Computer Vision

MagicFace achieves high-fidelity facial expression editing via AU control, preserving identity and background using a diffusion model and ID encoder, significantly outperforming existing methods.

Ingredients: A new framework customizes videos by blending multiple photos with video diffusion transformers, enabling realistic and personalized video generation while maintaining consistent identity…

VideoAnydoor: High-fidelity video object insertion with precise motion control, achieved via an end-to-end framework leveraging an ID extractor and a pixel warper for robust detail preservation and fi…

SeFAR: a novel semi-supervised framework for fine-grained action recognition, achieves state-of-the-art results by using dual-level temporal modeling, moderate temporal perturbation, and adaptive regu…

SeedVR: A novel diffusion transformer revolutionizes generic video restoration by efficiently handling arbitrary video lengths and resolutions, achieving state-of-the-art performance.

LightningDiT resolves the optimization dilemma in latent diffusion models by aligning latent space with pre-trained vision models, achieving state-of-the-art ImageNet 256x256 generation with over 21x …

Zero-shot image safety judgment is achieved using MLLMs and a novel method called CLUE, objectifying safety rules, and significantly reducing the need for human labeling.

VisionReward, a novel reward model, surpasses existing methods by precisely capturing multi-dimensional human preferences for image and video generation, enabling more accurate and stable model optimi…

Edicho: a novel training-free method for consistent image editing across diverse images, achieving precise consistency by leveraging explicit correspondence.

3to4D: Animate any 3D object with text prompts, preserving visual quality and achieving realistic motion!

VideoMaker achieves high-fidelity zero-shot customized video generation by cleverly harnessing the inherent power of video diffusion models, eliminating the need for extra feature extraction and injec…

PartGen generates compositional 3D objects with meaningful parts from text, images, or unstructured 3D data using multi-view diffusion models, enabling flexible 3D part editing.

Orient Anything: Learning robust object orientation estimation directly from rendered 3D models, achieving state-of-the-art accuracy on real images.

DiTCtrl achieves state-of-the-art multi-prompt video generation without retraining by cleverly controlling attention in a diffusion transformer, enabling smooth transitions between video segments.

DepthLab: a novel image-conditioned depth inpainting model enhances downstream 3D tasks by effectively completing partial depth information, showing superior performance and generalization.

SBS Figures creates a massive, high-quality figure QA dataset via a novel stage-by-stage synthesis pipeline, enabling efficient pre-training of visual language models.

Distilled Decoding (DD) drastically speeds up image generation from autoregressive models by using flow matching to enable one-step sampling, achieving significant speedups while maintaining acceptabl…

CLEAR: Conv-Like Linearization boosts pre-trained Diffusion Transformers, achieving 6.3x faster 8K image generation with minimal quality loss.

UIP2P: Unsupervised instruction-based image editing achieves high-fidelity edits by enforcing Cycle Edit Consistency, eliminating the need for ground-truth data.

Boosting autoregressive visual generation speed by 3.6-9.5x, this research introduces parallel processing while preserving model simplicity and generation quality.