Mingfei Guo

NVIDIA Omniverse NuRec

NVIDIA · Neural reconstruction · 2026

Re-render a recorded drive through a different camera rig — new mounting positions, intrinsics, and lens models (pinhole, fisheye, f-theta) — so perception for a new vehicle platform can be trained before that platform has real data. Scenes are 3D Gaussian splatting reconstructions of six-camera drives; the public Physical AI NuRec Dataset ships 1,500+ of them.

Training on the re-rendered data improved detection precision and recall in every object category over a zero-shot baseline. I work on NuRec’s reconstruction infrastructure and performance, and coauthored this walkthrough.

Apurv Naman, Mingfei Guo, Dominik Froehlich, Wonsik Han

Asset Harvester

NVIDIA · Coauthor · 2026

Neural reconstruction turns a driving log into a scene, but not into objects you can move or view from a new angle. Asset Harvester takes the few, limited-angle views a log has of a vehicle or pedestrian and produces a complete 3D Gaussian asset: SparseViewDiT generates the missing views, then a lifting stage fits Gaussians to them.

The system-level work is what makes it hold up on real data — large-scale curation of object-centric training tuples, geometry-aware preprocessing across sensors, augmentation, and self-distillation.

VRL / visual-rl

Training infrastructure · RL post-training

RL post-training for diffusion and flow models, built because LLM-RL frameworks assume a categorical token action. A visual policy’s step is a denoise transition over latents, its reward is computed on decoded pixels or video, and its conditioning is an image or video rather than a text prefix — so VRL keeps those semantics explicit behind one Collector → Evaluator → Algorithm loop that covers both full-sequence and chunk-autoregressive generation.

Model, reward (OCR, aesthetic, CLIP, PickScore, VideoReward), algorithm (GRPO, DPO, DiffusionNFT, Flow-DPPO), and distributed settings compose from YAML layers. 20+ image and video families are wired in — SD3.5, FLUX, Wan 2.1/2.2, HunyuanVideo, CogVideoX, Cosmos-Predict2 — and a recipe only counts as validated once a real run shows rising reward and changed weights. The SD3.5 OCR-GRPO run above is one.

3D Gaussian Splatting, from scratch

GPU programming · Differentiable rendering

The whole 3D Gaussian Splatting method — projection, rasterization, the backward pass, Adam, and densify / prune — reimplemented in Python with NVIDIA Warp instead of the reference CUDA. The same kernels run on CPU or GPU with a one-line config change, so the method fits in a handful of readable files and trains on a laptop.

Forward and backward follow graphdeco-inria’s reference implementation; densification follows gaussian-splatting-lightning, restructured to strip data preparation down to the minimum.