HR Beautify ComfyUI workflow - minsky91/HR-Beautify-workflow GitHub Wiki

HR Beautify ComfyUI workflow (SDXL): refine after upscale and go as high as 24K

HR Beautify Comfy workflow screenshot May 2025

HR Beautify (‘HR’ stands for Hires, or High resolution) is an advanced ComfyUI workflow to refine images after they have been upscaled, optionally beautifying them with a style transfer. Why such a narrow purpose range? For more than a year now, I have been experimenting with upscaling and enhancing hires images and came to the firm conclusion that, if one wants the best results possible, the refine or enhance step must be separated from the upscale one and have an entire workflow dedicated to it. And, as you can see from the screenshot, it turned out to be no small feat to accomplish.

The workflow includes two ControlNets, for the most powerful and flexible img2img guidance, as well as two IP Adapters, to enhance images with style transfer and composition or reference guidance. The ControlNets can be used simultaneously and can assist in processing images of up to 24K resolution (provided that all Comfy’s cappings on the image size are removed: a LINK to the page with instructions will be soon provided). The IP adapters, on the other hand, are best used at lower-to-mid resolutions (up to 4K).

The 1st of the ControlNets included (the Advanced version) can use its own control image and mask, while either of the IP Adapters can use its own Style, Reference or Composition image (you will need to unmute and/or expand these nodes before dropping input images into them). Note that mask images influence these components’ output in a different than regular clipping masks way, and they are called attention masks. Union ControlNets are also supported. (*)

For extra refinement power, the Power Lora Loader node is included. With it, you can add a potentially unlimited number of detail or style LoRas (though personally, I only used LCM and Hyper LoRas regularly, to enable corresponding samplers for non-Lightning checkpoints).

The tiling component used for hires refining is Tiled Diffusion (TD). At the time of writing, it is still in the beta testing stage, but I’ve found it to work just fine for all use cases I tried it with. Because of its tiling logic, TD makes it possible to process images of really large sizes (**), and very fast too. Also, due to the sophisticated denoising algorithm which it uses, TD can enhance images without leaving tile seam artefacts that plague output of other tiled upscaling components such as Ultimate SD Upscale, not to mention the additional ‘creative touch’ TD is so famous for. Click here to see examples of hires images produced with HR Beautify.

Note: of the 3 methods supported by TD, I recommend the Mixture of Diffusers one as the default choice, as it’s the most ‘creative’ and least artefact-prone of them all. Textures generated with it may appear slightly softer in comparison though. (If you still see tile seams, increase the tile overlap value.)

Also included in the workflow are three post processing components: enhancing image’s HDR range, advanced image sharpening and color-matching, all powered by dedicated custom nodes. In the shared version of the workflow, only color-matching is enabled by default. This has to do with the fact that img2img processing, when used with high denoise value, often introduces a color cast, usually a red tint, which is normally very hard to avoid or remove once it’s there. The other 2 options are of lesser importance, but can be easily enabled by setting a non-zero strength (via the ‘a’ parameter).

As mentioned above, the HR Beautify workflow was designed without an AI upscale model node on purpose. Still, I did include a node in it called Resolution multiplier that you can enable to have the image resized before processing. Since it uses the classical non-AI Lanczos method, you can only go as high as 2x with it. If used with a value above 1, it will increase working resolution of the image before refining, an internal sub-step that often may help to add extra detail to the output (but will also make the processing time substantially longer). By default, the image will be scaled back to the initial size after refinement and before post-processing; if you want to keep the output image size at the increased resolution, set the Upscale mode toggle above Resolution multiplier to True.

In contrast, Resolution multiplier values below 1.0, with the minimum of 0.5, are used in cases when your image resolution is above the checkpoint’s native one (1024x1024 for SDXL, for instance), and you see distortions in the output that are likely associated with such a resolution. Setting the multiplier to 0.7 or 0.5 decreases the working resolution closer to the native one and often allows to avoid distortions and other artefacts, by the price of slight detail loss. Alternatively, decrease the denoise strength (‘value’) in KSampler Advanced to avoid distortions.

To use the workflow, you’ll need to have the following custom nodes installed (in the order of their importance):

(*) Note that having all 4 guidance components (ControlNets & IP Adapters) enabled simultaneously is not a good idea, since this can lead to out-of-memory conditions or affect the output badly, particularly when used with hires images. The most efficient use scenario is to have only 2 of them enabled at once (one ControlNet and one IP Adapter, or two ControlNets), or 3 at most, with low-to-medium image resolutions.

(**) The workflow has been tested and extensively used on a Windows 11 PC equipped with an NVIDIA GeForce RTX 4070 Ti SUPER GPU with 16 GB VRAM. For a setup with a smaller amount of VRAM on the GPU, additional tweaking of the Tiled Diffusion parameters might be necessary, in particular the tile width / height and the tile batch size, to avoid Out of Memory or Allocation on Device failures of the Comfy server.