王通 | Tong Wang

Reference-guided video editing uses a source video, text instruction, and reference image. We observe that VLM layers encode complementary information hierarchically — early layers capturing spatial details and deeper layers encoding global semantics. MiVE repurposes VLMs as multiscale feature extractors, integrating hierarchical features from Qwen3-VL into a unified self-attention Diffusion Transformer, eliminating the modality mismatch inherent in cross-attention designs. MiVE achieves state-of-the-art performance by ranking highest in human preference.

We propose a method that constructs style and glyph prompts directly from the original image without introducing additional encoders. A two-stage training strategy is used: the diffusion transformer is first trained on large-scale self-supervised data and then refined using a small set of paired images. By leveraging the in-context learning capability of the Multi-Modal Diffusion Transformer (MM-DiT), it achieves open-vocabulary and style-consistent text editing with state-of-the-art performance across various languages.

A specialized glyph encoder for scene text editing that addresses the character structure modeling bottleneck in diffusion models. Achieves state-of-the-art generation quality with 18.02% improvement in sentence accuracy and 53.28% reduction in text-region FID.

Graph-structured dense correspondence learning method that improves cross-view matching robustness.