Daily digest — 2026-05-05

Diffusion-into-3D day. The thread: how Score Distillation Sampling (SDS) evolved from a clever trick into a default and is now being partially displaced.

DreamFusion (Poole et al., 2022) is the SDS paper. A 2D text-to-image diffusion model is used as a critic that backpropagates a per-view rendering of a NeRF toward "looks like the prompt". No 3D training data; the prior comes entirely from the 2D diffusion model. The ashawkey/stable-dreamfusion reimplementation is what most people actually ran, since the original used proprietary Imagen weights.

Magic3D (Lin et al., 2022) is the coarse-to-fine, mesh-export version: low-resolution NeRF first, then a high-resolution textured-mesh refinement stage. The two-stage pattern becomes standard.

Zero-1-to-3 (Liu et al., 2023) is a different lever: fine-tune Stable Diffusion to condition on camera pose relative to a reference image. Now your prior is geometric, not just appearance-level. cvlab-columbia/zero123 became the default novel-view prior for downstream image-to-3D. Magic123 (Qian et al., 2023) combines a 2D Stable Diffusion SDS loss with a 3D Zero-1-to-3 SDS loss — get the appearance richness of 2D priors plus the geometric consistency of 3D ones.

ProlificDreamer (Wang et al., 2023) reframes SDS as a particle-based variational approximation (Variational Score Distillation). The mode-collapse problem of vanilla SDS — overly smooth, oversaturated geometry — largely goes away. Best SDS-line read in the corpus.

The pivot is DreamGaussian (Tang et al., 2023): replace the NeRF backbone with 3D Gaussians for SDS. ~2 minutes instead of ~hours per asset, because rasterisation is faster than ray marching. The dreamgaussian/dreamgaussian release is small and readable; it's where I'd start anyone learning this sub-field today.

Six papers, ~55 min. Next week: pull these threads together.