Harhmix版本v4.0 (ID: 45999)

• Harhmix is a merge model which focus on anime style painting. it could also generate NSFW content.

• Features and improvements of Version 4 (My individual opinion):

  • Improvement of shade

  • Improvement of background

  • Better or worse hands

• How to use:

  • Clip: skip 2 or 1, In general, skip 1 is good at composition, but skip 2 is more creative.

  • VAE: Anything V3 VAE works very well, but other VAEs are triable.

  • Tags: Danbooru tag.

  • Negative prompts: Version 4 have no general negative prompts, now I use "(worst quality, low quality:1.4), bad eyebrow, fused eyebrow, bad teeth, poor drawn face" and change it depending on situations.

  • CFG: 7~11 is what I used in mostly time, lower CFG prone to Anything style, higher CFG prone to pastel-mix.

  • Sampler: DPM++ SDE Karras, Euler A perform better.

  • Highres fix: I perfer 20 steps and 0.45~0.65 Denoising strength.

    • Latent: Best performace but unstable.

    • R-ESRGAN 4x+ Anime6B: Stable but nomal performace

• List of merge source(Version 4): I only show the main merge source because of too many models and lora.

  • Anything-V3

  • Nutmegmix

  • AbyssOrangeMix2_hard

  • MIX-Pro-V3

  • Grapefruit

  • Novelai

  • Gape60

  • Sun and Shadow

  • Better light

  • Howls Moving Castle , Interior /Scenery LoRA ( Ghibli Style ) v3

描述:

• Harhmix is a merge model which focus on anime style painting. it could also generate NSFW content.

• Features and improvements of Version 4 (My individual opinion):

  • Improvement of shade

  • Improvement of background

  • Better or worse hands

• How to use:

  • Clip: skip 2 or 1, In general, skip 1 is good at composition, but skip 2 is more creative.

  • VAE: Anything V3 VAE works very well, but other VAEs are triable.

  • Tags: Danbooru tag.

  • Negative prompts: Version 4 have no general negative prompts, now I use "(worst quality, low quality:1.4), bad eyebrow, fused eyebrow, bad teeth, poor drawn face" and change it depending on situations.

  • CFG: 7~11 is what I used in mostly time, lower CFG prone to Anything style, higher CFG prone to pastel-mix.

  • Sampler: DPM++ SDE Karras, Euler A perform better.

  • Highres fix: I perfer 20 steps and 0.45~0.65 Denoising strength.

    • Latent: Best performace but unstable.

    • R-ESRGAN 4x+ Anime6B: Stable but nomal performace

• List of merge source: I only show the main merge source because of too many models and lora.

  • Anything-V3

  • Nutmegmix

  • AbyssOrangeMix2_hard

  • MIX-Pro-V3

  • Grapefruit

  • Novelai

  • Gape60

  • Sun and Shadow

  • Better light

  • Howls Moving Castle , Interior / Scenery LoRA ( Ghibli Style ) v3

训练词语: