Jinchao Zhang

RAVA reveals that incorporating geometric evidence can drastically enhance viewpoint alignment in image generation, outperforming traditional methods that rely solely on semantic embeddings.

SketchSong achieves superior song coherence and richness by explicitly planning arrangements before audio generation, outperforming strong post-trained models without extra optimization.

Diffusion models can bridge the semantic gap between abstract action labels and complex video content, leading to state-of-the-art performance in open-vocabulary temporal action detection.

Domain shift in open-vocabulary object detection is tamed with a parameter-efficient semantic augmentation framework that progressively enriches textual representations from coarse to fine-grained levels.

Ditch blurry FID scores: evaluating generative models directly in discrete token space reveals perceptual quality nuances that continuous feature metrics miss.