AvatarMe++ uses differentiable rendering and a GAN to infer render-ready facial geometry and BRDF from one in-the-wild image.

Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis et al.

Using DINO-ViT deep features as dense descriptors enables unsupervised semantic segmentation and correspondence with competitive performance.

Shir Amir, Yossi Gandelsman, Shai Bagon et al.

Plenoxels uses sparse voxels and spherical harmonics to replace NeRF, training about 100× faster with no visible quality loss.

Alex Yu, Sara Fridovich-Keil, Matthew Tancik et al.

Enhancing image restoration with Test-time Local Converter (TLC), improving single-image deblurring PSNR from 32.92 dB to 33.57 dB on GoPro dataset.

Xiaojie Chu, Liangyu Chen, Chengpeng Chen et al.

Ref-NeRF reparameterizes view-dependent radiance using reflection directions, greatly improving specular reflection realism.

Dor Verbin, Peter Hedman, Ben Mildenhall et al.

ProtoPool model achieves interpretable image classification via differentiable prototype assignment, excelling on CUB-200-2011 and Stanford Cars datasets.

Dawid Rymarczyk, Łukasz Struski, Michał Górszczak et al.

Mask2Former, a Transformer-based universal segmentation architecture, achieves SOTA on four datasets, outperforming specialized models with a novel masked attention mechanism.

Bowen Cheng, Ishan Misra, Alexander G. Schwing et al.

MViTv2 employs decomposed relative positional embeddings and residual pooling to enhance multiscale vision transformers, achieving 88.8% ImageNet accuracy and superior detection and recognition performance.

Yanghao Li, Chao-Yuan Wu, Haoqi Fan et al.

Dream Fields combines NeRF and CLIP for zero-shot text-guided 3D object generation, achieving 68.3% R-Precision without 3D supervision.

Ajay Jain, Ben Mildenhall, Jonathan T. Barron et al.

MonoScene introduces monocular 3D semantic scene completion using a dual 2D/3D UNet framework with FLoSP and CRP modules, outperforming existing methods.

Anh-Quan Cao, Raoul de Charette

Diffusion Autoencoders combine meaningful high-level semantics with detailed reconstruction, enabling attribute manipulation and efficient representation learning.

Konpat Preechakul, Nattanat Chatthee, Suttisak Wizadwongsa et al.

VQ-Diffusion combines VQ-VAE and DDPM for text-to-image generation, achieving 15x speed improvement.

Shuyang Gu, Dong Chen, Jianmin Bao et al.

Blended Diffusion combines DDPM and CLIP for localized text-guided image editing, ensuring background preservation with high realism.

Omri Avrahami, Dani Lischinski, Ohad Fried

Introduces OOD-CV benchmark to evaluate vision models' robustness against individual nuisances like pose, texture, weather; shows current methods have limited improvements.

Bingchen Zhao, Shaozuo Yu, Wufei Ma et al.

GMFlow replaces local regression with global matching, reaching Sintel EPE 1.08 after one refinement, better than 31-step RAFT.

Haofei Xu, Jing Zhang, Jianfei Cai et al.

VIOLET introduces an end-to-end video-language Transformer with Masked Visual-token Modeling, achieving SOTA on multiple tasks.

Tsu-Jui Fu, Linjie Li, Zhe Gan et al.

EvDistill combines bidirectional reconstruction and cross-modal distillation, reaching 58.02% mIoU on unlabeled DDD17 events.

Lin Wang, Yujeong Chae, Sung-Hoon Yoon et al.

mip-NeRF 360 leverages non-linear scene parameterization and online distillation, reducing error by 57% for unbounded scenes with high realism.

Jonathan T. Barron, Ben Mildenhall, Dor Verbin et al.

RedCaps leverages Reddit data, creating 12M image-text pairs, improving vision-language models' performance.

Karan Desai, Gaurav Kaul, Zubin Aysola et al.

DyTox achieves continual learning with dynamic token expansion, excelling on ImageNet1000.

Arthur Douillard, Alexandre Ramé, Guillaume Couairon et al.