Fluid-Gen-Zero wires physics simulators into frozen video models, cutting trajectory error ~81.5%
jiqizhixin · x · 2026-11-12
Fluid-Gen-Zero, a preprint from Simon Fraser University, Stony Brook University, Lawrence Berkeley National Laboratory, and Meta Reality Labs, injects simulator-driven motion into frozen video generators without retraining.
- Reconstruction & simulation: a reconstructed scene supplies fluid and object dynamics via a physical solver.
- Generation-time planning: a vision-language planner splits generation clips around key events (impact, splash) so physical moments aren't cut across segments.
- Region-aware latent guidance: strong constraints on object structure, softer control on fluid appearance. On a 43-sample benchmark, adding the method to Tora reduces object trajectory error from 133.20 to 24.61 (81.5%).
The authors note this measures agreement with simulation, not a general guarantee of physical correctness.
More from Multimodal
- MiniMax H3-Image merges up to 16 reference images into one concept scene — mhdfaran · 2026-11-12
- MOSS Image with 4-step LoRA runs on-device on iPhone: ~30s for 1024x1024 — woadwarrior · 2026-11-12
- Massive Gemini prompt specs a deterministic 3-minute looping violin concert film — CurieuxExplorer · 2026-11-12
- Seedream 5.5 launches on fal with multi-turn editing, in-image text edits, and layer separation — Jamez_Chi · 2026-11-12
- Runway adds 3D: text-to-3D via Tripo H3.1 with video workflow integration — runwayml · 2026-11-12
- AI artwork "Nox in Oculis" painted entirely in Python with Fae × Fable 5.2 — amplifiedamp · 2026-11-12