The Snap Research team has announced its participation in several major industry conferences and events in 2025, where it will present advancements in augmented reality (AR), generative artificial intelligence (AI), recommendation systems, and creative tools.
At SIGGRAPH 2025 in Vancouver, the team presented new methods for improving image generation and personalization. One of these is Nested Attention, which enhances identity preservation in image generation models by using a semantic-aware attention structure. This approach allows for more consistent images of specific subjects across various styles and scenes. Another project, InstantRestore, offers a single-step method for restoring degraded face images while retaining unique identity features.
The Set-and-Sequence framework addresses video generation with dynamic concepts—subjects defined by both appearance and motion patterns—enabling realistic personalization of videos that feature elements like ocean waves or flickering bonfires. DuetGen focuses on generating synchronized two-person dance motions from music input, supporting applications such as animation and digital performance.
Be Decisive introduces a neural network that predicts spatial layouts during the denoising process of image generation to accurately depict multiple distinct subjects within complex images.
At KDD 2025 in Toronto, Snap will highlight GiGL, an open-source library designed to train Graph Neural Networks (GNNs) at scale for applications including user growth and content ranking at Snapchat. The PRISM method replaces embedding weight decay with a simpler computation at training onset to improve efficiency in recommendation systems.
AutoCDSR aims to enhance cross-domain sequential recommendation by improving knowledge sharing between domains while reducing noise. SnapGen is a text-to-image model optimized for mobile devices that can generate high-quality images quickly without heavy computing resources. Its extension, SnapGen-V, enables five-second video creation on mobile devices within five seconds.
4Real-Video generates detailed 4D videos viewable from multiple angles for use cases such as immersive virtual reality experiences. Stable Flow allows users to edit photos by adding or removing objects without complex training or hardware requirements.
Omni-ID creates holistic representations of faces across different angles and expressions to support more realistic AI-generated imagery. PrEditor3D streamlines 3D model editing with minimal input, making it easier for creators to develop AR content efficiently.
MM-Graph introduces a benchmark combining visual and textual data for evaluating multimodal graph learning models. Video Alchemist enables video generation from text prompts and reference images without extensive tuning.
Mind the Time gives creators control over the timing of events within AI-generated videos to support structured storytelling. Video Motion Transfer uses diffusion models to transfer motion between videos easily.
Wonderland constructs detailed 3D scenes from single photographs, speeding up design processes without requiring many resources or viewpoints. AC3D improves camera movement control within video diffusion transformers for more realistic generated scenes.
According to Snap Research: “All models and work outlined here is for research purposes only.”
