Lipman, Havasi, Holderrieth, Shaul, Le, Karrer, Chen, Lopez-Paz, Ben-Hamu and Gat, FAIR at Meta, MIT CSAIL and Weizmann Institute (2024) — ARXIV-2412.06264. Read at v1 (9 Dec 2024), the only version.
What was read: §§1–2 in full; §3 (flow models, continuity equation) for its theorems, not every proof; §4 in full, the claim-making core (losses, paths, the OT bound, affine and Gaussian paths, scheduler change, couplings, guidance); §5 (Riemannian) in full; §§6–7 (CTMC and discrete flow matching) in full; §§8–9 (general Markov processes, Generator Matching) for statements, skimming the jump-process derivations; §10 (diffusion) in full; Appendix A’s proofs not checked.
Key takeaways
- One loss family (§4.5, Prop. 1, Thm. 4). The conditional loss has the marginal loss’s gradient for any Bregman divergence, not just squared error, because the divergence’s gradient is affine in its first argument. That is what licenses KL losses for the discrete case (§7.5) and the generalized KL that is also an ELBO (Shaul et al. 2024).
- “Sample t, don’t weight by it” (§4.5): mathematically equivalent, but sampling t ∼ ω “leads to better performance than using weights ω(t) in large scale image generation”, citing SD3. This is an attribution, not a result of the guide.
- What “OT” means for the straight path (§4.7). The linear conditional flow minimizes an upper bound on kinetic energy over all conditional flows (Eqs. 4.46–4.48). It is the true OT solution only when the target is a single point (Thm. 5). The marginal’s kinetic energy is at most the coupling’s transport cost (Eq. 4.49), so a cheaper coupling gives a cheaper flow (§4.9).
- Schedules are interchangeable after training (§4.8.2). A velocity trained under one affine scheduler converts in closed form to any other with the same coupling, and in theory all schedulers reach the same samples at t = 1 (Eq. 4.68, after Shaul et al. 2023). For Gaussian paths the marginal velocity is a gradient, so it is kinetic-optimal for its own marginal path (Eq. 4.78).
- Couplings (§4.9). Paired data (masked image to full image) trains inpainting or super-resolution as transport. Minibatch OT couplings lower transport cost and straighten paths as the batch grows.
- Guidance (§4.10). Classifier and classifier-free guidance carry over to velocities through the score conversion. “The exact distribution which CFG samples from is unknown.”
- Other state spaces (§§5–7, 9). Geodesic conditional flows train simulation-free on manifolds with closed-form exp/log maps. Premetric flows need simulation. Discrete flow matching uses factorized velocities (d·K outputs, not K^d) and mixture paths, and a divergence-free term added at sampling time leaves the marginals unchanged, which is what a corrector is (§§6.3.1, 7.5).
- Diffusion is a case of it (§10). Training a diffusion model is flow matching with an affine-drift Gaussian path and score or ε parameterization. The probability-flow ODE and the SDE family share marginals. The backward-time reversal is not needed for generation.
Where the hedges are
Per DP-010:
- No experiments. Every claim is a derivation or an attribution. The sentences that sound empirical are other papers’ results: SD3 on timestep sampling, EDM on the probability-flow ODE being “state-of-the-art for sampling with a low number of neural network evaluations” (§10.5), Liu et al. 2023 and Albergo et al. 2024 on paired-data couplings beating guided diffusion.
- “State-of-the-art performance across various domains” (abstract, §1) is adoption: SD3, Movie Gen, speech, audio, proteins and π₀ are listed as users, not tested against alternatives (DP-005).
- The authors review their own line. Flow Matching, Discrete Flow Matching, Multisample Flow Matching, Riemannian Flow Matching and Generator Matching are all by authors of this guide, and they get the most space.
- Manifold singularities are dismissed as measure-zero, with “at-least in use cases we are aware of” (§5.6).
- Nothing on few-step generation. Reflow, distillation, consistency and flow-map models are absent. Straightness appears only as a property that makes ODE solving cheaper (§4.7).
Standing in the anthology
It is the reference for Flow Matching (LIT-630) and its extensions as their authors frame them two years on. A reader of LIT-630, of the minibatch-OT couplings (LIT-789, LIT-798) or of Discrete Flow Matching (LIT-792) can come here for the unified statement and the library. Rectified Flow (LIT-636) and the stochastic interpolants (LIT-644, LIT-645) appear as concurrent origins and as the two-sided conditioning variant.
On SOTA-266 and THEORY-106 it sharpens what the straight path is. §4.7 proves what THEORY-106 says in prose: the line is optimal per pair and bounds the marginal’s kinetic energy, but the marginal flow is OT only for a one-point target. §4.8.2 adds a reason the practice’s “No mechanism has been isolated” is the right hedge. If every scheduler reaches the same endpoint under the exact ODE, a path’s advantage has to come from what it does to training (loss weighting, target variance) and to discretization error. That is THEORY-106’s decomposition. The guide’s own recommendation is the cond-OT path, and it backs the logit-normal half only by citing SD3.
It has no bearing on the practices about few-step students, because it does not cover them.
Filed without a NOTE: the coverage is stated above. This is a reference text, and the note records what it asserts and on whose authority.