Face Restoration XL

Mainly used for face restoration, this is not a simple merge, but adds some images to the excellent checkpoints to address some of the pain points of face quality (img2img).

CFG 6-6.5

dpm++2m/dpm++2m_sde + Karras(It makes the skin look less smooth and more detailed.)

dimm(It produces perfect, not-so-natural looking skin.)

It responds positively to the following prompt words:

1. features

[Iranian/Arab/European/Asian/Indian/African/Eurasian] features

2. expression

pain, aroused, scared, suffering, afraid, smiling, loving, sad looking, weeping, crying, laughing, angry, , surprised, hateful, disgusted, excited, sorrowful, tears

3. ClipVision

If the face of the original image is at some unique angle, then using SAI's ClipVision feature as part of the POSITIVE ensures that the face angle remains unaltered despite the high strength denoise.

Also uploaded a comfyui workflow

It contains models for:

LEOSAM's HelloWorld XL 5.0

I've been using it for face restoration since Ver4. But up until Ver5, it still looked like it lacked a dataset of portraits of people at certain camera angles and face orientations. That was my pain points in using it, hence this release.

JuggernautXL_RunDiffusion_TensorV7

dpo-sdxl-text2image-v1

and 2 lora files that I trained myself on this basis

The training material uses the following models:

Expressions XL

merge method:

1. JuggernautXL_RunDiffusion_TensorV7 add dpo-sdxl-text2image-v1

2. train lora (Close-up pics of faces from various angles and ethnicities. LyCORIS-Full)

3. merge to LEOSAM's HelloWorld XL 5.0

4. train lora (Pics with various expressions from multiple angles and ethnicities. LyCORIS-Full)

5. Merge lora to checkpoint