Jiaye Li

Hui Li, Baoyou Chen, Liwei Zhang et al.

Diffusion transformers (DiTs) adopt Patchify, mapping patch representations to token representations through linear projections, to adjust the number of tokens input to DiT blocks and thus the computation cost. Instead of a single patch size for all the timesteps, we introduce a Pyramidal Patchification Flow (PPFlow) approach: Large patch sizes are used for high noise timesteps and small patch sizes for low noise timesteps; Linear projections are learned for each patch size; and Unpatchify is accordingly modified. Unlike Pyramidal Flow, our approach operates over full latent representations other than pyramid representations, and adopts the normal denoising process without requiring the renoising trick. We demonstrate the effectiveness of our approach through two training manners. Training from scratch achieves a $1.6\times$ ($2.0\times$) inference speed over SiT-B/2 for 2-level (3-level) pyramid patchification with slightly lower training FLOPs and similar image generation performance. Training from pretrained normal DiTs achieves even better performance with small training time. The code and checkpoint are at https://github.com/fudan-generative-vision/PPFlow.

Wenzhen Zhang, Debo Cheng, Guangquan Lu et al.

Deep neural networks are highly susceptible to overfitting noisy labels, which leads to degraded performance. Existing methods address this issue by employing manually defined criteria, aiming to achieve optimal partitioning in each iteration to avoid fitting noisy labels while thoroughly learning clean samples. However, this often results in overly complex and difficult-to-train models. To address this issue, we decouple the tasks of avoiding fitting incorrect labels and thoroughly learning clean samples and propose a simple yet highly applicable method called Adaptive Label Refinement (ALR). First, inspired by label refurbishment techniques, we update the original hard labels to soft labels using the model's predictions to reduce the risk of fitting incorrect labels. Then, by introducing the entropy loss, we gradually `harden' the high-confidence soft labels, guiding the model to better learn from clean samples. This approach is simple and efficient, requiring no prior knowledge of noise or auxiliary datasets, making it more accessible compared to existing methods. We validate ALR's effectiveness through experiments on benchmark datasets with artificial label noise (CIFAR-10/100) and real-world datasets with inherent noise (ANIMAL-10N, Clothing1M, WebVision). The results show that ALR outperforms state-of-the-art methods.