Flow directly from source to target images for efficient inversion-free instructional image editing.

A discrete representation autoencoder with hierarchical vision alignment for generative image tokenization.

Generate network weights via a diffusion model, which parameterize an implicit neural representation and self-decode to synthesize a video.

Single-stage one-shot video motion customization via constraining within temporal-dominant early denoising timesteps.

A zero-shot, scene graph-based conditioning mechanism for compositional text-to-image generation with soft visual guidance.

A holistic video tokenizer that integrates a light weight autoregressive transformer to align the latent space with AR generative models.

Customize pre-trained video diffusion models from a single reference video and adapt its motion to new subjects and scenes with temporal diversity.

Edit the color tone of an image with the editing styles spatially adaptive to its content.

A video implicit neural representation that encodes large-scale and diverse videos faithfully and efficiently.

A frame-wise video implicit neural representation that encodes a video into a dedicated neural network's parameters

A two-stage MMI2I method via style encoder pre-training for simplifed losses and stablized training.