Enjoying my work? Consider supporting me with a coffee!
Introduction
Furthermore is a LoRA created by merging two detail-enhancing models (listed below) and customizing individual block weights until achieving the desired result. It was initially designed for furry art, but with enough tweaking I was able to make it function with most checkpoints.
My issue with other detail LoRAs is entirely personal; they simply don't accomplish the specific result I wanted to see. Additionally, none of the detailers I've worked with produced images that I felt were universally detailed. For example, More Details was great with objects and structures but struggled with fur and character organics, while Furtastic was the opposite. This was forcing me to choose between which group of elements I wanted to improve.
I did discover that despite their individual shortcomings, both models worked decently together when using a very specific set of strengths (More Details at 0.5 combined with Furtastic at 0.1). This was the inspiration for merging them into a LoRA with customized block weights.
Furthermore aims to accomplish these goals:
Add details while being minimally destructive.
Fix structure that could produce artifacts or deformities.
Improve both organic and non-organic details.
Work with a wide variety of styles and checkpoints.
Function as intended at a strength of 1.0 with little need for dampening.
Capable of being used in txt2img and img2img depending on needs.
Dramatically improve backgrounds, subtly improve character features.
This project is a work in progress. Please feel free to contact me with suggestions, I'm always open to feedback.
Suggested Use
Furthermore should work at a strength of 1.0 with no need for dampening, but you can lower this number as you see fit. Since the model's weights have been adjusted by hand it shouldn't cause problems. Strengths between 0.6 and 1.0 will add noticeable details, while 0.5 is a "safe zone" that only makes minor changes.
Most samplers work fine, but I have had the most success with DPM++ 2M Karras (potentially biased, I used it for a lot of testing). I also highly recommend you use Hires. fix with the Latent (nearest-exact) upscaler. Pure Latent can work as well, especially if your image is borked and you want to fix without regenerating the seed, but I've found it deviates a lot. If you want to generate without Hires Fix, make sure you add a few steps to compensate.
If you're having trouble with quality try lowering the LoRA strength. I know that I said it worked perfectly at 1.0, but some seeds are just bad. A strength between 0.6 and 0.9 will fix most problems while also keeping the new details we all love. If something is truly broken, try 0.5.
You can also use this detailer in img2img. I recommend starting with a denoise strength of 0.4. You can increase it to 0.45 for a moderate boost in details at the cost of small changes in accuracy, but anything above 0.5 will likely make sweeping changes to your base image (which is fine, if you're okay with it).
Don't be afraid to experiment, part of the fun is being brave and exploring. If you find anything cool let me know, enjoy!
Credits
Merged with: More Details and Furtastic