Overview of our VidClothEditor framework. First, we preprocess the reference video and the target upper garment through segmentation and pose detection. This process obtains a clothing-agnostic sequence and a pose sequence. We apply full-body masking and obtain the lower garment image from the first frame. Next, the target upper garment and the extracted lower garment are inputted into the GarmentNet to extract hierarchical features. Finally, the UNet incorporate aligned garments features and the conditional sequence into the denosing progress to produce vivid video try-on results.

Overview of the region-guidance multiple garment fusion. We use (a) spatial attention to integrate garment features into the denoising UNet. To prevent feature interference, we employ (b) region guidance to properly align the attention maps.

Overview of the video consistency learning. Given a source video \(z_0\) in the latent space, we perform a forward SDE to add noise on it, resulting in \(z_{t_{n+1}}\). \(z_{t_n}\) is the one-step estimation from \(z_{t_{n+1}}\), which is obtained by the garment-augmented PF-ODE solver. \(z_{t_{n+1}}\) and \(z_{t_n}\) should be enforced to satisfy the self-consistency property.