This paper introduces the Image Processing Transformer (IPT), a pre-trained model based on the transformer architecture for low-level image processing tasks such as super-resolution, denoising, and deraining. The IPT model is designed with multiple heads and tails to handle various tasks, along with a shared transformer body. To maximize the transformer's capabilities, the model is trained on a large-scale dataset generated from the ImageNet benchmark, where images are corrupted using various degradation methods. The model also incorporates contrastive learning to adapt to different image processing tasks. The pre-trained IPT model outperforms state-of-the-art methods on multiple benchmarks, demonstrating its effectiveness after fine-tuning. The model is trained using a combination of supervised and self-supervised learning, with the contrastive loss helping to learn universal features. Experimental results show that the IPT model achieves superior performance in tasks such as super-resolution, denoising, and deraining, and exhibits strong generalization ability across different tasks. The model's architecture is detailed, and its performance is evaluated on various datasets, showing that it outperforms traditional CNN-based models. The paper also discusses the impact of data percentage, contrastive learning, and multi-task training on the model's performance. The results indicate that the IPT model is effective for pre-training and can be fine-tuned for various image processing tasks.This paper introduces the Image Processing Transformer (IPT), a pre-trained model based on the transformer architecture for low-level image processing tasks such as super-resolution, denoising, and deraining. The IPT model is designed with multiple heads and tails to handle various tasks, along with a shared transformer body. To maximize the transformer's capabilities, the model is trained on a large-scale dataset generated from the ImageNet benchmark, where images are corrupted using various degradation methods. The model also incorporates contrastive learning to adapt to different image processing tasks. The pre-trained IPT model outperforms state-of-the-art methods on multiple benchmarks, demonstrating its effectiveness after fine-tuning. The model is trained using a combination of supervised and self-supervised learning, with the contrastive loss helping to learn universal features. Experimental results show that the IPT model achieves superior performance in tasks such as super-resolution, denoising, and deraining, and exhibits strong generalization ability across different tasks. The model's architecture is detailed, and its performance is evaluated on various datasets, showing that it outperforms traditional CNN-based models. The paper also discusses the impact of data percentage, contrastive learning, and multi-task training on the model's performance. The results indicate that the IPT model is effective for pre-training and can be fine-tuned for various image processing tasks.