notproniss/pixelanchoredremaster

Pixel Anchored Remaster is a custom node for ComfyUI, the popular AI image generation tool. It is designed to improve images that have already been through a high-resolution upscaling pass, by running them through a three-step cleanup and rebuild process. The goal is a better-looking final image, not a pixel-perfect copy of the input. The process works like this: first, the image is downscaled slightly in regular pixel space, which removes some of the harsh artifacts and noisy textures that come from aggressive upscaling. Second, the cleaned-up image is encoded into the internal format the AI model works with (called latent space), and then scaled back up to the original dimensions. Third, the AI sampler is run to rebuild fine details, reinterpreting the image rather than simply stretching it. The node is built for 2x upscaling workflows only. The author found that 4x runs produce unreliable results, and explicitly scoped the tool to 2x. It is not a replacement for a high-resolution fix pass, and it is not a tool for preserving every edge and texture exactly. Some small visual drift is expected and considered acceptable if the final image looks better overall. Installation is simple: drop the folder into the ComfyUI custom_nodes directory and restart ComfyUI. The node takes one image in and gives one image out. It includes controls for the sampling method, scheduler, sampling strength, and tiled VAE settings, though default settings are reported to work well for most cases. The node works best with SDXL, Pony, and Illustrious-style model families. The README specifically calls out good results with anime-style models, especially portraits. It performs poorly with some other model architectures such as Anima and Chroma, so those are not recommended use cases.