A group of researchers has published PhysStream on arXiv — an autoregressive model for image-to-video generation that allows controlling object motion directly during generation. Instead of a full motion plan before the start or per-pixel signals, the model uses sparse velocity increment signals encoding physical quantities, and structural scene memory from positional maps and object tracking maps. Details are on
the paper's page on arXiv. The authors did not release any code.
What it means
PhysStream is a research work, and our reference does not yet have a tool with this approach. The closest video generators on the topic from the
AI Video section work differently.
Maestro is a local studio based on the WanGP pipeline, with an image-text-to-video task. The MiniMax H3 weights take up 464.1 GB, the largest file is 9.7 GB, and 12—24 GB of VRAM is required. The code is open source, but the WanGP license permits only non-commercial use of the program itself; the MiniMax H3 weights are available for commercial use up to $20 million in revenue, but the license does not apply in the EU, the UK, Korea and the US. Interactive motion control in the middle of generation is not in the fact sheet.
LTX-Video from Lightricks handles the image-to-video task, weighs 236.4 GB, the largest file is 26.6 GB, and only 1 GB of VRAM is stated. The code is under Apache-2.0, the weights are under the LTXV Open Weights License with a commercial use threshold of $10 million in annual revenue. Motion physics control is not mentioned in the fact sheet.
CogVideoX is a text-to-video model from Zhipu AI, weights 20 GB, the largest file is 9.2 GB, runs on CUDA. The code is under Apache-2.0, the weights are under The CogVideoX License: commercial use after registration, free up to 1 million visits per month. There is no physically grounded motion control in the description.
PhysStream is presented as a step from text prompts to controlling the physical dynamics of a scene, but without published code and weights it cannot be tested on your own hardware. For local video generation with motion control, the tools from our
AI Video section, which do not have such a mode in their fact sheets, remain the only option for now.
