Generative Video Model Predicts Stress, Strain and Fracture, Generalizes to Unseen Objects
anand_bhattad · x · 2026-10-12
Part 6 of a research thread. After training on a limited set of objects, the model generalizes to new, unseen objects with different appearances and material properties. Demos: cake and bread pulled vertically vs. horizontally, where changing loading conditions changes the direction of separation; rubber toys stretch far while remaining connected. The model jointly generates stress, strain, flow and damage maps, letting the authors inspect the predicted physical state as the visible deformation unfolds.
Training (from part 5): simulated videos and physical maps are encoded and noised, with the input image plus loading, material and fracture conditions; following the controlled video generation recipe, Wan is adapted with LoRA to predict video and physical maps jointly, plus a physics-consistency loss relating strain, stress and more.
Related event: JHU Releases FracGen: Physics-Realistic Fracture Video Generation(2 posts)→
More from Multimodal
- vivo and Xidian Propose StereoBind for Spatially Consistent Stereo Audio in Video Generation — Hanmo Chen · 2026-10-12
- MiniMax voice cloning trouble: cloned character babbles and speaks too early — Nokersoek · 2026-10-12
- Dev building an AI DJ desktop app that matches BPM/key and fixes AI music drift — AIandDesign · 2026-10-12
- AI-generated thriller made with Seedance 2.5 on Pixio, no camera or crew, prompts public — tsi_org · 2026-10-12
- Code Video Model uses code as the engine for controllable, physics-grounded video generation — angrypenguinPNG · 2026-10-12
- Have US labs given up on video generation? Reddit debate points to Seedance — Appropriate_Bend_602 · 2026-10-12