Asap
When using DMD2 Speed LoRA, I use the following settings:
Clip Skip: 1 or 2
Steps: 9 (smaller values are possible)
LoRA Weight: 0.7 or 1.0
Sampler: LCM
Scheduler: Normal
Compression Generator: 1.2 or 1.4
Noise Multiplier for img2img: 1.1
I enable ADetailer but disable the ADetailer detector.
Three settings are available for normal generation.
1.
Skip Clip: 1 or 2
Sampler: DPM++ 2M SDE
Templates: Karras
Steps: 42
CFG: 5.5
Noise Multiplier for img2img: 1.1
I enable ADetailer
The ADetailer detector for "face_yolov8n.pt", but you can use any other detector.
2.
Skip Clip: 1 or 2
Sampler: DPM++ 2M SDE
Templates: Karras
Steps: 31
CFG: 4.0
Noise Multiplier for img2img: 1.1
I enable ADetailer
The ADetailer detector for "face_yolov8n.pt", but you can use any other detector.
3.
Skip Clip: 1 or 2
Sampler: DPM++ 2M SDE
Templates: Karras
Steps: 30
CFG: 6.0
Noise Multiplier for img2img 1.1
I enable ADetailer
Enable the ADetailer detector "face_yolov8n.pt", but any other detector will do.
I'm enhancing the image.
Recommended Settings
Highres.fix: Scaling 1.5
Scaling 4x_foolhardy_Remacri (1xDeNoise_realplksr_otf or OmniSR_X4_DIV2K,4xLSDIRplusC)
High Resolution Steps: 10
Enable ADetailer
Enable the ADetailer detector "face_yolov8n.pt", but any other detector will do.
Noise reduction level: 0.1 or 0.28; 0.4; 0.48; 0.50; 0.58 (I usually use 0.1).
But feel free to experiment, no one's stopping you from using other settings!
Description
An intermediate test model. I decided to publish it. I recommend using DMD2 Speed LoRA [SDXL, Pony, Illustrated]. For correct generation, be sure to set Clip skip=2, otherwise people's faces will break (although this is not certain). The model doesn't like brackets in tooltips, but not always. If people's faces break during normal image generation, try changing the sampling step and CFG scale.
FAQ
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.