Inference-Time Pointwise Velocity for Diffusion Acceleration
ACM Transactions on Graphics · 09.11.2026 · 45 (6) · 1-22
DOI: 10.1145/3842513 ↗Kurz gefasst
Fachartikel in der Zeitschrift ACM Transactions on Graphics, online erschienen am 09.11.2026. Verfasst von Itay Hazan und 1 weiteren Autor:innen. Die Arbeit nennt 146 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Large rectified-flow image generators such as FLUX and Z-Image deliver state-of-the-art quality, but their sampler advances every latent coordinate under a single global time. We argue that this uniform schedule is wasteful: at any intermediate noise level some image regions are already structurally settled while others still carry most of their stochasticity, and the settled regions would benefit the unsettled ones if allowed to advance further. We introduce ITPV, an inference-time, training-free method that assigns each latent coordinate its own denoising speed. ITPV generalises adaptive Layer Normalization to accept a per-token time map, reusing the pretrained time embedder and modulation weights without architectural changes. At every step it constructs a single speculative future latent through a Tweedie-style extrapolation, queries the model once at this accelerated time, and grades each coordinate by the cosine similarity between its current and speculative velocities. Coordinates whose direction would not change after the jump are advanced aggressively; coordinates whose direction would change stay on the canonical schedule. The result is a strict generalisation of standard sampling: it recovers the canonical trajectory exactly in the limit of zero per-coordinate speedup, requires no caching or profiling, and integrates with classifier-free guidance through a single extra forward pass per step. On FLUX.2-Klein-4B-Base, Z-Image, and Stable Diffusion 3, ITPV at 16 sampling steps matches the perceptual quality of 50-step classifier-free guidance, as confirmed by automatic distribution, quality, and diversity metrics on 30K MS-COCO captions and by the GenEval2 compositional benchmark. At a matched compute budget, ITPV improves over every training-free accelerator we compare against, including state-of-the-art samplers, caching, and region-adaptive methods. We further validate our results with a blinded pairwise user study with 14 participants on two of the backbones.
Zitieren
APA
Hazan, I., & Wolf, L. (2026). Inference-Time Pointwise Velocity for Diffusion Acceleration. ACM Transactions on Graphics, 45(6), 1-22. https://doi.org/10.1145/3842513
Vancouver
Hazan I, Wolf L. Inference-Time Pointwise Velocity for Diffusion Acceleration. ACM Transactions on Graphics. 2026;45(6):1-22. doi:10.1145/3842513
BibTeX
@article{Hazan2026,
author = {Hazan, Itay and Wolf, Lior},
title = {Inference-Time Pointwise Velocity for Diffusion Acceleration},
journal = {ACM Transactions on Graphics},
year = {2026},
volume = {45},
number = {6},
pages = {1--22},
doi = {10.1145/3842513}
}
EUVIMED