Scarlett's Mix [2b2157a1]

Heads Up!

This mix can not be completely recreated by the Checkpoint Merger panel in the AUTOMATIC1111 webui. You must use the previous merge model script instead.

Download

https://mega.nz/folder/IOVHkSQC#27H1hODY_b9btNin21uZSg
sha256sum ceb5e3160c9c01ab7c14b4611970db744e363e153dd95938ec7ebd630688dc6b

Assembly Instructions

Be sure to read the instructions on how to use merge.py.

Requirements

Stage 1

Take HD16 and merge it with F111 at 50%.

python3 merge.py path/to/hd_16.ckpt path/to/f111.ckpt --alpha 0.5 --output path/to/stage1

The script automatically appeds the alpha value and ".ckpt" to your output file name. The resulting file should be named stage1-50.ckpt and have a model hash of 86a12278.

Stage 2

Take Stage 1 and merge r34_e4 into it at 33.33%.
(This is where Checkpoint Merger currently fails. It does not accept 0.333 as a valid value or is doing something else weird or wrong. I may investigate.)

python3 merge.py path/to/stage1-50.ckpt path/to/r34_e4.ckpt --alpha 0.333 --output path/to/stage2

The resulting file should be named stage2-3330.ckpt and have a model hash of 4acb41e9.

Stage 3

Take Stage 2 and merge trinart2_115k into it at 25%.

python3 merge.py path/to/stage2-3330.ckpt path/to/trinart2_115000.ckpt --alpha 0.25 --output path/to/stage3

The resulting file should be named stage3-250.ckpt and have a model hash of 189fe4d9.

Stage 4 (Final)

Then Stage 3 and mix NAI NSFW into it at 70%.

python3 merge.py path/to/stage3-250.ckpt path/to/nai_nsfw.ckpt --alpha 0.25 --output path/to/scarlett_mix

The resulting file should be named scarlett_mix-70.ckpt and have a model hash of 2b2157a1.

Checksums

1
2
3
4
5dd2eb106d439890d8547a857041af60b75ffec5ab289041e1bd13d63570e1b1  stage1-50.ckpt
5aa7c3d42b20242350ad1c3153855f37e6c9b529fb6fb181741eacddd9ea5839  stage2-3330.ckpt
98336aaf36fa97fa9e961e2ce4bef82814b05958c31b7ef2283c59787cf179ce  stage3-250.ckpt
ceb5e3160c9c01ab7c14b4611970db744e363e153dd95938ec7ebd630688dc6b  scarlett_mix-70.ckpt

Checkpoint files are zip archives. Minor variances in the data may produce a wildly different checksum. If the sha256sum of your model does not match, but the model hash matches, you're probably fine. Probably.

Effective Prompting (with bonus Chinese Magic)

(work in progress)
Magic Negative Prompt:

duplicate, morbid, mutilated, mutated, poorly drawn, blurry, disfigured, text, error, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, long neck, long body, mutation, poorly drawn, malformed

Example Prompts and Results

(work in progress)

Edit
Pub: 29 Oct 2022 02:58 UTC
Edit: 29 Oct 2022 08:37 UTC
Views: 4355