1.5 model training, it is recommended to use a two-dimensional base model, refer to the example picture, you can try it yourself,
Sampling Method recommends using DPM++2MKarras, Eulera, or Eulera.
Sampling Steps recommend 20-30, there is no need for too many step values to waste time running the image;
CFG Scale recommends 6 to 9;1.5 model training, it is recommended to use a two-dimensional base model, refer to the example picture, you can try it yourself,
Sampling Method recommends using DPM++2MKarras, Eulera, or Eulera.
Sampling Steps recommend 20-30, there is no need for too many step values to waste time running the image;
CFG Scale recommends 6 to 9;
Generated image sizes: 768x512, 512x768, 512x512, or 512xN, Nx512 (N represents any number). That is, keeping one value as 512 for width or height will yield better results. Because the original material of the Stabledifussion 1.5 base model training is 512x512, keeping one value as 512 will yield better results!
Hiresfix (high-definition restoration), after extracting a satisfactory effect, then open Hiresfix (high-definition restoration). This can save a lot of time.
Upscaler recommends using 4x-AnimeSharp, R-ESRGAN4x+, R-ESRGAN4x+Anime6B;
Sampling Steps 25-35 are sufficient; DenoisingStrength recommends 0.3-0.4 (too high a value will deviate from the original image)
Lora weight: 0.7~1
No trigger words, natural language description,
Generated image sizes: 768x512, 512x768, 512x512, or 512xN, Nx512 (N represents any number). That is, keeping one value as 512 for width or height will yield better results. Because the original material of the Stabledifussion 1.5 base model training is 512x512, keeping one value as 512 will yield better results!
Hiresfix (high-definition restoration), after extracting a satisfactory effect, then open Hiresfix (high-definition restoration). This can save a lot of time.
Upscaler recommends using 4x-AnimeSharp, R-ESRGAN4x+, R-ESRGAN4x+Anime6B;
Sampling Steps 25-35 are sufficient; DenoisingStrength recommends 0.3-0.4 (too high a value will deviate from the original image)
Lora weight: 0.7~1
No trigger words, natural language description,