arXiv：cs.AI· George Matthew (University of Exeter, United Kingdom), Guoqiang Zhang (University of Exeter, United Kingdom), Kenta Niwa (NTT Communication Science Laboratories, Japan)·· 19 小时前
TraceSampler：让加速生成采样的误差可预测、可校正
TraceSampler: Making Sampling Errors Predictable and Correctable for Accelerated Generative Sampling
AI 导读
TraceSampler 是一种后训练方法，无需修改预训练模型，即把少步采样的端点误差建模为样本特定的预测问题，并用两阶段学习残差校正来修正。在 CIFAR-10 与 SiT-XL/2 上，10 步采样相对匹配步数 Euler 将轨迹 MSE 降低 90.5% 和 85.6%，FID 达 2.692 与 2.189，接近 120 步 Heun 参考且主干评估少 24 倍。
正文
Authors:George Matthew (1), Guoqiang Zhang (1), Kenta Niwa (2) ((1) University of Exeter, United Kingdom, (2) NTT Communication Science Laboratories, Japan)
Abstract:Accurate sampling from continuous-time generative models can require many model evaluations. Low-step sampling reduces cost, but coarse updates introduce endpoint errors that can accumulate and degrade generation quality. We introduce TraceSampler, a post-training method that corrects these errors without modifying the pretrained model. Rather than optimising low-step updates directly, TraceSampler models endpoint error as a sample-specific prediction problem. A higher-step reference integration over the same interval provides the supervised target, while a compact representation makes high-dimensional error predictable from the sampling trace. This representation combines shared residual structure, trace-derived directions, and a low-rank remainder, so shallow per-step MLP heads predict sample-dependent coefficients. These coefficients drive a two-stage learned residual correction: the first stage predicts the full endpoint residual, while an intermediate backbone evaluation provides new trace information to predict the remaining residual; this evaluation is reused by the following step. TraceSampler is evaluated with Rectified Flow on CIFAR-10, SiT-XL/2 on ImageNet-256, and SD3.5 Medium for text-to-image generation. At 10 steps, it reduces trajectory MSE relative to matched-step Euler by 90.5% on CIFAR-10 and 85.6% on ImageNet-256, while achieving FID scores of 2.692 and 2.189, close to 120-step Heun references of 2.601 and 2.056 with 24$\times$ fewer backbone evaluations. On SD3.5 Medium, TraceSampler improves image quality and prompt following; at 10 steps, ImageReward rises from 0.752 to 0.906 while trajectory MSE falls by 70.7%. These gains preserve Euler's backbone-evaluation budget, add less than 0.4% inference FLOPs, and incur negligible wall-clock overhead, showing that coarse-step sampling error can be made predictable from the sampling trace and efficiently corrected.
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.14332 [cs.CV] |
| (or arXiv:2610.14332v1 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2610.14332 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: George Matthew [view email]
[v1]
Fri, 9 Oct 2026 21:18:54 UTC (6,444 KB)
来源：arXiv：cs.AI · arxiv.org