The paper "Guiding a Diffusion Model with a Bad Version of Itself" by Tero Karras explores the challenges and improvements in image generation using diffusion models. The authors focus on the trade-offs between image quality, variation, and alignment with given conditions, such as class labels or text prompts. They observe that classifier-free guidance (CFG) improves image quality by focusing on high-probability regions but reduces variation. To address this, they propose a novel method called *autoguidance*, which uses a smaller, less-trained version of the model to guide the generation process. This approach allows for better control over image quality without compromising variation. The method is validated through various synthetic and practical tests, achieving record-breaking FID scores for ImageNet generation. Additionally, the method is applicable to unconditional diffusion models, significantly improving their quality. The paper also includes a detailed analysis of why CFG improves image quality and provides insights into the mechanisms behind the proposed autoguidance technique.The paper "Guiding a Diffusion Model with a Bad Version of Itself" by Tero Karras explores the challenges and improvements in image generation using diffusion models. The authors focus on the trade-offs between image quality, variation, and alignment with given conditions, such as class labels or text prompts. They observe that classifier-free guidance (CFG) improves image quality by focusing on high-probability regions but reduces variation. To address this, they propose a novel method called *autoguidance*, which uses a smaller, less-trained version of the model to guide the generation process. This approach allows for better control over image quality without compromising variation. The method is validated through various synthetic and practical tests, achieving record-breaking FID scores for ImageNet generation. Additionally, the method is applicable to unconditional diffusion models, significantly improving their quality. The paper also includes a detailed analysis of why CFG improves image quality and provides insights into the mechanisms behind the proposed autoguidance technique.