ADR-066: Two subtopics under generative modeling

Context

The owner asked how the anthology covered flow methods, then for the gaps to be filed, and suggested the large catch-all topics might take a subtopic or two. The count bears that out. generative-modeling carries 169 documents, 115 of them as primary, and its blurb (“diffusion, samplers, text-to-image, conditioning and control”) holds GANs, tokenizers, video systems, evaluation and flow theory alike. inference-optimization carries 89, and few-step generation is split between the two: progressive distillation and consistency models sit under one, DMD2 and CausVid under both, DPM-Solver under the first only.

The flow-methods batch adds about twenty more documents to exactly this seam, so a reader asking “what does the anthology hold on flow matching” or “on one-step samplers” would have to read the whole of generative-modeling to find out.

Decision

Add two words to topics, in the secondary-only section beside tiny-models and capability-thresholds:

• flows-and-transport: generative models as learned transport between distributions — flow matching, rectified flow, stochastic interpolants, normalizing flows and neural ODEs, and the paths, couplings and time schedules they are trained on. Gaussian diffusion is the special case (THEORY-106), so a diffusion paper takes it only when its argument runs through the flow or transport formulation.
• few-step-generation: getting a sample in one or a few network evaluations — progressive, distribution-matching and adversarial distillation, consistency, flow-map, shortcut and mean-flow models, and fast ODE solvers.

They are subtopics by convention, not by mechanism. A document carries one beside its parent topic, never as its first tag, so primary_topic and every chain invariant that compares it are unchanged. Nothing checks this; it is the same unchecked convention tiny-models lives under.

Retagged now, each with the subtopic placed directly after its first tag, so no primary topic moved:

• flows-and-transport: Flow Matching (LIT-630), Rectified Flow (LIT-636), both stochastic-interpolant papers (LIT-644, LIT-645), “Diffusion Meets Flow Matching” (LIT-678), SiT (LIT-447), SD3 (LIT-449), Score SDE (LIT-722), FLUX.1 Kontext (LIT-573), Movie Gen (LIT-626) and Wan (LIT-619); SOTA-263, SOTA-265, SOTA-266, SOTA-392 and THEORY-106.
• few-step-generation: Consistency Models (LIT-093), progressive distillation (LIT-067), DPM-Solver (LIT-076), DMD (LIT-643), DMD2 (LIT-646), CausVid (LIT-631), Self Forcing (LIT-629) and DMAD (LIT-770); SOTA-206, SOTA-289, SOTA-392 and SOTA-394.

The flow-methods batch filed in the same contribution takes them as well.

Alternatives considered

• Promote topics to a scheme, so each term can name a parent. This is what luria recommends once values need a hierarchy, and it would let the views nest subtopics under their parent and the lint check the convention. It is a structural migration, needs a newer luria than the 0.33.3 this record pins, and touches every document’s tags. Worth doing if subtopics multiply; not for two.
• Split generative-modeling into primary topics, so that a flow paper no longer takes generative-modeling first. That moves the primary topic of dozens of documents and every chain that compares it, to fix a browsing problem that a secondary word fixes without moving anything.
• One word, diffusion-and-flows. Most of generative-modeling is diffusion, so the word would be nearly as large as its parent and browse no better.