The Snap Research team has announced its participation in several major industry conferences and events in 2025, highlighting advancements in augmented reality (AR), generative artificial intelligence (AI), recommendation systems, and creative tools.
At SIGGRAPH 2025 in Vancouver, Canada, the team presented multiple research papers. One of these was “Nested Attention: Semantic-aware Attention Values for Concept Personalization,” which introduces a method to improve identity preservation in image generation models. The approach allows for more consistent and accurate images of specific subjects across various styles and scenes by using a semantic-aware attention structure.
Another paper, “InstantRestore: Single-Step Personalized Face Restoration with Shared-Image Attention,” details a technique for restoring degraded face images efficiently while retaining identity-specific features. This supports identity-aware restoration for portrait photo enhancement.
The team also introduced “Set-and-Sequence,” a framework aimed at video generation models that can handle dynamic concepts—entities defined by both appearance and unique motion patterns over time. This enables realistic personalization of videos involving subjects like ocean waves or flickering bonfires.
“DuetGen: Music Driven Two-Person Dance Generation via Hierarchical Masked Modeling” addresses generating synchronized two-person dance motions from music input. It is designed to model interactive choreography between dance partners for use in animation, virtual avatars, and digital performances.
In their work titled “Be Decisive: Noise-Induced Layouts for Multi-Subject Generation,” the researchers developed a neural network that predicts spatial layouts during image denoising. This helps accurately generate multiple distinct subjects within complex images without blending errors.
At KDD 2025 in Toronto, Ontario, Canada, Snap Research highlighted additional projects. Among them is GiGL, an open-source library used at Snap Inc. for training graph neural networks on large-scale graphs supporting hundreds of millions of nodes and billions of edges. Applications include user growth analysis, content ranking, and advertising optimization.
The paper “On the Role of Weight Decay in Collaborative Filtering: A Popularity Perspective” introduces PRISM—a strategy that eliminates embedding weight decay during recommendation model training. Instead, it uses a single computation at the start of training to improve efficiency.
“Revisiting Self-Attention for Cross-Domain Sequential Recommendation” presents AutoCDSR, which aims to enhance prediction accuracy across different user interaction domains by promoting effective knowledge sharing while reducing noise from irrelevant signals.
SnapGen is another development—a high-performance text-to-image research model capable of running directly on mobile devices to generate high-quality images quickly with reduced computational demands. Its extension, SnapGen-V, brings similar capabilities to video generation on mobile devices.
Other notable research includes 4Real-Video for generating detailed 4D videos suitable for immersive VR experiences; Stable Flow for easy image editing without complex training; Omni-ID for comprehensive facial representation across angles; PrEditor3D for efficient 3D shape editing; MM-Graph as a benchmark combining visual and textual data in multimodal graph learning; Video Alchemist enabling streamlined personalized video creation; Mind the Time introducing temporal control into AI-generated videos; Video Motion Transfer allowing realistic movement transfer between videos; Wonderland creating detailed 3D scenes from single photos; and AC3D improving camera control within video diffusion transformers.
According to Snap Research, all models and technologies described are intended solely for research purposes at this stage.
