Research.
We publish notes and blogs that improve inside real production workflows.
From Minutes to Seconds: Making 4D Generation Practical for Physical AI
Turbo4DGen removes redundant attention work to accelerate multi-view 4D generation by 9.7x, with 24.2% lower peak memory and broadly preserved visual quality.
EAServe: Up to 4.3x Higher Goodput for Multimodal LLM Serving
EAServe turns the multimodal encoder into the pipeline's control point, increasing SLO-compliant goodput while reclaiming idle GPU capacity.
GaussianSpa: Compact and High-Quality 3D Reconstruction through Constrained Optimization
GaussianSpa makes the target Gaussian count part of training, delivering up to 10.9x fewer primitives while matching or improving rendering quality.