← Home

Why Few-Step Diffusion Fails for Text but Works for Images

By James Trappett · 2 July 2026

4 min read

Practitioners building continuous-latent text diffusion models have long observed an uncomfortable asymmetry: the same few-step deterministic generation techniques that produce excellent images in one to four network evaluations produce incoherent, token-flipping garbage when applied to text. The standard explanations invoke training budget, model scale, or scheduling choices. This paper argues those explanations are wrong, and proves it. The failure is geometric and irreducible within the deterministic-continuous class. Full paper: arXiv:2606.30705.

The Core Problem

Image decoders map continuous latents to pixel values through smooth functions. Text decoders, by contrast, terminate in an argmax over a vocabulary embedding matrix, partitioning the latent space into sharp Voronoi-like cells. A deterministic transport map, however accurate, delivers its output only to within some residual blur. For images, that blur is absorbed gracefully because the decoder is smooth. For text, the same blur, if it lands near a decision boundary, flips a token. The question is not whether the transport is accurate on average, but whether it can guarantee placement within the correct argmax cell for every token position simultaneously.

The paper formalises this intuition through Theorem 3, which states that the token flip rate at the posterior-mean terminal step scales as the latent mass within an O(s(t)) tube around decision boundaries, where s(t) is the transport residual magnitude. Critically, this rate is set by decoder sharpness, not by how well the flow was trained. You cannot train your way out of it while remaining in the deterministic-continuous class.

Key Contributions

Methodology and Empirical Validation

The theoretical framework distinguishes two regimes: an overlapping regime where text autoencoder latents from different tokens genuinely intermingle, and an idealised separated regime used to derive clean scaling laws. Theorem 3 operates in the overlapping regime under smooth density assumptions and exact decoding, with Proposition 16 extending coverage to learned denoisers whose terminal active-normal residual is stable and O(s).

The empirical work is careful about what it is and is not claiming. DABI is measured using a single margin-normal probe applied identically to all codecs, with 95% bootstrap confidence intervals. The image-text gap is not a marginal statistical finding; boundary-aligned perturbations flip 47-77% of tokens in text decoders while isotropic perturbations of identical norm flip essentially zero. The authors also handle potential counterexamples explicitly.

FMLM, a one-step deterministic flow model for text, appears at first glance to refute the entire argument. The paper addresses this directly: FMLM interpolates toward one-hot simplex targets and applies argmax decoding, which is categorical commitment by definition (CCI=1). The authors measure DABI on the official FMLM checkpoint and find the same sharp-readout signature as autoregressive models. FMLM succeeds because one-hot targets provide large margins, not because it achieves smooth continuous transport. Similarly, FastDiSS's training-time perturbation regularisers and Loopholing's deterministic feature pathway are correctly classified as orthogonal to or consistent with the taxonomy rather than counterexamples.

The survey of 2025-2026 continuous-latent text generators finds that the deterministic-ODE, smooth-continuous, few-NFE cell of the taxonomy is empty across all surveyed literature. Every few-step system uses either SDE sampling or categorical commitment. This is a strong empirical regularity, and the theory explains why it should hold.

Limitations and Open Questions

The formal results assume smooth densities and exact clean decoding. Learned generators only approximate posterior-mean transport, and the bounds apply strictly to deterministic samplers. The authors note that label noise, finite-perimeter irregularities, and codimension-2 boundary effects change constants but not the linear-in-s scaling, which gives the first-order law reasonable robustness to these idealisations.

A more substantive open problem is a Wasserstein theory for overlapping latents. The current sharp transport laws (Theorems 5-7) are proved in the separated regime, and the coarea identity of Theorem 17 connects them to the overlapping tube, but a full within-core W_p theory remains open. This matters practically because real text autoencoder latents are not well-separated.

The paper is also explicit that its results constrain the deterministic-continuous class and say nothing about the quality ceiling achievable by SDE or categorical generators. The finding is a lower bound on what deterministic transport cannot do, not an upper bound on what alternatives can achieve. Whether SDE sampling or categorical commitment introduces its own quality costs at very low NFE is a separate question left open.

For practitioners, the implications are fairly direct. Continuous-latent text diffusion models requiring fewer than roughly 16 steps need either stochastic sampling or a categorical commitment mechanism. Training longer or scaling up will not resolve the geometric constraint. The DABI and CCI diagnostics are cheap to compute on published checkpoints and provide a principled basis for classifying new systems before committing to a generation strategy.

Diffusion ModelsText GenerationGenerative ModelsNLPMachine Learning Theory

Related Articles

When Does Learned Early Stopping Beat Simple Thresholds?When Does Feedback Actually Help? Separating Signal from NoiseRSEA: Safer LLM Agent Self-Evolution via Held-Out Selection