Albergo, Boffi and Vanden-Eijnden, New York University (2023) — ARXIV-2303.08797. Read as v4, the JMLR 26 (2025) version. Authors listed alphabetically, with equal contribution.
Key takeaways
- The latent term is what is new. x_t = I(t, x0, x1) + γ(t)z with γ(0) = γ(1) = 0 (Def. 1). Remark 2 names this term, and an optional coupling of x0 and x1, as the differences from ARXIV-2209.15571. The term smooths ρ(t) and the velocity. The paper shows this analytically on Gaussian mixtures, where spurious intermediate modes disappear (Fig. 4, §4.2).
- The noise level is decoupled from training. The density solves the transport equation with velocity b and, for any ε(t) ≥ 0, the forward and backward Fokker-Planck equations with drift b ± εs (Cor. 10). So one learned b and s define an ODE and a whole family of SDEs with the same marginals (Cor. 18). ε “can be specified after learning” (Fig. 3 caption). v1 already says so (§1.2).
- The losses bound SDE likelihood and not ODE likelihood. For the ODE the KL needs Fisher-divergence control that the regression does not give (Lemma 21). For an SDE with ε > 0 the KL is bounded by the drift error over 4ε (Lemma 22, eq. 2.42). So the b and s losses bound it (Thm. 23).
- Diffusion is a one-sided interpolant. With a Gaussian base, x_t = α(t)z + β(t)x1 (Def. 32, eq. 4.15). The score comes from a denoiser E[z|x_t] (eq. 3.17), and the velocity comes from that same denoiser via α η_z + β η_1 = x (eqs. 4.18–4.19). VP diffusion is α = √(1−t²), β = t after a change of time (eq. 5.5). Its singularity at the endpoint comes from compressing an infinite-time SDE, not from the density path (§5.1).
- Rectification does not change the map. Refitting on the learned coupling gives straight trajectories with the same X_1 (Thm. 47). Straight paths are “necessary … but not sufficient” for optimal transport (Remark 48).
- The SDE wins in low dimension. On a 2-D checkerboard, every ε > 0 beats the ODE (Figs. 7–8). On a 5-mode Gaussian mixture in d = 128 the best ε is nonzero for all four choices of learned pair. Learning b and the denoiser is best (Fig. 12). The KL is measured between KDEs of the first two coordinates (§7.2).
Where the hedges are
Per DP-010:
- “Illustrated on numerical examples” is the whole image story. §7.3 trains one-sided linear and trigonometric interpolants on 128×128 Oxford Flowers and shows samples and nearest neighbours only (Figs. 13–15). FID is left “for a future study”. No version reports FID or a diffusion baseline. Table 3 lists an ImageNet-32 configuration that has no result anywhere in the paper.
- The optimal ε cannot be computed as stated. ε* in eq. 2.49 is the square root of a ratio of excess losses, L − min L. The minima are unknown in practice. SiT’s w_KL (its eq. 13) is a different, computable choice, derived from Lemma 22.
- v1 claimed more than later versions keep. v1 §1.2 says the latent amplitude and the noise strength “can both be tuned as model hyper-parameter after training”. Changing γ changes ρ(t), and so changes what must be learned. By v3 only ε is post hoc.
- Eq. 4.9 is misprinted. It writes α = t, β = 1 − t, which violates the boundary conditions (4.2) and Table 1.
Standing in the anthology
SiT (LIT-447) extends this paper. SiT shares three of its authors. SiT’s Appendix A says “most proofs are derived from [2]”, and [2] is this paper. Its own text shows what it takes:
- the one-sided interpolant x_t = α_t x* + σ_t ε, reversed in time (its §2.1 and eq. 12)
- the velocity and score as conditional expectations (its eqs. 3 and 5)
- score-from-velocity through x = E[x_t | x_t = x] (its eqs. 8–9, this paper’s eq. 4.18)
- the free diffusion coefficient w_t, choosable after learning (its §2.4)
- the SDE KL bound it cites as “Lemma 2.22 in [2]” (this is Lemma 22 in v4). SiT uses it to derive w_KL (its App. A.5) and to explain why SDE beats ODE (its Table 5).
SiT’s GVP interpolant is ARXIV-2209.15571’s trigonometric path. Its linear path is Liu et al.’s. SiT adds the cost-regularized w_KL,η and the ImageNet measurements.
This moves where SOTA-265 comes from. The diffusion-coefficient identity
it credits to LIT-447 (introduced_by) and calls something “the record has
nothing like” is Corollary 10 here, with the likelihood argument in Theorem
23. SiT supplies the measurement on a frozen model. The earlier interpolant
paper, ARXIV-2209.15571, is filed alongside. This paper extends it.
From LIT-644 it keeps the objective, a velocity regressed onto the time derivative of an interpolant between two arbitrary densities with no ODE solved in training, and the trigonometric path among its interpolants. It adds the latent term γ(t)z and an optional coupling of the endpoints (Remark 2), and with them a learned score beside the velocity, a family of SDEs with the interpolant’s marginals, and a KL bound for those SDEs where LIT-644 bounded the ODE’s transport error in W2.
Filed without a NOTE: the takeaways come from one full reading of v4’s
main text and Appendix C done for this filing. The proofs in Appendix B were
skimmed and not checked. v1 and v3 were checked only for the claims above
and for image metrics.