How to read Diffusion Lighthouse
A curated, dataset-first map of diffusion research. Not a leaderboard. Not an archive.
Diffusion Lighthouse helps you navigate the diffusion literature by making editorial judgment explicit. It is not an exhaustive index and it does not rank papers by performance. Instead, it surfaces work that shaped how diffusion models are conceptualized and used.
Benchmarks are context. Datasets define regimes. Ideas are the map.
What “dataset-first” means
Papers are interpreted relative to the data distributions they target (images, text–image pairs, scale regimes, modalities), not primarily by benchmark scores.
Dataset focus helps clarify where a method applies, which claims are comparable, and why certain ideas mattered at the time they were introduced.
The inclusion bar
Most papers are excluded. Inclusion is editorial: a paper must be central to diffusion and deliver a lasting conceptual, methodological, or dataset-level contribution.
Read the full policy on Editorial Policy.
Publication labels
- Peer-reviewed: accepted at a major conference or journal; canonical proceedings/journal links are shown.
- Canonical preprint: included only by exception when a technical report defined a direction before formal publication and has clear downstream lineage. These entries are explicitly labeled and include an editorial note.
Relations: why the map matters
Relations are editorial claims (e.g., “builds on”, “unifies”, “simplifies”). They turn a list into a map: follow how ideas propagate from foundations → refinements → systems.
Citations are context
“Cited by X” is a best-effort snapshot (and may come from different sources). It is not real-time, and it is not a ranking signal. Lighthouse uses citations only as historical context.
Tip: sort by citations for a rough “gravity well” view — then open papers to read the why and the relations.