Diffusion Lighthouse

A curated, dataset-first map of diffusion research.

Not a leaderboard. Not an archive.

0 papers
0 results

How to read Diffusion Lighthouse

A curated, dataset-first map of diffusion research. Not a leaderboard. Not an archive.

← Back to index

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.

← Back to index

Editorial Policy

How papers get included in Diffusion Lighthouse — and why we intentionally say “no” more than “yes”.

← Back to index

Diffusion Lighthouse is a curated, dataset-first map of diffusion research. Inclusion is editorial, not automatic. The goal is clarity, not completeness.

Scope

Lighthouse includes papers where diffusion models are the primary modeling framework (or a documented direct ancestor), and where the contribution meaningfully advances understanding, methodology, or applicability of diffusion-based generative modeling.

Inclusion Criteria

A paper must satisfy all Core Criteria and at least one Contribution Criterion.

Core Criteria (required)

  1. Peer-reviewed-first — canonical proceedings or journal links are shown whenever available. Canonical preprints are included only by exception and are explicitly labeled with an editorial note.
  2. Central use of diffusion — diffusion is the main modeling approach, not a minor baseline or appendix comparison.
  3. Original contribution — introduces or meaningfully refines a method, theory, modeling assumption, or widely adopted practice. Pure scaling reports without conceptual insight are excluded.
  4. Clear dataset grounding — clearly specifies the data distribution it models and motivates why the approach applies to it. Purely toy or synthetic setups without broader relevance are excluded.

Contribution Criteria (≥1 required)

  • Foundational contribution — new diffusion formulation, objective, or theoretical interpretation.
  • Methodological advancement — improvements in sampling, conditioning, efficiency, stability, or training dynamics.
  • Dataset-level impact — enables diffusion on a new modality, or makes a previously infeasible dataset tractable.
  • Conceptual synthesis — unifies prior work, clarifies relationships between methods, or becomes a standard reference.

Canonical preprint exception

A paper may be included without peer-reviewed publication only if it is field-defining, has clear downstream lineage in later peer-reviewed work, and has no peer-reviewed substitute. These entries are explicitly labeled as Canonical preprint and include an editorial note.

The exception is rare by design — it does not apply to routine arXiv submissions, under-review work, or short-lived benchmark trends.

What Lighthouse is not

  • Not a leaderboard
  • Not an automatic crawler
  • Not a complete bibliography
  • Not a substitute for reading the papers

Explicit Exclusions

Papers are excluded if they are primarily:

  • Benchmark chasing with marginal gains and no new insight
  • Minor architectural tweaks without analysis
  • Narrow task-specific tricks that do not generalize
  • Redundant reformulations of existing ideas

Exclusion is a feature, not a limitation.

Editorial Philosophy

  • Lighthouse is dataset-first and idea-centric: benchmarks and metrics are annotations, not organizing principles.
  • Citation counts provide context, not authority.
  • Judgment is explicit, documented, and revisable.

The goal is not completeness — it is clarity.

← Back to index