| name | pyramidal-diffusion |
| description | Multi-scale/pyramidal approaches to denoising diffusion models. Covers the Ryu & Ye 2022 pyramidal DDPM paper and related coarse-to-fine diffusion architectures. Relevant to subdivision mode / spatial budget training and hierarchical latent backbones in the PDLI experiment. |
Pyramidal Diffusion Models
Mental Model
Pyramidal diffusion models process images at multiple scales/resolutions within a single reverse diffusion process. Instead of running all denoising steps at full resolution, they start with coarse (low-resolution) denoising and progressively upsample/fine-tune as the process continues. This allocates most compute to low-resolution steps where it's cheaper, reserving high-resolution compute for refinement. The key architectural enabler is a single score function conditioned on a positional embedding that tells the network which resolution level it's currently processing.
Coverage
Documented: Ryu & Ye 2022 pyramidal DDPM paper; coarse-to-fine diffusion paradigm
Not yet documented: PyramidalWan (Korzhenkov et al. 2026), Pyramid Diffusion for 3D scenes, PyDiff for low-light enhancement
Last extended: 2026-06-22
Key Findings
Pyramidal Denoising Diffusion Probabilistic Models (Ryu & Ye 2022)
- Location:
docs/upstream/papers/ryu2022pyramidal.md
- ArXiv: 2208.01864
- What: Canonical pyramidal DDPM paper. Uses a single score function with positional embeddings to operate across multiple resolution scales. During reverse diffusion, starts at coarse resolution and progressively upsamples through the pyramid.
- Why it matters: Directly relevant to PDLI's subdivision mode / spatial budget training and hierarchical latent backbone. The single-model multi-scale approach provides a precedent for efficient compute allocation in diffusion models.
Positional Embedding for Scale Conditioning
- Location: Abstract of ryu2022pyramidal
- What: The score function is conditioned on which pyramid level it's processing via a learned positional embedding. This is the key innovation that allows one model to handle all scales.
- Why it matters: Simple architectural addition (positional embedding) enables multi-scale capability. Could inform how PDLI's subdivision governor conditions on resolution/latent size.
Related Papers Discovered
- Pyramid Diffusion for Fine 3D Large Scene Generation (Liu et al., ECCV 2024, arXiv:2311.12085) — 3D extension of pyramid diffusion using a "Pyramid Discrete Diffusion" (PDD) model with scale-varied diffusion.
- Pyramid Diffusion Models For Low-light Image Enhancement (Zhou et al., IJCAI 2023, arXiv:2305.10028) — PyDiff uses pyramid diffusion sampling with progressively increasing resolution.
- PyramidalWan (Korzhenkov et al., 2026, arXiv:2601.04792) — Pipeline to convert pretrained diffusion models into pyramidal ones via low-cost finetuning; investigates step distillation in pyramidal models.
Critical Invariants
- Pyramidal diffusion is about resolution-varying diffusion within a single reverse process, not about training separate models per scale.
- The efficiency gain comes from running most denoising steps at low resolution — only refinement happens at high resolution.
- Positional (or scale) embeddings are the standard mechanism for conditioning on resolution level.
Sources
- arXiv API search:
all:pyramidal AND all:diffusion (114 results)
- Paper pages: arXiv:2208.01864, arXiv:2311.12085, arXiv:2305.10028, arXiv:2601.04792
- Google Scholar citation count: 36 citations for Ryu & Ye 2022
