Image Generation Network Combining High-efficiency Residual and Self-conditioning

In the field of image generation, deep - learning methods based on diffusion models have achieved remarkable progress. However, when dealing with images of Thangka ritual implements, which have fine textures and unique shapes, existing models suffer from slow inference speed, as well as deficiencies in texture clarity and pattern integrity.To address these issues, based on the DDIM, we propose an image - generation network that combines the high - efficiency residual and self - conditioning. This network utilizes the ConvNext v2 Block to enhance the ability to extract and utilize image features. Meanwhile, the self conditioning is introduced to enable the model to better utilize its own existing information during the generation process.Experimental results show that, on the Thangka ritual implement dataset, compared with the original DDIM, the method proposed in this paper reduces the FID by 1.169, increases the IS by 0.23, and improves the subjective evaluation score by 6.2%.