11-22
- $48B robotics boom has a hidden risk: 90%+ of rare earth magnets come from China — MarwaEldiwiny · 2026-11-22