1.1 version

The dataset: I choose 160 images generated from NoobAI-XL epsilion-pred-0.5 and Illustrious x Pony Mix with this Lora (1.0 version), then adjusting them.

The 1.1 version trained on NoobAI-XL Epsilon 1.1 version

All sample of images are generated by NoobAI-XL epsilion-pred-1.1 version and upscaled by R-ESRGAN 4x+ Anime6B.

Settings I use:

Sample: Euler a (I recommend Sinusoidal SF Schedule type)

Steps: 24-30

CFG Scale: 3.5-10

Clip Skip: 2

Negative Prompts: displeasing, low quality, worst quality, realistic, bad hands, bad eyes, bad mouth, overexposure, fused pupils

Upscale: R-ESRGAN 4x+ Anime6B

Resolution: 1104*1380, 1380*1104, 1448*1024, 1024*1448

1.0 version

The dataset: I choose 128 images generated from HotaruLabs models with some Lora, then adjusting the brightness, saturation and contrast of these images.

This version trained on Illustrious-XL works well in NoobAI-XL Epsilon 0.5 version

All sample of images are generated by NoobAI-XL epsilion-pred-0.5 version and upscaled by R-ESRGAN 4x+ Anime6B.

Settings I use:

Sample: Euler a or DPM++ 2M

Steps: 24-30

CFG Scale: 6-9

Clip Skip: 2

Negative Prompts: bad quality, low quality, worst quality, bad hands, fused pupils, fused fingers(optional)

Upscale: R-ESRGAN 4x+ Anime6B

Resolution: 1104*1380, 1380*1104, 1448*1024, 1024*1448