Skip to content
Bud's Wiki

Awesome Diffusion Categorized

TL;DR

Section titled “TL;DR”

A long-running awesome-list of diffusion-model papers organized by sub-area: acceleration / distillation, image restoration (colorization, face restoration, compression, super-resolution), text-guided editing, virtual try-on, drag editing, personalization (ID encoder, multi-concept, decomposition), in-context learning, image translation, segmentation/detection/tracking, inpainting, layout, count guidance, color, illusion, and poster generation. Each entry is paper title + venue badge + arXiv/project/code links with no synthesis — the value is breadth and the categorical taxonomy, not commentary. The list spans roughly 2023–2026 and is updated continuously by the maintainer.

Key claims

Section titled “Key claims”
  • The repo’s organizing principle is sub-area taxonomy rather than chronology: 20+ top-level categories, several with two-level nesting (e.g. Accelerate → Train-Free / AR model / VAR model; Image Restoration → Colorization / Face Restoration / Compression / Super Resolution / Personalized Restoration; In Context Learning → Multi-Concept / Decomposition / ID Encoder / General Personalization / AR-based / Video Customization) [Contents].
  • The Accelerate section is the largest single category and mixes distillation (DMD1/2, LCM, PCM, SwiftBrush, SDXL-Turbo/Lightning, InstaFlow, Score-identity Distillation, pi-Flow), training-free caching (TeaCache, AdaCache, FasterCache, MagCache, TaylorSeer, EasyCache, Block Caching), and architectural acceleration (DC-Gen, SANA-Sprint, Skip-DiT) into a single list without sub-grouping by technique [§ Accelerate].
  • Entries are not commit-dated and not chronologically sorted within sections — venue tags (CVPR / NeurIPS / ICLR / ECCV / ICML / Website / arXiv) are the only freshness signal embedded in the listing itself [list format].
  • The list cross-references several papers also filed independently in this wiki: pi-Flow (arXiv:2510.14974), Autoregressive Distillation of Diffusion Transformers (arXiv:2504.11295), Decoupled DMD (arXiv:2511.22677), TurboDiffusion (arXiv:2512.16093), ArcFlow (arXiv:2602.09014), TDM-R1 (arXiv:2603.07700) [§ Accelerate].
  • The repo is currently maintained as a personal-but-public resource; the README shows historical mentions of a postdoc-recruitment header for the LAMP group, suggesting the maintainer treats it as part of their research-visibility surface [README top].

Method

Section titled “Method”

A flat README-driven index. Each entry follows the pattern **Title** \ [[Venue/Website](arxiv-url)] [[Project](project-url)] [[Code](code-url)]. Internal anchor links power the table of contents. No metadata schema beyond venue tag — no date, no author list, no abstract excerpt, no per-paper notes. New entries appear to be appended within their section without re-sorting, which is what makes git-log-based chronological browsing painful (the motivation behind Tucker’s companion website, awesome-diffusion-papers.vercel.app — time-sorted, searchable browser for awesome-diffusion-categorized).

Results

Section titled “Results”

The categorical taxonomy aligns closely with several of this wiki’s existing concept pages — Accelerate / Train-Free maps to Diffusion Distillation and Diffusion serving optimization; Image Restoration → Super Resolution maps to Video Super-Resolution / Restoration via Diffusion Priors; In Context Learning → ID Encoder / Multi-Concept maps to Parameter-Efficient Finetuning; Layout / Inpainting / Drag Edit are adjacent to Layered Image/Video Decomposition. As an indexing surface for “what has the field done on X sub-problem”, the README is more thorough than any single survey on the wiki, at the cost of zero synthesis.

Why it’s interesting

Section titled “Why it’s interesting”

This is the diffusion counterpart to Awesome World Models — same shape (community-maintained README index, no synthesis, sub-area taxonomy) but for image/video diffusion sub-fields rather than world models. For the team it’s most useful as a discovery surface when a new editing / restoration / acceleration paper arrives and the question is “what was already done in this exact sub-area”; the Accelerate section in particular is dense enough that searching it before filing a new distillation paper would catch most prior art faster than arxiv search. The list does not synthesize, so any claim that “technique X dominates sub-area Y” still has to come from reading the underlying papers, not from this repo’s structure.

See also

Section titled “See also”