MixFlow 训练：以减速插值混合缓解暴露偏差
MixFlow Training: Alleviating Exposure Bias with Slowed Interpolation Mixture
针对扩散模型训练-测试不一致（暴露偏差）问题，提出用「减速时间步」上的真实插值混合来后训练预测网络，可直接套用到 SiT、REPA、RAE 等扩散/流匹配骨干上，在 ImageNet 256/512 上把 FID 压到 1.1 左右，是提升现有生成模型采样质量的即插即用后训练手段。
原文依据 MixFlow leverages the interpolations at the slowed timesteps, named slowed interpolation mixture, for post-training the prediction network at each training timestep.
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论本文研究用于改进扩散模型的训练-测试不一致（又称暴露偏差）问题。在训练时，预测网络在训练时间步的输入是对应真值噪声数据，即噪声与数据的插值；而在测试时，输入是生成的噪声数据。我们提出一种新的训练方法 MixFlow，以提升训练性能。该方法的动机来自慢流现象：在给定采样时间步上，与生成的噪声数据最接近的真值插值被观察到对应于一个噪声更高的时间步（称为减速时间步），即对应的真值时间步比采样时间步更慢。MixFlow 利用这些减速时间步上的插值（称为减速插值混合），在每个训练时间步对预测网络进行后训练。在类条件图像生成（包括 SiT、REPA 和 RAE）与文本到图像生成上的实验验证了该方法的有效性。我们的 MixFlow 在 RAE 模型上于 ImageNet 取得强劲的生成结果：256×256 下为 1.43 FID（无引导）和 1.10（有引导），512×512 下为 1.55 FID（无引导）和 1.10（有引导）。
英文原文摘要 1316 字符
This paper studies the training-testing discrepancy (a.k.a. exposure bias) problem for improving the diffusion models. During training, the input of a prediction network at the training timestep is the corresponding ground-truth noisy data that is an interpolation of the noise and the data, and during testing, the input is the generated noisy data. We present a novel training approach, named MixFlow, for improving the training performance. Our approach is motivated by the Slow Flow phenomenon: the ground-truth interpolation that is the nearest to the generated noisy data at a given sampling timestep is observed to correspond to a higher-noise timestep (termed slowed timestep), i.e., the corresponding ground-truth timestep is slower than the sampling timestep. MixFlow leverages the interpolations at the slowed timesteps, named slowed interpolation mixture, for post-training the prediction network at each training timestep. Experiments over class-conditional image generation (including SiT, REPA, and RAE) and text-to-image generation, validate the effectiveness of our approach. Our approach MixFlow over the RAE models achieve strong generation results on ImageNet: 1.43 FID (without guidance) and 1.10 (with guidance) at 256 x256, and 1.55 FID (without guidance) and 1.10 (with guidance) at 512 x512.