#2 🚧 Everything-Mix Experimental versions (for Hen and Mix V3)

Enviado em: 6 de dez. de 2023, 08:42
Usos: 9.1k
Avaliações: Excelente (1)
Passos de Amostragem: 27
Método de Amostragem: Euler a
CFG scale: 7
Negativo: (bad-artist-anime:0.9), (by bad-artist:0.8), (worst quality, low quality:2), (bad quality:1.4), normal quality, (EasyNegativeV2, easynegative:1), (bad eye, unclear eyes, undetailed eyes, missing limbs, fused limbs, broken limbs, fewer digits, extra digits, mutation, deformity:1.3), (interlocked fingers:1.3), (bad anatomy, white background:1.1), monochrome, watermark, signature, ng_deepnegative_v1_75t, bad-hands-5, verybadimagenegative_v1.3, green skin
Permissões: Permitir geração e compartilhamento de imagens
Permitir que usuários baixem seu modelo
Usos comerciais

Descrição

My idea for Everything-Mix v3 basically will be to try to mix the concept from v1 and v2 with a mix between 2D and 2.5D style and maybe make a better 2D style, but it is very hard. Models in progress: · v3_experimental is a version that tries to mix 2D and 3D. This is intended to be a mix between v1 and v2, but not as exactly as that. · v3_realis is another version evolved from v2_human and v2_realis that generates 2.75D images. It is still an experimental version. · v3_blurness is a very different version that gives blur effects to the images and it is also good in closeups (best eyes yet in my models, but not on full body). You can lower steps in closeups but the quality will be worse. Sometimes it doesn't add blurness but it is part of the lightning. · v3_Reverse is the version that tries to be the best in every aspect: face, N-SFW, hands... It is very close to the final v3 version I want to make, but it suffers from some issues. I named it that way because it has the opposite blocks of my v3 final version. · v3_2D is the pre-released version (that evolved into Everything-Hen v3). Very good with prompts like black iris, detailed eyes... It's better with 1024x resolution. It is the model that I mixed up with a lot of my N-SFW models until I got something interesting out of it. It has theorically: Improvement in hand and face generation, improvement in genital/N-SFW, flexible with more samplers than my HEN versions. But I noticed while testing it that it needs just a few improvements. Recommended resolution: 1024x612. Samplers not tested enough.

I learned how to make better certain things with a guide on blocks weights and I can improve my models even more from now on, but it will require much more experimentation. Everything-Mix v3 will be hardly experimental for me in that sense so it doesn't necessarily mean that it will be better than v2, but I'll try it.

Issues with v3_EX: Eyes, full body. Hands are like v2, sometimes worse and sometimes good. Issues with the rest of versions. I only recommend here the model v3_2D.

Recommended use of LoRAs. Sometimes it does very good work with 20 steps, but I'll recommend 30 to be safe.