Yu-Lun Liu

Static 3D worlds now possess vivid, looping motion, seamlessly transitioning from any viewpoint without explicit motion capture.

Flowcharts break silently across aspect ratios, but a new agentic pipeline with VLM critics now guarantees connectivity and editability.

Most video generation models fail to accurately render fine-grained, program-specified world events, with PROWBench revealing significant discrepancies between intended logic and visual output.

Online 3D reconstruction failure on long videos is not a representation collapse but an artifact of single-anchor pose extrapolation—and querying relative poses across multiple keyframes fixes it with just 1% parameter overhead.

LATCH accelerates diffusion language model decoding by up to 17.8x without compromising accuracy, redefining how we approach early exits in generative tasks.

LongE2V achieves unprecedented temporal coherence and zero-shot generalization in video reconstruction from sparse event data, outperforming all existing methods.

Achieving dual-semantic 3D visual illusions in mere minutes, JanusMesh outshines traditional methods in both speed and quality.

By merging physical modeling with generative priors, BRDFusion achieves unprecedented control and quality in urban scene rendering, setting a new standard for inverse rendering applications.

Reroute reveals that visual token importance is dynamic, allowing for improved grounding in VLMs without sacrificing performance through aggressive token reduction.

Benchmarking reveals that a diffusion-based 2D enhancer can significantly elevate the performance of distractor-free radiance field methods, achieving nearly 1 dB PSNR improvement.

Training generative inverse renderers on AAA game footage closes the realism gap, enabling superior generalization and controllable generation compared to models trained on existing synthetic datasets.