LTX-Video and CogVideoX Still Don't Solve Multi-Object Generation in a Single Video

A comparative study of three approaches to multi-object video generation without fine-tuning has been published on arXiv: direct, parallel, and sequential generation. The authors analyze how each method handles preserving object appearance, motion consistency, and interactions between objects, and describe characteristic failures. The authors did not release the code. More details are in the paper on arXiv.

What it means

In our reference on video generators, there are three tools with the image-to-video or text-to-video task, but none of them claims support for multiple objects in a single clip. LTX-Video handles the image-to-video task and works with a single reference image. As described by the author, there is no mention of splitting the scene into separate objects or composing multiple videos. Direct generation from the study is closest to what LTX-Video does: the model receives the full image and the prompt as a whole. CogVideoX is listed as text-to-video, although image-to-video is also mentioned in the repository description. There is no information about working with multiple objects or their separate generation in the fact sheet. Pyramid Flow is also text-to-video, and its description says nothing about multi-subject generation or scene decomposition. The study describes paradigms that are not yet implemented in the tools in our reference. The direct approach partially overlaps with LTX-Video’s current capabilities, but parallel and sequential generation are separate techniques not present in the fact sheets. If the authors later release the code, it will be possible to test how applicable these methods are to the open-source models in our catalog.