BulletTime ablation: Time-RoPE alone beats RoPE + AdaLN on PSNR, zero new parameters
GordonWetzstein · x · 2026-09-09
A surprising finding from BulletTime: changing positional encoding can outperform adding learned conditioning. In time-control ablations, Time-RoPE alone beat standard RoPE + AdaLN on PSNR with no learnable parameters; combining both worked best. BulletTime also decouples frame index from world time for explicit time/camera control.
Related event: Stanford's BulletTime Decouples Frame Index from World Time(2 posts)→
More from Multimodal
- Horror video prompt for Seedance 2.5: the scariest customer wasn't even alive — umesh_ai · 2026-09-09
- NVIDIA Cosmos powers 4 of top 5 winners in ECCV AI City Challenge forecasting track — NVIDIAAI · 2026-09-09
- Astra, Opus and Fable near completion of their collaborative pixel-art short film — RileyRalmuto · 2026-09-09
- GPT-6 Astra Tops Design Arena's 3D Leaderboard With Elo 1495, 65 Pts Ahead of Kimi K3 — airesearch12 · 2026-09-09
- OpenAI's GPT-Image-2.5 Sweeps All Three LMArena Image Leaderboards, +81 Pts in Multi-Image Edit — airesearch12 · 2026-09-09
- Gradium Tops Voice Arena TTS Latency Benchmarks With 231ms Time-to-First-Audio — rohanpaul_ai · 2026-09-09