LobotomyMix - The semi-realistic mastermind

Overview

LobotomyMix aims to generate a very specific beautiful semi-realistic art style. This model specializes in creating stunning characters, particularly women, with an emphasis on extreme detail and sharpness. It retains the diverse anatomical knowledge and character variety from its Danbooru-inspired foundation.

The model merges layers from multiple Illustrious models and LoRAs. A custom LoRA, which was specifically trained for this model’s target art style, has been integrated to refine the output.

Strength

• Semi-realistic, highly beautified aesthetics

• Diverse fictional characters and anatomical positions

• Extreme detail aswell as sharp eyes and mostly accurate hands even without upscaling

• Stability when generating at higher base resolutions

Weakness

• Struggles with non-specialized styles

• Tends to add somewhat excessive detail ("noise" or artifacts), especially when upscaling

• Sometimes inconsistent results with male characters, especially when upscaling

• Kinda repetitive facial features across outputs based on the focussed art style

Prompting

Try to avoid artist/style tags, since this model is optimized for its built-in aesthetic. Characters, complex ???????, and anime-inspired traits (e.g., colored hair) work well despite the semi-realistic style.

Positive Prompt Template

Use structured tags or even try natural language. This is simply the prompt schema I used while testing the model and also like to use in general.

[primary subject: f.e. 1girl], [specific character, if wanted: f.e. raiden_shogun], [additional things: f.e. beach], masterpiece, best quality, detailed, beautiful, 8k, highres, absurdres, incredibly_absurdres,

Negative Prompt

This is the negative prompt I use pretty much everywhere. Just add more specific things, if you want to exclude anything from your generation. Also feel free to use your own negative prompt if you like.

signature, text, monochrome, rough sketch, fewer digits, extra digits, worst quality, low quality, sketch, interlocked fingers, comic, bad ???????, lowres, bad hands, greyscale, watermark, uneven eyes, lazy eye, framed, outside_border,

Generation Parameters

CFG Scale

• 2-6 (sweet spot: ~4 for realism/coherence)

Sampler

• DPM++ 3M SDE Dynamic ETA (or other DPM++ SDE samplers)

• Any sampler should work just fine.

Scheduler

• GITS scheduler

• Any scheduler should work just fine.

Steps

• ~20 steps (sweet spot for high quality and still manageable speed)

Resolution

• This model works with very diverse resolutions and works especially great for higher base resolutions, as long as one side roughly is around 1024 in length.

• Works great

  • 1024x1024, 1280x1024, 1024x1280, 1536x1024, 1024x1536

• Works often

  • 1792x1024, 1024x1792

• Works sometimes

  • 1920x1080, 1080x1920, 2048x1024, 1024x2048

• Different resolutions in between those listed also should work.

• Extremely small resolutions might sometimes require an additional detailer for the face and hands.

v1.0 Perun - [04.02.2025]

• Initial release

• Core semi-realistic style

• Retains anime-inspired anatomical and character knowledge

• Enhanced detail and sharpness

• Sharp eyes and anatomically correct hands

• Stability for high resolution generation