Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification
Training-free sparse attention for quality-preserving video generation, delivering over 2× end-to-end acceleration and up to 5× with Sol-Engine.
Publications
Publications
Research across foundation models, generative vision, and automated AI systems.
SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation
Open-source 5B/14B hybrid video DiTs for high-quality 720p generation, accelerated to 13.06 seconds on a single H100.
GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation
SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer
An open-source world model for controllable minute-long 720p video generation on a single GPU.
UltraFlux: Data-Model Co-Design for High-quality Native 4K Text-to-Image Generation across Diverse Aspect Ratios
Native 4K text-to-image generation across diverse aspect ratios, matching or surpassing Seedream 4.0.
PosterCraft: Rethinking High-Quality Aesthetic Poster Generation in a Unified Framework
A unified framework for high-quality aesthetic poster generation with precise text rendering, strong layouts, and visual-language refinement.
LucidFlux: Caption-Free Universal Image Restoration via a Large-Scale Diffusion Transformer
A caption-free 14B restoration model delivering photorealistic results beyond leading commercial systems.
GlassWizard: Harvesting Diffusion Priors for Glass Surface Detection
MagicInfinite: Generating Infinite Talking Videos with Your Words and Voice
Technical report for Hedra Inc.’s Character-3 model.
PromptHaze: Prompting Real-world Dehazing via Depth Anything Model
AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement
Learning Diffusion Texture Priors for Image Restoration
Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis
The first SDXL-level high-resolution non-AR T2I model.
