GeoGuide: Geometric guidance of diffusion models

Diffusion models have emerged as powerful tools for image generation, offering flexibility in generating images conditioned on specific classes or properties. Unlike GANs, diffusion models can be conditioned during training with relative ease. However, adapting pre-trained diffusion models to generate images from new, unlabeled data remains a significant challenge. The ADM-G approach addresses this by guiding diffusion models to generate images from a given class, but it often produces results of lower quality compared to models originally trained with class-specific conditioning. For instance, the ADM-G-guided model achieves an FID score nearly three times worse than that of a class-conditioned guidance. We identify that this performance gap arises partly because ADM-G provides minimal guidance during the final stages of the denoising process. To overcome this limitation, we introduce GeoGuide, a novel guidance method that improves the model's trajectory alignment with the data manifold. GeoGuide refines the backward denoising process by applying normalized adjustments to the model's output. Experimental results show that GeoGuide significantly outperforms ADM-G in both FID scores and the visual quality of the generated images.