⭐3wolfMondAI-SDG⭐

An improved version of realMondAI-SDG to provide better realism and bring back toony styles support.

Prompt recommendations
Generation recommendations
Sampler/Scheduler comparison
Style examples
Artist examples
RescaleCFG comparison
Example Generations

🟨Model Description

Developed by: Big Lasagna
Funded by: /sdg/
Also available on civitai.com
Also available on tensor.art

🟨Uses

This merge combines the best of StableMondAI-SDG and RealMondAI-SDG into one model. Its base style is a stable semi-realistic look while the realism tags turn on the realistic generations from RealMondAI, and using artist tags and style tags bring back the toony styles. Some can even be combined but some toony styles are trained so strong that they overwrite the realism completely. Play around with different weights to find a middle ground.

This model is in V-PREDICTION mode.

Prompts need to use danbooru/e621 tags, no Natural Language Support is added at this point. Use noobAI quality tags at your own discretion.

See below for examples and comparison grids.

🟨Prompt Recommendations

Start with a basic prompt and work from there. All IllustriousXL/noobAIXL quality tags are recognized but these models work better with leaner prompts.

Starter Positive prompt:

masterpiece, best quality, newest,

Starter Negative prompt:

worst quality,

To increase the realism, add these tags to the positive prompt

(photorealism, realistic,) detailed background,

and these to the negative prompt

blurry background, blurry foreground, simple background,

If you want increased fur detail, try adding some of these tags into the prompt:

(detailed fur:1.2), fluffy, shaggy fur,

The realism combined with some tags will try to pull your generations toward human features, if this happens you can add "human (feature)" to the neg prompt, e.g.

human hands, human lips,

🟨Generation Recommendations

The following settings are recommended starting points:

Sampler: Euler A
Scheduler: SGM Uniform
Steps: 30
CFG: 5

RescaleCFG of 0.7 is recommended.

Euler A & DPM++ 2s ancestral as well as normal Euler and DPM++2M seem to work best, SDE samplers tend to generate more artifacts on some schedulers. Karras is not recommneded.