, close-up, u — копия.jpeg)
, __mature female, dark-skinned — копия.jpeg)
Hello again my friends!
Intro
Since the LAX Lab team released the second Pony killer I also started making loras for the NoobAI-XL V-pred model. I was skeptical but now I think it's a very good model.
This model listens better to your promt than Epsilon-type models (NoobAI Epsilon and Pony) and makes dark scenes more pleasant. However, using this model requires using additional parameters.
Since V-prediction model need extra settings I made little guide for you.
This guide for local webui but I hope civitai will add these options to site image generator because NoobAI-XL V-pred is game changer.
! If V-pred settings are difficult for you or you use site image generator which doesn't support V-prediction and rescaling CFG it is better to use NoobAI-Xl Epsilon 1.1 and Cyberfix v3 (or other mixes if NoobAI / Cyberfix are not available).
I make my loras for both versions. NoobAI-XL Epsilon is also very good model!
---------------------------------------------------------------
Optimal settings for NoobAI-XL V-pred:
I. WEBUI
You need web user interface which supports V-prediction sampling and zero SNR.
I recommend reForge or Forge webui. Also comfyUI works.
A1111 also supports V-pred but in dev branch. Since it slower than forge / reForge I stopped to use A1111.
You can find detailed information about supported webui on noobai-xl v-pred page in description.
My personal choice - reForge. Forge is also good (but it has broken checkpoint hash calculating which prevent civitai to see checkpoints which you used in art).
---------------------------------------------------------------
II. Models
I can recommend 2 main models for generating on NoobAI-XL V-pred:
I prefer Cyberfix because according to description in this merge SDXL 1.0 was replaced by Cyberrealism XL model which has better anatomy than SDXL 1.0 base by StabilityAI. But anime / furry style of NoobAI-XL is saved.
The only side effect of this merge - scenes are brighter than in NoobAI-Xl V-pred 1.0.
I confirm that the anatomy on the cyberfix is much better than on the basic noobai-xl v-pred model.
By the way, this statement also applies to the cyberfix for Epsilon - it also improves the anatomy of the base NoobAI-XL Epsilon model.
But when you need darker scenes it is good idea to use NoobAI-XL V-pred 1.0.
---------------------------------------------------------------
III Settings for generating images
Txt2img settings:
Euler, shedule type: automatic or SGM Uniform, 30 steps.
Euler a with SGM Uniform is also very good! Try it!
CFG 4-6 (I use 5).
Hires fix: Upscaler Foolhardy_Remacri_4x, Hires steps 12-18, denoising strength 0.25 - 0.3.
Adetailer: 0.3 denoising strength, separate resolution 1024x1024.
---------------------------------------------------------------
Necessary settings for V-pred:
Since V-pred model sometimes is oversaturated I recommend to use one of the following extensions (already installed in reForge / Forge).
YOU NEED ONLY ONE! DON'T USE THEM TOGETHER:
Rescale CFG 0.3 - 0.7 (reForge webui);
Latent Modifier - rescale CFG Phi 0.3 - 0.7 (Forge or reForge webui);
Dynamic Thresholding - mimic cfg 4-5, cfg - 5-6 (Forge or reForge);
APG is now your CFG (reForge, experimental. Better use extensions above)
---------------------------------------------------------------
IV Quality prompt (very important):
Put "masterpiece, good quality, best quality" in positive and put "worst aesthetic, worst quality, low quality, bad quality" in negative.
"very awa, very as2" also can help but start without it.
You don't need to use artist tags if you use my character loras or style loras but if you want to use it check danbooru for artist tags!
Also for everything (art styles, poses, body, faces, actions, backgrounds, etc) use danbooru wiki (for anime and everything) and e621 (for furry and monster tags).
---------------------------------------------------------------
If you have any questions left, I am ready to answer them.