The Snap Research team is highlighting its latest advances in augmented reality (AR), generative artificial intelligence, recommendation systems, and creative tools. The group’s work will be presented at several major industry conferences in 2025.
At SIGGRAPH 2025 in Vancouver, Canada, the team introduced several new methods and models. “Nested Attention: Semantic-aware Attention Values for Concept Personalization” is a method that improves how image generation models maintain consistent identities of subjects across different styles and scenes. According to the team, this approach enables the creation of personalized images that can combine multiple subjects into one picture.
Another project, “InstantRestore: Single-Step Personalized Face Restoration with Shared-Image Attention,” presents a method for restoring degraded face images using a single step in a diffusion model. The technique focuses on keeping identity-specific features intact during restoration.
“Dynamic concepts personalization from single videos” introduces Set-and-Sequence, a framework that learns motion patterns over time to personalize video content with dynamic subjects like ocean waves or flickering bonfires.
For dance choreography applications, “DuetGen: Music Driven Two-Person Dance Generation via Hierarchical Masked Modeling” offers a system that generates synchronized dance motions for two people directly from music input.
“Be Decisive: Noise-Induced Layouts for Multi-Subject Generation” describes an approach to generate images with multiple distinct subjects without blending errors by refining spatial layouts during image generation.
At KDD 2025 in Toronto, Ontario, Snap Research will showcase further developments. GiGL is an open-source library designed for training large-scale graph neural networks (GNNs) on data sets with hundreds of millions of nodes and billions of edges. It is used at Snap for tasks such as user growth analysis and advertising optimization.
The team also introduced PRISM (Popularity-awaRe Initialization Strategy for embedding Magnitudes), which streamlines recommendation model training by removing the need for embedding weight decay—resulting in faster and more efficient systems.
AutoCDSR is another innovation aimed at improving cross-domain sequential recommendation by enhancing knowledge sharing while reducing irrelevant signals, leading to better personalization accuracy.
SnapGen is a high-performance text-to-image research model optimized to run on mobile devices and capable of generating high-quality images quickly with reduced computational requirements. Its extension, SnapGen-V, brings fast video generation capabilities to mobile devices as well.
4Real-Video enables realistic four-dimensional video creation viewable from multiple angles, potentially benefitting immersive virtual reality experiences and new forms of storytelling.
Stable Flow allows users to edit photos by adding or removing objects without complex training or advanced hardware needs. Omni-ID provides comprehensive facial representations for more realistic AI and AR generations.
PrEditor3D supports quick editing of 3D models with minimal input, aiming to simplify content creation workflows for animators and Lens creators. MM-Graph introduces a benchmark combining visual and textual data for multimodal graph learning evaluations.
Video Alchemist allows users to generate videos from text prompts and reference images without extensive tuning. Mind the Time gives precise control over event timing in AI-generated videos for improved storytelling structure.
Other projects include Video Motion Transfer using diffusion transformers; Wonderland’s single-image-based 3D scene construction; AC3D’s improvements to camera control in video generation; and additional research models designed solely for research purposes within Snap Inc.’s ecosystem.
All these efforts are part of ongoing research activities by Snap Research focused on advancing AR, AI-powered personalization, content creation tools, and machine learning infrastructure.
