Synced · WeChatMay 21, 2026, 11:03 UTC+8

VAST and Tsinghua propose density-controlled 3D Gaussian generation for SIGGRAPH 2026

VAST+清华提出3D生成新范式，空间智能密度控制「把算力花在刀刃上」| SIGGRAPH 2026

VAST and Tsinghua propose DeG, a 3D Gaussian generation method that samples Gaussian centers from a learned density distribution and trains density control with a render loss contribution gradient; in some settings, it reaches TRELLIS-like visual quality with less than half the Gaussian count.

Why it matters: HKR-H/K/R pass: DeG offers a concrete mechanism and a testable efficiency claim, reaching TRELLIS-like quality with under half the Gaussians in some scenes. SIGGRAPH research has some technical depth, but no hard-exclusion rule applies.

Read the original ↗Export Markdown

Take

DeG attacks resource allocation, which is a better 3D generation problem than another single-image quality bump. It samples Gaussian centers from a learned density field, then trains that density with a render loss contribution gradient. The concrete claim is strong: in some settings, it reaches TRELLIS-like visual quality with under half the Gaussian count.

I’m cautious about the “new paradigm” framing. 3D Gaussian Splatting already lives on densification and pruning; DeG’s contribution is moving that heuristic into a trainable generator. Against fixed-structure lines like TRELLIS and LGM, the pitch is clean: one model can sample different Gaussian budgets at inference. But latency, training overhead, and cross-category robustness are not pinned down in the snippet. Half the Gaussians does not automatically mean half the system cost.

Source: Synced · WeChat · mp.weixin.qq.com
Published: May 21, 2026, 11:03 UTC+8 · May 21, 2026