Ever wish you could turn your video generator into a controllable physics simulator? . We're thrilled to introduce Force Prompting! Animate any image with physical forces and get fine-grained control, without needing any physics simulator or 3D assets at inference. 🧵(1/n)

RT @davisblalock: Deep learning training is a mathematical dumpster fire. But it turns out that if you *fix* the math, everything kinda ju….

RT @Koven_Yu: 🤩 WorldScore is accepted to #ICCV2025. Benchmark your 3D/4D/video world models with WorldScore!.

RT @Koven_Yu: #ICCV2025 🤩3D world generation is cool, but it is cooler to play with the worlds using 3D actions 👆💨, and see what happens! —….

RT @frankzydou: Check out 🌟Vid2Sim: Generalizable, Video-based Reconstruction of Appearance, Geometry & Physics for Mesh-Free Simulation #C….

RT @zhang_yunzhi: (1/n) Time to unify your favorite visual generative models, VLMs, and simulators for controllable visual generation—Intro….

RT @sainingxie: Had a great time at this CVPR community-building workshop---lots of fun discussions and some really important insights for….

RT @ShijieWang20: I'm in #CVPR2025!. Fri, 13 Jun, 4-6 PM CAT, poster session 2. At ExHall D Poster #230. come and have a chat about our wor….

RT @openworldlabs: In this blog post we will summarize some of our findings with training autoencoders for diffusion! We also share some nu….

RT @drsrinathsridha: Existing 3D human manipulation datasets are valuable, but are limited in scale and diversity. At #CVPR2025, we will in….

alternative/better paper title: "Use The Force!" (thanks @danbgoldman)

This was a collaboration between @BrownCSDept and @GoogleDeepMind, joint work with @mik3fr33man @dakshces @DeqingSun @jesu9 + two other collaborators who aren't on X yet, Charles and Evan!.

We've released all datasets, code, and model weights to foster further research on interactive and physically plausible video generation models. Explore more on our project page including fun interactive demos like this 👶 (n/n)

Force prompting can recreate some demos for prior works that use a physics simulator at inference! Our method doesn’t require any 3D asset, or simulator, at inference time–just upload your image, and choose wind force/angle, or poke-ing location and force/angle. (7/n)

Beyond mass, Force Prompting also seems to understand physical affordances (e.g. a train follows its track when pushed) and object atomicity (poke any part of a train, and the whole train moves). This opens up exciting avenues for interactive world models. (6/n)

Two important design choices we uncovered during our ablation studies: use of motion-specific keywords during training (e.g. "wind"/"blow"/"breeze"), as well as strategic diversity of synthetic training data (e.g. multiple backgrounds, balls/flags with different colors) (5/n)

And we've seen some emergent "intuitive physics" too! Our model demonstrates a degree of mass understanding, where a lighter object will move farther than a heavier one under the same force. Check out the examples! 👇 (4/n)

Force Prompting offers diverse control! One model handles localized point forces (like a precise poke 👆) and another handles global wind force fields (like wind sweeping across a scene 💨). Full control, right in your hands. (3/n)

How does it work? We found that even with limited but strategically generated synthetic training data (think: Blender videos of flags + rolling balls) video generation models can learn to generalize and apply physical forces across diverse scenes. It's surprisingly robust! (2/n)