NovelAI Diffusion V2:

Resolution: Native 1024px (unlike standard SD1.5's 512px)
Recommended aspect ratios: 1024x1024, 896x1152, 832x1216, 768x1344, 640x1536
CLIP skip: Set to 2
Best samplers:
euler_ancestral (most stable)
dpmpp_sde (best texture/stability balance)
gradient_estimation (faster convergence)
Scheduler: "simple" recommended
Tag order: "1boy, 1girl, characters, series, other general tags..."
Quality tags: best quality, amazing quality, great quality, etc.

>Specs:
High-resolution native support: Generates at 1024px natively (can push to 1024x1536px), unlike typical SD1.5 models limited to 512px
Clean U-Net training: Trained only on U-Net without text encoder contamination, making it ideal for fine-tuning and merging
Superior tag recognition: Rivals modern SDXL models despite being SD1.5-based
Improved VAE: Enhanced colors and eliminated the fading issues from v1
Exceptional detail rendering: Particularly strong at eyes and small accessories
>Versatile use cases:
Direct text-to-image generation
Merging with existing SD1.5 models
Enhancing other models via img2img

Extensive concept knowledge: Already understands many concepts without additional training

Despite being based on the older SD1.5 architecture, the model's training quality puts it ahead of many newer models in terms of output quality and flexibility.

>LINK:
https://civitai.com/models/1772131/novelaidiffusionv2