As a 3D generalist and CGI artist, I often find myself in the need of complex and specific moodboards, and, if you ever did one, you most certainly know that this is almost impossible. Pinterest, ArtStation, Behance, and general Google Search can get you only as far as someone already imagined - so the inspiration is either limited or unconsistent.

This makes my take on AI art too: a huge oasis of unlimited moodboards customized and fine-tuned my your own need. But for this tool to really do wonders and to be a trusty sidekick even in a production environment, a lot of tweaking is required.

M.OS Realistic Diffusion aims to be the fine-tuned model that can be used by artists of various industries to kickstart their creations even if they work in automotive, photography, game, fashion, or architecture.

The Model

... is trained using CC0 libraries, stock images, and my own copyrighted work.

Recommended settings:

Clip Skip 2;
VAE: StabilityAI ft-MSE 850000 EMA (How to);
Dynamic Thresholding (more samples and details without clipping);
Hires. Fix (for weird aspect ratios);
Optimized for 512px and 768px;
Sampler: DPM++ SDE Karras;
Soft / cinematic looks: 2-6 CFG + 10-20 Steps (fast);
Realistic looks: 6-10 CFG + 20-40 Steps (balanced);
Dramatic / high-detail looks: >10 CFG + >40 Steps (slower);

Samplers (based on no. of samples required to produce good images, in order - CFG 7):

Fast (<10 steps): DPM++ SDE Karras, Euler, Euler a, DPM ++ SDE, DDIM;
Balanced (<20 steps): DPM++ 2M, DPM++ 2M Karras, Heun;
Slow (<30 steps): DPM++ 2M SDE Karras, PLMS, LMS.

Take a moment to leave a rating and fav. M.OS Realistic Diffusion. It will give me a greater chance to reach a broader audience and fine-tune the model further.