The result that kicked off the deep learning era: in 2012 a deep CNN cut ImageNet top-5 error from 26% to 15%, showing that GPU-trained networks with ReLU and dropout could beat decades of hand-engineered computer vision features.
Frames generative modeling as a two-player game: a generator forges data while a discriminator learns to spot fakes, training both by backpropagation alone — no Markov chains, no inference networks. The adversarial pressure yields sharp samples.
Introduced dilated (atrous) convolutions, which expand a filter's receptive field exponentially with no loss of resolution and no extra parameters — the trick that let dense-prediction networks see wide context while keeping per-pixel detail.
Re-examines residual blocks and shows that pure identity skip connections plus pre-activation (BN-ReLU before each conv) let gradients flow cleanly enough to train a 1001-layer ResNet, hitting 4.62% error on CIFAR-10.
Enables real-time (≥30 fps) 1080p novel-view synthesis by representing scenes as optimized anisotropic 3D Gaussians plus a visibility-aware splatting renderer; provides the paper's reference implementation, pretrained models and viewers — high-quality training requires CUDA GPU and significant VRAM.
AI Model2023
Estimates and tracks 6D poses of novel objects without per-object fine-tuning — supports both model-based (CAD) and model-free (few reference images) setups. Trained on large-scale synthetic data with a transformer-based architecture and contrastive learning; CVPR 2024 highlight with demos and pretrained weights.
AI Image2025
Converts document images—scans, photos, born-digital PDFs—into structured text in two stages: first map layout and reading order, then parse each element (text, tables, formulas, figures) in parallel, each guided by its own task prompt.
AI Video Papers2025
Argues a single web-scale generative video model handles vision tasks zero-shot the way LLMs handle language. Probes Veo 3 on segmentation, edge detection, image editing, physical and affordance reasoning, and puzzles like maze solving and symmetry.
Reduces object-driven shortcut learning in zero-shot compositional action recognition by enforcing temporal verb cues and regularizing against frequent object-verb co-occurrence priors. Proposes RCORE with Co-occurrence Prior Regularization (treats frequent co-occurrences as hard negatives) and Temporal Order Regularization. Evaluated on Sth-com and EK100-com with improved compositional generalization.
AI Video Papers2026
Provides a diagnostic suite that audits video-understanding benchmarks to find samples solvable without visual or temporal input, filters those shortcuts, and produces a distilled video-native testbed that reveals major capability gaps in current Video-LLMs.
AI Model2026
Generates high-fidelity 3D assets from a single image by back-projecting pixel-aligned features into 3D, preserving fine geometry and PBR textures; includes inference code and a Hugging Face demo—best suited for single-view object reconstruction.
AI Model2026
Transforms pretrained latent-diffusion priors into pixel-space diffusion models by removing the VAE and training shallow pixel layers on LDM-generated synthetic images — enabling fast convergence, native 4K output, and low-data training on 8 GPUs.