Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution

Artifact-free super-resolution (SR) aims to translate low-resolution imagesinto their high-resolution counterparts with a strict integrity of the originalcontent, eliminating any distortions or synthetic details. While traditionaldiffusion-based SR techniques have demonstrated remarkable abilities to enhanceimage detail, they are prone to artifact introduction during iterativeprocedures. Such artifacts, ranging from trivial noise to unauthentic textures,deviate from the true structure of the source image, thus challenging theintegrity of the super-resolution process. In this work, we proposeSelf-Adaptive Reality-Guided Diffusion (SARGD), a training-free method thatdelves into the latent space to effectively identify and mitigate thepropagation of artifacts. Our SARGD begins by using an artifact detector toidentify implausible pixels, creating a binary mask that highlights artifacts.Following this, the Reality Guidance Refinement (RGR) process refines artifactsby integrating this mask with realistic latent representations, improvingalignment with the original image. Nonetheless, initial realistic-latentrepresentations from lower-quality images result in over-smoothing in the finaloutput. To address this, we introduce a Self-Adaptive Guidance (SAG) mechanism.It dynamically computes a reality score, enhancing the sharpness of therealistic latent. These alternating mechanisms collectively achieveartifact-free super-resolution. Extensive experiments demonstrate thesuperiority of our method, delivering detailed artifact-free high-resolutionimages while reducing sampling steps by 2X. We release our code athttps://github.com/ProAirVerse/Self-Adaptive-Guidance-Diffusion.git.