The Snap Research team is highlighting its latest developments in augmented reality (AR), generative artificial intelligence (AI), recommendation systems, and creative tools for 2025. The group is presenting its research at major industry conferences, including SIGGRAPH 2025 in Vancouver and KDD 2025 in Toronto.
At SIGGRAPH 2025, the team introduced several advancements. “Nested Attention: Semantic-aware Attention Values for Concept Personalization” is a method that enhances identity preservation in image generation models by using a semantic-aware attention structure. This allows the model to maintain subject identity across various styles and scenes.
Another project, “InstantRestore: Single-Step Personalized Face Restoration with Shared-Image Attention,” presents a technique for restoring degraded face images efficiently while retaining individual features. The method supports portrait photo enhancement through a single pass of a diffusion model.
The “Set-and-Sequence” framework addresses video generation challenges involving dynamic concepts—entities characterized by unique motion patterns over time. This enables realistic personalization of videos by learning how subjects move across different scenes.
In dance animation, “DuetGen: Music Driven Two-Person Dance Generation via Hierarchical Masked Modeling” generates synchronized dance motions for two people based on music input. It models interactive choreography and physical interactions between dancers for use in animation and digital performance.
The “Be Decisive” project introduces a neural network that refines spatial layouts during image generation to accurately place multiple distinct subjects within complex images, reducing visual inaccuracies or blending.
At KDD 2025, Snap Research presented GiGL, an open-source library designed to train and run Graph Neural Networks (GNNs) on large-scale graphs with hundreds of millions of nodes and billions of edges. GiGL supports machine learning applications at Snap such as user growth, content ranking, and advertising.
Another paper describes PRISM (Popularity-awaRe Initialization Strategy for embedding Magnitudes), which replaces embedding weight decay in recommendation system training with a simpler computation at the start of training, improving efficiency.
AutoCDSR was introduced as a way to improve cross-domain sequential recommendation systems by promoting effective knowledge sharing while minimizing irrelevant signals. This leads to more accurate personalization in recommendations.
SnapGen is described as a high-performance text-to-image model that runs directly on mobile devices and generates high-quality images quickly. Its extension, SnapGen-V, can generate five-second videos on mobile devices within five seconds.
Other projects include 4Real-Video for creating realistic 4D videos viewable from multiple angles; Stable Flow for editing images without complex training or hardware; Omni-ID for building comprehensive facial representations; PrEditor3D for efficient editing of 3D models; MM-Graph as the first benchmark combining visual and textual data in graph learning; Video Alchemist for streamlined personalized video creation; Mind the Time introducing temporal control into AI-generated videos; Video Motion Transfer using diffusion transformers to transfer motion between videos; Wonderland generating detailed 3D scenes from single images; and AC3D improving camera control within video generation models.
All research presented is intended solely for research purposes.
