# Back to Basics: Let Denoising Generative Models Denoise

**Authors:** Tianhong Li, Kaiming He  
**arXiv:** [2511.13720](https://arxiv.org/abs/2511.13720)  
**Date:** 2025-11-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2511.13720v1/

## Key points

- Authors Tianhong Li and Kaiming He address limitations of current denoising diffusion models
- Current models focus on predicting noise or a noised quantity, not clean images
- Distinguishing between noise and natural data is crucial in generative modeling tasks
- Proposed approach by Li and He involves models directly predicting clean data
- Simple large-patch Transformers are used as strong generative models without tokenizers or pre-training
- Approach, called "Just image Transformers" (JiT), achieves competitive results on ImageNet at resolutions of 256 and 512 with patch sizes of 16 and 32
- Research focuses on mapping back to basics of the manifold assumption for Transformer-based diffusion on raw natural data
- JiT framework driven solely by plain Transformers without additional losses or pre-training, making it computationally efficient and scalable across different resolutions

## Summary

In their paper "Back to Basics: Let Denoising Generative Models Denoise," authors Tianhong Li and Kaiming He address the limitations of current denoising diffusion models. These models do not directly predict clean images but instead focus on predicting noise or a noised quantity. The authors argue that predicting clean data and predicting noised quantities are fundamentally different tasks. They highlight the importance of distinguishing between noise and natural data in generative modeling tasks. The proposed approach by Li and He is a novel one where models directly predict clean data. This enables apparently under-capacity networks to effectively operate in high-dimensional spaces. The authors demonstrate the effectiveness of simple large-patch Transformers on pixels as strong generative models without the need for tokenizers, pre-training, or extra loss functions. This approach, dubbed "Just image Transformers" (JiT), yields competitive results on ImageNet at resolutions of 256 and 512 using patch sizes of 16 and 32. By focusing on mapping back to the basics of the manifold assumption, this research offers a self-contained paradigm for Transformer-based diffusion on raw natural data. The authors underscore that their approach is driven solely by plain Transformers without additional losses or pre-training, making it computationally efficient and scalable across different resolutions. The paper also provides qualitative results showcasing the efficacy of JiT-H/32 on ImageNet 512×512 images and compares its performance with previous methods. In conclusion, this study sheds light on the significance of addressing noise differently from natural data in generative modeling tasks. By leveraging basic principles and avoiding unnecessary complexities, the proposed JiT framework offers a promising direction for advancing diffusion models in high-dimensional spaces without compromising computational efficiency.