V4 IS AVAILABLE!
What's new:
Replaced about 50% of dataset for better quality images;
More effects, more characters, more stable 3D style;
Added activation tag.
BEST MODELS TO USE 3D LORA:
NoobAICyberfix1.1 perpendicular
NoobAI-XL Epsilon (v. 0.75, v 1.0 or 1.1)
LoRA was trained on NAI Epsilon 0.75 because loras from 1.0 are oversaturated and loras from 1.1 have worse style but you can easily use this lora on 1.1 checkpoint (including cyberfix 1.1).
FOR NOOBAI-XL V-PRED (CYBERFIX) USE THIS LORA. If you can use rescale CFG option (local webui forge, comfyUI, Comfy on Tensorart) better use my V-pred version.
Please don't use it on anime mixes because they kill 3D style!
If you
NoobAI-XL / IL tags:
Main tag:
FOR VERSION 4 and Illustrious 0.1:
3d, realistic (For Resident Evil, The Witcher, Mass effect, etc. r34 3D).
OR
3d, cartoonish (For Overwatch, Fortnite, WoW r34 3D).
FOR VERSION 3: 3d \(realistic\)
FOR VERSIONS 1 - 2: 3d
Body effects and tags:
shiny skin, wet, oiled, toned, curvy, plump, muscular, tan, tanlines, cellulite
sidelighting, backlighting, underlighting, bloom, shadow, light paricles, sunlight, dappled sunlight, light rays, moonlight
FOR V1, V1.1:
To prompt known franchise characters use prompt order: {character name}, {francise name}, other tags.
FOR NEWER VERSIONS:
DON'T NEED FRANCHISE TAG. BETTER DON'T USE IT.
All characters in my LoRA are tagged with danbooru tags! So use tags from this site:
My preview (from all versions) can help.
NoobAI-XL 3D quality settings:
Put "masterpiece, good quality, best quality" in positive and put "worst aesthetic, worst quality, low quality, bad quality, lowres, extra digits" in negative.
Recommended generation settings:
Euler A SGM Uniform. Personally for NAI Epsilon Euler is better than Euler A while on NAI V-pred euler A is better.
30 steps, CFG 5-6.
Highres fix:
0.25-0.3 denoising strength, 12-15 hires steps,
Upscalers: Foolhardy Remacri 4x or (VERY powerful but slow. His main advantage - saving many small details. It is not so critical for anime but very critical for realistic 3D).
Adetailer:
Denoising strength 0.25 - 0.3. Detection model confident threshhold: 0.7-0.8.
Inpaint only masked padding pixels - 70-90 pix for face models, 156-256 for hands models.
Please don't forget to rate my LoRA. This motivates me to do more!