trackdiffusion_ytvis

TrackDiffusion Model Card

<!-- Provide a quick summary of what the model is/does. -->

TrackDiffusion is a diffusion model that takes in tracklets as conditions, and generates a video from it.

Model Details

Model Description

TrackDiffusion is a novel video generation framework that enables fine-grained control over complex dynamics in video synthesis by conditioning the generation process on object trajectories. This approach allows for precise manipulation of object trajectories and interactions, addressing the challenges of managing appearance, disappearance, scale changes, and ensuring consistency across frames.

Uses

Direct Use

We provide the weights for the entire unet, so you can replace it in diffusers pipeline, for example:

pretrained_model_path = "stabilityai/stable-video-diffusion-img2vid"
unet = UNetSpatioTemporalConditionModel.from_pretrained("/path/to/unet", torch_dtype=torch.float16,)
pipe = StableVideoDiffusionPipeline.from_pretrained(
    pretrained_model_path, 
    unet=unet,
    torch_dtype=torch.float16,
    variant="fp16",
    low_cpu_mem_usage=True)