#generative model

EdgeConnect uses a two-stage adversarial model for image inpainting, focusing on the accurate reproduction of image details by first generating edges and then using an inpaint network. Utilizing datasets like Places2, CelebA, and Paris Street-View, and built with Python and PyTorch, this model reconstructs missing image sections. It supports different levels of training and testing, provides pre-trained models, detailed configuration guides, and includes alternative edge detection, serving both researchers and developers interested in advanced inpainting methods.

Explore SEINE, a short-to-long video diffusion model integral to the Vchitect system, leveraging Stable Diffusion v1.4. It offers transformations of images into videos, supporting I2V and video transition tasks. The platform includes a Text-to-Video feature with customizable settings, suitable for creative animations. SEINE is available for research and commercial use under the Apache-2.0 license.

Discover a cutting-edge approach to conditional 3D image synthesis, facilitating realistic image creation from 2D label maps with neural radiance fields. The model enables control through multiple viewpoints and provides a real-time editing interactive system. Utilizing widely-available monocular images and label maps, labels are assigned to 3D points for simultaneous image and label map rendering. This PyTorch-based implementation is compatible with multiple datasets, ensuring easy integration with pre-trained models and detailed guidance for data preparation and training, proving valuable for researchers and developers in 3D image synthesis.

The MM-Interleaved model is a pioneer in interleaved image-text generative modeling, featuring a multi-modal feature synchronizer for high-resolution recognition. It supports tasks like visual storytelling, visual question answering, and text-to-image generation. With zero-shot and finetuning capabilities, it offers excellent performance across multiple benchmarks. Access pretrained models for versatile application adaptation.

ThemeStation offers a sophisticated method for creating 3D assets that maintain theme coherence and diversity, utilizing a two-stage framework and dual score distillation. It supports 3D-to-3D applications with minimal examples, with promising results from experiments and user studies. Learn more about its latest innovations.