Kaede Shiohara

Autoregressive latent priors learned directly from data, not fixed Gaussians, unlock VAEs for high-fidelity sequential generation with diffusion-level quality and faster inference.

GUSH3R achieves real-time dynamic human-scene reconstruction with photorealistic quality, outperforming traditional optimization methods in efficiency.

Wild3R achieves competitive performance against traditional optimization methods while eliminating the need for time-consuming scene-specific adjustments.

A single transformer model now beats specialized architectures at generating floorplans from diverse inputs, thanks to a unified markup language.