SR-Alpha aims to combine comic art style with semi-realistic textures.
I highly recommend inpainting for best results and to unlock the full potential of the merge. Separate inpainting model not required.
To help with inpainting human features, I have also been tinkering with this extension called !After Detailer, which automatically detects and "fixes" the portion of the image being generated via txt2img depending on the "ADetailer" model used, notably the models for the hands and face which I use most often.
PS:
I no longer remember the exact models used for versions 1-1.2. I believe that with the new update, the old versions would be rendered obsolete.
V2 and V3 Mix Recipe:
Recommended Settings:
V 1.0:
Sampling Method: DPM++ SDE Karras
Sampling Steps: 50
V1.1 and above:
Sampling Method: DPM++ 2M Karras, Euler A (Inpainting)
Sampling Steps: 20-30
CFG = 7-10
Denoising Strength = 0.4-0.5 (general), 0.2-0.3 (inpainting hands)
Workflow (used in V3 samples):
txt2img
>Initial dimensions 512x615 (WxH)
>Hi-res fix by 1.25x to get 640x768 dimensions>Adetailer enabled using either 'face_yolov8n' or 'mediapipe_face_mesh'
inpainting
>Inpaint areas to improve quality. Do it part by part so that focused area will be altered with best results.
Upscaling (optional): Upscale images by 2.25x
Method A.:
Upscaler 1 (U1) = Lanczos
U2 = R-ESRGAN General WDN 4xV3
U2 Visibility = 0.5
GFPGAN visibility = 0.3
Method B.:
U1 = 4xNMKDSuperscale_4xNMKDSuperscale
Suggested Embeddings:
veryBadImageNegative - veryBadImageNegative_v1.3 | Stable Diffusion Embedding | Civitai
https://cdn.discordapp.com/attachments/1032948846197747731/1069660323709190195/bad-hands-5.pt
Deep Negative V1.x - V1 75T | Stable Diffusion Textual Inversion | Civitai
Bad artist Negative embedding - Bad artist | Stable Diffusion Textual Inversion | Civitai
Async's Negative - v12 | Stable Diffusion Embedding | Civitai
SimpleNegative | Negative TI - V3 | Stable Diffusion Embedding | Civitai
VAE:
Credits to the respective original creators of the resources used.