My goal was to generate some nice anime-like images with decent details (on SD 1.5 models).

This is my end result after trying out various custom nodes.

Nothing too crazy here.

1. The start is standard with the initial image generation to upscaling to twice the resolution.

2. From there, it goes through a series of detailers.

3. One cycle in a person detailer, one cycle in a face detailer, and 2 cycles in an eye detailer.

4. After that, into Ultimate SD Upscale for the final upscale to 4 times the original size.

My workflow has a few custom nodes from the following:

Impact Pack (for detailers)

Ultimate SD Upscale (for final upscale)

Crystools (for progress and resource meters)

ComfyUI Image Saver (to show all resources when uploading images to CivitAI) - Added in v2

In addition to those four, I also use an eye detailer model designed for adetailer to use with the FaceDetailer node from Impact Pack. Eyes have been a problem for me with many models, so this in particular is a godsend for me.

The denoise numbers in the detailers have to be adjusted for the eyes and faces sometimes. I usually use 0.4 to 0.5.

Also, I am using the FreeUv2 node. This is a node that is included with comfyUI. If you are not able to find it, you may need to update your comfyUI.

Updated on March 4, 2024 (v2):

I started using ControlNet nodes in this version.

For the detailers I use: control_v11p_sd15_normalbae.pth

For Ultimate SD Upscale I use: control_v11f1e_sd15_tile.pth

It seems to give me better results, but could just be copium.

I am also using a different eye detailer model. The link above has been adjusted. This model details both eyes at once. This has led to better results for me.

On average, it takes me about 5 to 6 minutes from start to finish on a 12GB 3060.