Computer Vision

Adaptive Blind All-in-One Image Restoration (ABAIR) efficiently handles diverse image degradations, generalizes well to unseen distortions, and easily incorporates new ones via efficient fine-tuning.

AC3D achieves precise 3D camera control in video diffusion transformers by analyzing camera motion’s spectral properties, optimizing pose conditioning, and using a curated dataset of dynamic videos.

WF-VAE boosts video VAE performance with wavelet-driven energy flow and causal caching, enabling 2x higher throughput and 4x lower memory usage in latent video diffusion models.

Omegance: One parameter precisely controls image detail in diffusion models, enabling flexible granularity adjustments without model changes or retraining.

MARVEL-40M+ & MARVEL-FX3D: 40M+ high-quality 3D annotations & a fast two-stage text-to-3D pipeline enabling high-fidelity 3D model generation within 15 seconds.

ConsisID achieves high-quality, identity-preserving text-to-video generation using a tuning-free diffusion transformer model that leverages frequency decomposition for effective identity control.

DreamMix enhances image inpainting by disentangling object attributes for precise editing, enabling both identity preservation and flexible text-driven modifications.

DreamCache enables efficient, high-quality personalized image generation without finetuning by caching reference image features and using lightweight conditioning adapters.

Collaborative Decoding (CoDe) dramatically boosts visual auto-regressive model efficiency.

ChatGen-Evo automates text-to-image generation from freestyle chatting, simplifying the process and significantly improving performance over existing methods.

AnchorCrafter animates cyber-anchors selling products via human-object interacting video generation, achieving high visual fidelity and controllable interactions.

VisualLens leverages user visual history for personalized recommendations, improving state-of-the-art by 5-10% and exceeding GPT-4’s performance.

SplatFlow: A novel multi-view rectified flow model enabling direct 3D Gaussian splatting generation & training-free editing for diverse 3D tasks.

SAR3D: Blazing-fast autoregressive 3D object generation and understanding using a multi-scale VQVAE, achieving sub-second generation and detailed multimodal comprehension.

Image editing is revolutionized by Frame2Frame, which uses video generation to produce seamless and accurate edits, preserving image fidelity.

OneDiffusion: A single diffusion model masters image synthesis & understanding across diverse tasks, from text-to-image to depth estimation, pushing the boundaries of AI.

Self-supervised learning of 3D representations from procedurally generated synthetic shapes achieves comparable performance to models trained on real-world datasets, highlighting the potential of synt…

FQGAN revitalizes image generation by introducing Factorized Quantization, enabling scalable and stable visual tokenization with state-of-the-art performance.

DREAMRUNNER generates high-quality storytelling videos by using LLMs for hierarchical planning, motion retrieval, and a novel spatial-temporal region-based diffusion model for fine-grained control.

BootComp: generate realistic human images wearing multiple garments using a novel synthetic data pipeline & diffusion model, enabling diverse applications like virtual try-on.