Make Good Pic XL - comfy workflow

New in v1.3:

Great improvement in generation speed for hi-res images:
- 25 seconds per 2K image
- 55 seconds per 4K image
  (on 4090)
Fewer Custom Nodes Used
ToDo:
- Find way to reduce prompt (color) bleed without slowing down generation or requiring per-prompt fine-tuning (like cut-off does).
- Further improve image quality at higher resolutions.

New in v1.2:

2K-4K high quality upscale using Mikey's custom Tiled Sampler.
- On 4090 I get:
  - 30-40s for batch of 2x1K base images
  - 60-80s to upscale both images to 2K or 4K
    (100-120s total)
  - For now, I focused on highest quality, so generation time can probably be optimized.
Re-wrote much of the workflow to use fewer custom nodes.
Removed custom KSampler, using Mikey's instead.
To Do: add workflow inline comments and tips.
Notes:
- It's possible Mikey's tiled sampler can be used on its own using noise instead of base image, this could potentially greatly speed up generation time of >2K images, but initial tests still unconclusive.
- Use your preferred text-debug node if you want to see the prompts being auto-generated.
- This workflow is intended to be open and easy to understand - using clear layout and colors with all options easily exposed in their native form.
  This makes it very easy to "reverse-engineer" and customize to any use case.

New in v1.1:

More control over generated images.
Better and more consistent image quality.
Simpler to use.
Notes added inline for better usability.
Simple yet effective layout with everything exposed and explained, no secrets!
It really does produce consistently good images - makes good use of the generate-forever checkbox.

SDXL is like old amateur radio - need to find a good frequency among the static. Usually, you can tell it is good when the workflow responds well to small changes in prompt or settings.
This workflow is such a "good frequency", with strong reception - it produces consistently good results.

Some notes about the workflow features and my experience with it and with SDXL so far:

Mix different XL base models on the fly - very powerful feature.
Add any Loras - my experience is that in most cases even very low rates like 0.2 and 0.4 are noticeable.
Giving a base prompt and then adding SDXL Auto-Prompter prompt can add a lot of variation with surprising and often pleasing results.
Generation time for me is 42 secs for batch of 2, on RTX4090.
Using DPM SDE Exponential for sampler is slow but produces best results in my experience, using it with low steps count gives better results than samplers with much higher steps count.
Upscaling using SD always loses detail for me so i don't do it, 3rd party commercial software does it better.
NSFW is no issue - ie no twisted nipples even with refiner.
Some models are better for NSFW - in my experience AltXL is the current leader in this category.