Mingyang Song

Ph.D. Student between DisneyResearch|Studios & ETH Zürich.
Dynamic scene reconstruction and compression;
Video denoising;
Digital camera noise modeling.

Portrait of Mingyang Song on a snow-covered mountain path.

Selected Projects

Motion-vector visualisation of a dynamic 3D-Gaussian scene: overlapping colourised poses of an animated character, shown for two compression settings side by side.

SmoothMotionVectors

Optimizing your content for video codecs in free-view video compression. A compressed 4D Gaussian scene, decoded and rasterized entirely in your browser — no GPU, no install.

SIGGRAPH ’26 · Conference Track Open the live viewer
A spline trajectory through knots in a canonical 3D frame: canonical points map through per-timestep deformation fields to positions and velocities at time t and t+1.

Spline Deformation Field

To reconstruct dynamic scenes from images, we revisit classic splines and blend them with modern coordinate neural networks. This fusion combines the best of both, achieving smoother temporal transitions and better spatial coherence.

SIGGRAPH ’25 · Conference Track Playground
The same photographic crop three times: clean, then regrained as an LG G4 at ISO 800 and at ISO 1600, the noise visibly coarser and more coloured at the higher setting.

Digital Camera Noise Synthesis

A generative model for real sensor noise, distilled down to two measured ingredients — an intensity-to-variance curve and a spatial correlation kernel. Regrain your own photographs as five smartphones across fourteen ISO levels, live, on your own CPU.

arXiv ’23 · Preprint Playground

Contact

Personal misong9756@gmail.com ETH Zürich mingyang.song@inf.ethz.ch

Elsewhere

Pixiv Digital illustration View gallery