3D generation, computer vision, graphics, CUDA inference, and implementation details between commits.
Why generic compiler paths underperform on dynamic sparse 3D generative models—and how null-context attention removal and model-aware CUDA kernels cut inference latency.
Read post
Why naive GEMM becomes memory-bound, how tiled reuse changes arithmetic intensity, and where Tensor Cores, schedulers, and libraries fit
How a fast-follower team tracked the evolution of 3D AI to build an in-house generative 3D service—from VecSet VAEs to Voxel-conditioned 3D DiTs.
A rigorous derivation of why a signed distance field satisfies the Eikonal equation, from its geometric definition to gradient bounds
Why a Straight Line? Unpacking the Physics and Geometry of Rectified Flow
How Generalized Winding Numbers turn imperfect meshes into robust inside–outside tests for signed distance field construction
Comparing vecset and sparse-voxel VAE–DiT architectures, from 3D representation and compression to shape-generation design
Preparing training-ready 3D data through watertight mesh conversion, surface sampling, and salient-edge-aware preprocessing
Comparing Google model-viewer with a custom Three.js pipeline for embedding, controlling, and presenting 3D assets on the web
Tracing 3D latent diffusion from ShapeVAE to TRELLIS and Hunyuan3D, and comparing how modern systems represent and generate shape
Applying NeRF and Gaussian Splatting beyond photography to X-ray tomography, plus a lightweight viewer workflow built with Viser.
A practical Three.js guide to embedding an interactive Gaussian Splatting scene in a website, with no web-graphics background required
How NVIDIA’s 3D Gaussian Ray Tracing replaces splat rasterization with differentiable ray tracing—and what that changes.
A ground-up guide to Gaussian Splatting and differentiable rasterization, tracing the mathematics through its CUDA implementation—from projection and covariance
Comparing VGGSfM and MASt3R as learned Structure-from-Motion pipelines for camera estimation and radiance-field reconstruction.
How 2D Gaussian Splatting replaces volumetric blobs with oriented disks to improve surface geometry and mesh reconstruction
Evaluating diffusion-estimated depth by turning Marigold predictions into textured meshes and inspecting their geometric consistency.
Testing whether video generation models encode enough 3D consistency to reconstruct radiance fields from their synthesized views
Examining the quality, rendering, and production constraints that limit NeRF in games—and the workarounds available today
Reviewing Instant-NGP’s multi-resolution hash encoding and rebuilding its NeRF pipeline step by step in PyTorch
Review of Fourier Features Let Networks Learn High-Frequency Functions