Light passes through the Magic Mirror to create a personalized virtual world~ It generates identity-preserved videos from reference images using a conditional adaptive normalization module for faster convergence in video Diffusion Transformers.

This work proposes a light-weight controllable module for various base models (SD1.5, SDXL, SD3, SVD) and tasks (image / video generation with various conditions).

Mining potential of open-source VLMs! Mini-Gemini is a novel framework ranges from 2B to 34B VLMs for hi-resolution image understanding. It has a impressive OCR capability, and can generate HQ images powered by its multi-modal reasoning ability.

Control text generation by highlighting our prompt! Prompt Highlighter is a training-free inference pipeline, which facilitates token-level user interactions for customized generation. Our method is compatible for both LLMs and VLMs.

Given an image as the prompt, we can generate its variations by aligning the diffusion inversion chain. The variations are diverse and controllable.

Add 'Lego' attribute to the child, an edited video is generated. Powered by a novel video inversion process and cross-attention control. We also find that a Decoupled-Guidance strategy is essential for video editing.

We present a controllable scene stylization method utilizing radiance fields to stylize a 3D scene, with a single stylized 2D view taken as reference.

A speech input generates an authentic 3D facial animation based on representations with discrete motion prior.

R²Former decomposes the relationships into intra-region ones and inter-region ones. With such modifications, the performance of mask classification can be improved on widely used semantic segmentation benchmarks.

We propose the Continuous Refinement Model (CRM) for the ultra high-resolution mask refinement. CRM aligns the feature with the refinement target and aggregates them to reconstruct image details.

We propose a novel proxy-based contrastive learning method, which replaces the original sample-to-sample relations with proxy-to-sample relations, significantly alleviating the positive alignment issue.