PhysStream Brings Physics-Grounded Mid-Generation Control to Streaming Video
Chuhao Chen · hf · 2026-09-16
PhysStream is an autoregressive image-to-video model enabling interactive, mid-generation control over multi-object rigid-body scenes — a capability prior methods lack.
- Structured scene memory: positional and object-tracking maps derived online from generated frames maintain physical consistency.
- Motion control: sparse velocity-increment signals encode physical quantities so the model learns underlying dynamics rather than dictating pixel positions.
- Two-stage training: a bidirectional model finetuned with motion control, then a causal autoregressive model with scene memory.
- Results: reduces FVMD by 33% and trajectory error by 12% over strongest baselines; preferred by human evaluators in over 85% of in-the-wild comparisons.
More from Multimodal
- Pixio launches AE plugin: chat agent builds native layers and keyframes in your timeline — tsi_org · 2026-09-16
- Midjourney style ref + Gemini photorealism: a two-step image workflow — michaelrabone · 2026-09-16
- Seedance 2.5 recreates a GTA-style stealth mission with flawless phone tracking — SimplyAnnisa · 2026-09-16
- Nex-N2.5 Pro builds an interactive 3D mechanical garden by iterating on its own output — nikola_mr64990 · 2026-09-16
- Sora's Diffusion Transformer explained by hand in 14 steps: prompts enter only as scale and shift — ProfTomYeh · 2026-09-16
- ElevenLabs in one week: UMusic deal, Music v2.5, MCP multimodal, $1M AI SDR — lukeharries · 2026-09-16