Simple workflow with ControlNet tiled upscale

This is my main workflow for SDXL/Pony/Illustrious that uses ControlNet tile upscaling instead of Ultimate SD. Feel free to DM me if you have any questions about the workflow.

v1.2 update

• Added CFG Skimming, after more testing I think it's an interesting tool. Can be easily disabled in generation settings options (enable_skimming). Change CFG to recommended checkpoint values when you disable it.

• Saved images should now be compatible with Civitai. I was getting tired of filling some fields manually. Requires manual installation of Save Images With Metadata Universal Extension

• Combined generation settings into a single subgraph
  

• For MetaData saver to work properly, run Rules Scanner Subgraph once after installation or when you change nodes in ComfyUI (I don't make the rules)
  

• By default set tile_batch_size to 1 in HiRes subgraph to safely run on <12GB VRAM. If you have more than 12GB, set it to 2 (can be even higher with 24GB, but I won't recommend it).

• Removed redundant switch for Adetailer.

v1.1 update

• Replaced Lora Loaders to non-Clip ones

• Replaced KSampler to default ComfyUI node

• Cleaned up wiring

• Promoted additonal options in HiRes subgraph for easier access

  • Initial upscale model selector (i.e. Remacri/UltraSharp/ESRGAN)

  • ControlNet tile checkpoint selector

  • tile batch size (higher values increase speed at the cost of VRAM)

  • blend factor (color correction)

• Also added some options in the face detailer subgraph

• Moved nodes closer together and grouped them for easier bypass

Features

• Slower than Ultimate SD upscaler but produces better quality in my opinion

• Allows for higher denoise without adding weird details

• Easily switch HiRes and FaceDetailer subgraphs on and off

• Save upscaled output as PNG and JPG, I usually post JPG images for faster loading. PNG versions will include workflow in a metadata.

• Can be VRAM hungry, only tested it on my 12GB card

• Uses additional ControlNet tile model:

  • Realistic version, but it works well even for stylized outputs - good results with fabric, fur and skin texture, but tends to oversharpen hair (can play with strength settings to fix this): https://huggingface.co/TTPlanet/TTPLanet_SDXL_Controlnet_Tile_Realistic

  • Alternative model for anime and stylized models - sharper and cleaner details: https://huggingface.co/xinsir/controlnet-union-sdxl-1.0

  • Put these models into ComfyUI/models/controlnet folder

• KSampler can be replaced with a ComfyUI one, I'm using Inspire's node for different noise generation options.

Control upscale size and denoise in the HiRes subgraph

For higher upscale values (3-4x) use lower denoise (0.2 or even lower).

Custom Adetailer prompt

In Use custom prompt node set value to true and type positive/negative prompts which will be used for faces. Userful when some Loras overpower Adetailer and change face color too much, or if you just want to adjust face expression. With additional nodes you can set it up to use different Loras during this step. Increase feather value if you get visible edges around faces.

Low VRAM

To save VRAM inside a HiRes Upscale subgraph try this:

1. In Tiled Diffustion node Set batch size to 1. This will reduce VRAM load considerably at the cost of speed.

2. Set VAE Decode tile size to 512.

3. If it's still VRAM hungry, try setting tile width/height in Tiled Diffusion and VAE encode to 512x512. Might want to change overlap as well, but I haven't tested this.