Custom Workflow for Krita's AI Diffusion Plugin. Works directly from the visible canvas. Upscale, Enhance the image, Downscale to same size. Can be used as normal ComfyUI workflow with a bypass, just connect it like you need it. There 2 two groups that work in sequence, so you do not need to use the Krita (imbedded custom workflow) if you want to use a part of this idea in another comfy workflow. ComfyUI Manager recommended to get missing nodes, as they usually not part of the Local installation of Krita's Comfy backend.
Following additional nodes are used, (Upscaler), (JKNodes), (WASNodes), (EasyUse), (CustomScripts) for Math, [(RGthree) Downscale], and of course the Krita Required custom nodes.
What the benefit? Better details, included eye fix without enlarging the image. Also works great on already upscaled ones. At least i have better results as the build in Krita Upscaler feature (it noises a lot), without destroying already present details. At the standard values are discovered by experimentation, and seems to do for my taste. At least you do not need to sharpen afterwards.
For multiple runs it can over sharpen or darken the image, as the upscale works chunk by chunk, it recommended to adjust the settings or in post to counteract this. The default 'seam' parameters work fine for the most cases, but it can produce chunk shadows on larger upscales values. Larger Upscale values works better for indicate fine details, while higher steps smoothen details. Overall works best with 30-50% denoise, depends on the picture/prompts. The Result blend very good with the source, and can be masked for fine control.
Add a revision V2 with some changes, see changelog. V1 is still good for oneshot 'txt to img', V2 is better suited for 'img to img' as its softer in the defaults.
V3 has some needed changes; see changelog. Mirror and description of the Text to Img workflow can be found on Deviant Art. Happy Drifting.
Note: If you have problems with losing Color, saturation or brightness, use another VAE.
Description
Change Default Values, to reduce the over sharpen effect. As the last pass can be done in post.
Pass 1 Edge: 20->15, Pass 1 Sharpness: 60->30, Pass 2 Sharpness: 30->20.
Expose New Parameters: Seamfix, Seamfix Denoise. It might help at larger Upscales to avoid seams.
Upscale VAE (Some Build in VAE mud the output)
Sort the Settings.
Add a simple "txt to img" workflow with a compare viewer. Check it yourself.
2.1 forget to expose downscale method: Lanczos (optimal), bicubic (softer), bilinear (softest), nearest (hardest, pixelart).
2.2: Add a divider for Tilesize, that is usual the image size. It useful to reduce the tilesize on a larger image to stay in range of the checkpoint resolution. It it useful when you work with higher then 2048x2048 (already upscaled). Usually its 4 tiles (with 2x upscale) with the default '1', with '2', smaller ones, 16 Tiles and so on. Note: Float math can produce rounding errors, if not using half numbers. Note: More tiles mean more borders, if you use the seam fix; it add more work, but can safe some memory in cost of time. It makes possible to process extrem images on smaller vram.