Snap Inc.

Sergey Tulyakov

Jian Ren

Menglei Chai

Aleksei Stoliar

Aliaksandr Siarohin

Hsin-Ying Lee

Kyle Olszewski

Christopher Yale Crutchfield

Andrew Wan

Sergey Demyanov

Sergei Kotcur

Zeng Huang

Willi Menapace

Ivan Skorokhodov

Oleksandr Pyshchenko

Pavel Savchenkov

Maksim Gusarov

Sergei Korolev

Aleksei Podkin

Arnab Ghosh

Roman Furko

Huan Wang

Qing Jin

Ju Hu

Walton Lin

Andrew Cheng-min Lin

Jiayao Yu

Enxu Yan

Chen Cao

Jaewook Chung

Emre Yamangil

Oliver Woodford

Igor Kudriashov

Shubham Vij

Yanping Chen

Eric Buehl

Aymeric Damien

Theresa Barton

Artem Bondich

Richard Leider

Zhenglin Geng

Yinghao Xu

Francesc Xavier Drudis Rius

Angus Kong

Benjamin Dodson

Kavya Venkata Kota Kopparapu

Jordan Hurwitz

Yanyu Li

Daniel Moreno

Dhritiman Sagar

Georgios Evangelidis

Riza Alp Guler

Roman Golobokov

Sergey Smetanin

Yang Wen

Nikita Demidov

Ivan Babanin

Nikita Belosludtcev

Vladimir Gordienko

Timur Zakirov

Dor Ayalon

Denys Makoviichuk

Vladislav Shakhrai

Ligong Han

Aleksandr Zakharov

Sofya Savinova

Tsai-Shien Chen

Evangelos Ntavelis

Pavlo Chemerys

Colin Eles

Ergeta Muca

Yuwei Fang

Geng Yuan

Zezhou Cheng

Shervin Minaee

Francesco Barbieri

Yash Mukund Kant

Chieh Lin

Peihao Zhu

Rameen Abdal

Zhengfei Kuang

Junli Cao

Panagiotis Achlioptas

Hierarchical patch-wise diffusion models (HPDMs) use a diffusion paradigm that learns a hierarchical distribution of patches instead of whole videos for efficient patch-wise training of diffusion models. To enforce consistency between the patches, deep context fusion may be used to propagate the context information from low-scale to high-scale patches in a hierarchical manner. To accelerate patch-wise training and inference, adaptive computation also may be used to allocate more computational resources and network capacity towards coarse image details and to cheapen synthesis of high-frequency texture details. All the processing stages are jointly trained to provide spatially aligned global context to the higher levels of the cascade. As a result, the model does not operate on the full-resolution inputs, which allows the model to be trained on high-resolution video datasets in an end-to-end fashion.

Hierarchical patch-wise diffusion models for high-resolution video generation

Systems and methods for generating static and articulated 3D assets are provided that include a 3D autodecoder at their core. The 3D autodecoder framework embeds properties learned from the target dataset in the latent space, which can then be decoded into a volumetric representation for rendering view-consistent appearance and geometry. The appropriate intermediate volumetric latent space is then identified and robust normalization and de-normalization operations are implemented to learn a 3D diffusion from 2D images or monocular videos of rigid or articulated objects. The methods are flexible enough to use either existing camera supervision or no camera information at all—instead efficiently learning the camera information during training. The generated results are shown to outperform state-of-the-art alternatives on various benchmark datasets and metrics, including multi-view image datasets of synthetic objects, real in-the-wild videos of moving people, and a large-scale, real video dataset of static objects.

Autodecoding latent 3D diffusion models

Automatic captioning pipelines and methods for automatically annotating video data with subtitles, which can be obtained using automatic speech recognition (ASR). An automatic captioning pipeline with inputs of multimodal data scales up the dataset of high-quality video-caption pairs. The automatic captioning pipeline generates video-caption pairs by establishing and using a large video-language dataset along with an automatic captioning approach leveraging multimodal inputs, such as textual video description, subtitles, and individual video frames.

Captioning videos with multiple cross-modality teachers

A neural light field (NeLF) that runs real-time on mobile devices for neural rendering of three dimensional (3D) scenes, referred to as MobileR2L. The MobileR2L architecture runs efficiently on mobile devices with low latency and small size, and it achieves high-resolution generation while maintaining real-time inference for both synthetic and real-world 3D scenes on mobile devices. The MobileR2L has a network backbone including a convolutional layer embedding an input image at a resolution, residual blocks uploading the embedded image, and super-resolution modules receiving the uploaded embedded image and rendering an output image having a higher resolution than the embedded image. The convolution layer generates a number of rays equal to a number of pixels in the input image, where a partial number of the rays is uploaded to the super-resolution modules.

Real-time neural light field on mobile devices

A 3D-aware generative model for high-quality and controllable scene synthesis uses an abstract object-level representation (i.e., 3D bounding boxes without semantic annotation) as the scene layout prior, which is simple to obtain, general to describe various scene contents, and yet informative to disentangle objects and background. An overall layout for the scene is identified and then each object is located in the layout to facilitate the scene composition process. The object-level representation serves as an intuitive user control for scene editing. Based on such a prior, the system spatially disentangles the whole scene into object-centric generative radiance fields by learning on only 2D images with global-local discrimination. Once the model is trained, users can generate and edit a scene by explicitly controlling the camera and the layout of objects' bounding boxes.

Spatially disentangled generative radiance fields for controllable 3D-aware scene synthesis

An environment synthesis framework generates virtual environments from a synthesized two-dimensional (2D) satellite map of a geographic area, a three-dimensional (3D) voxel environment, and a voxel-based neural rendering framework. In an example implementation, the synthesized 2D satellite map is generated by a map synthesis generative adversarial network (GAN) which is trained using sample city datasets. The multi-stage framework lifts the 2D map into a set of 3D octrees, generates an octree-based 3D voxel environment, and then converts it into a texturized 3D virtual environment using a neural rendering GAN and a set of pseudo ground truth images. The resulting 3D virtual environment is texturized, lifelike, editable, traversable in virtual reality (VR) and augmented reality (AR) experiences, and very large in scale.

Infinite-scale city synthesis

Described is a system for improving machine learning models. In some cases, the system improves such models by identifying a performance characteristic for machine learning model blocks in an iterative denoising process of a machine learning model, connecting a prior machine learning model block with a subsequent machine learning model block of the machine learning model blocks within the machine learning model based on the identified performance characteristic, identifying a prompt of a user, the prompt indicative of an intent of the user for generative images, and analyzing data corresponding to the prompt using the machine learning model to generate one or more images, the machine learning model trained to generate images based on data corresponding to prompts.

Text-to-image diffusion model rearchitecture

Methods and systems are disclosed for performing operations for generating a 3D model of a scene. The operations include: receiving a set of two-dimensional (2D) images representing a first view of a real-world environment; applying a machine learning model comprising a neural light field network to the set of 2D images to predict pixel values of a target image representing a second view of the real-world environment, the machine learning model being trained to map a ray origin and direction directly to a given pixel value; and generating a three-dimensional (3D) model of the real-world environment based on the set of 2D images and the predicted target image.

3D modeling based on neural light field

A method of generating an image for use in a conversation taking place in a messaging application is disclosed. Conversation input text is received from a user of a portable device that includes a display. Model input text is generated from the conversation input text, which is processed with a text-to-image model to generate an image based on the model input text. The coordinates of a face in the image are determined, and the face of the user or another person is added to the image at the location. The final image is displayed on the portable device, and user input is received to transmit the image to a remote recipient.

Text-guided cameo generation

Invertible Neural Networks (INNs) are used to build an Invertible Neural Skinning (INS) pipeline for reposing characters during animation. A Pose-conditioned Invertible Network (PIN) is built to learn pose-conditioned deformations. The end-to-end Invertible Neural Skinning (INS) pipeline is produced by placing two PINs around a differentiable Linear Blend Skinning (LBS) module using a pose-free canonical representation. The PINs help capture the non-linear surface deformations of clothes across poses and alleviate the volume loss suffered from the LBS operation. Since the canonical representation remains pose-free, the expensive mesh extraction is performed exactly once, and the mesh is reposed by warping it with the learned LBS during an inverse pass through the INS pipeline.

Invertible neural skinning

Unsupervised volumetric 3D animation (UVA) of non-rigid deformable objects without annotations learns the 3D structure and dynamics of objects solely from single-view red/green/blue (RGB) videos and decomposes the single-view RGB videos into semantically meaningful parts that can be tracked and animated. Using a 3D autodecoder framework, paired with a keypoint estimator via a differentiable perspective-n-point (PnP) algorithm, the UVA model learns the underlying object 3D geometry and parts decomposition in an entirely unsupervised manner from still or video images. This allows the UVA model to perform 3D segmentation, 3D keypoint estimation, novel view synthesis, and animation. The UVA model can obtain animatable 3D objects from a single or a few images. The UVA method also features a space in which all objects are represented in their canonical, animation-ready form. Applications include the creation of lenses from images or videos for social media applications.

Unsupervised volumetric animation

A framework trains game-engine-like neural models from annotated videos to generate a Learnable Game Engine (LGE) that maintains states of the scene, objects and agents in it, and enables rendering the environment from a controllable viewpoint. The LGE models the logic of the game and the rules of physics, making it possible for the user to play the game by specifying both high- and low-level action sequences. The LGE also unlocks a director's mode where the game is played by plotting behind the scenes, specifying high-level actions and goals for the agents using text-based instructions. To implement the director's mode, a trained diffusion-based animation model navigates the scene using high-level constraints, to enable play against an adversary, and to devise the strategy to win a point. To render the resulting state of the environment and its agents, a compositional neural radiance field (NeRF) representation is used in a synthesis model.

Plotting behind the scenes with learnable game engines

A three-dimensional (3D) scene is generated from non-aligned generic camera priors by producing a tri-plane representation for an input scene received in random latent code, obtaining a camera posterior including posterior parameters representing color and density data from the random latent code and from generic camera priors without alignment assumptions, and volumetrically rendering an image of the input scene from the color and density data to provide a scene having pixel colors and depth values from an arbitrary camera viewpoint. A depth adaptor processes depth values to generate an adapted depth map that bridges domains of rendered and estimated depth maps for the image of the input scene. The adapted depth map, color data, and scene geometry information from an external dataset are provided to a discriminator for selection of a 3D representation of the input scene.

3D generation of diverse categories and scenes

A system to enable 3D hair reconstruction and rendering from a single reference image which performs a multi-stage process that utilizes both a 3D implicit representation and a 2D parametric embedding space.

Single image three-dimensional hair reconstruction

A multimodal video generation framework (MMVID) that benefits from text and images provided jointly or separately as input. Quantized representations of videos are utilized with a bidirectional transformer with multiple modalities as inputs to predict a discrete video representation. A new video token trained with self-learning and an improved mask-prediction algorithm for sampling video tokens is used to improve video quality and consistency. Text augmentation is utilized to improve the robustness of the textual representation and diversity of generated videos. The framework incorporates various visual modalities, such as segmentation masks, drawings, and partially occluded images. In addition, the MMVID extracts visual information as suggested by a textual prompt.

Video synthesis via multimodal conditioning

Growing community of inventors

Marina Del Rey, CA, United States of America

Sergey Tulyakov