LoRA training cost 500 BUZZ. If you want more LoRAs like that, and more LoRAs trained on CivitAI, please give me BUZZ (or give me your reactions and posts; every your reaction under my images give me 2 BUZZ, every your image-post under my models give me 50 BUZZ).

About idea and dataset creation:

From times of H.R.Giger and till nowadays most artworks in biomechanoid genre are pale, monochrome or one-colored. So I decide to make them more colorful. For that I:

1. Collected black and white sketches and artworks of my favorite artists in cyberpunk and biomechanoid genre (mostly Shichigoro Shingo),

2. Collaged this sketches together to make 60 new images in 768x768 resolution (most original images were smaller so sometimes I collage together up to three images),

3. Colorized the resulted images with https://palette.fm,

4. Manually fixed colorization of some parts and add a lot of bright toxic blue eyes,

5. Put this sketches on my own handmade colorful dirty grunge backgrounds.

This was my hardest ever work on dataset, but I believe that I created something new. It's not a captured style of Shichigoro or other artist. It's my own interpretation of biomechanoid.

Training setup:

LoRA for SD1.5 trained with CivitAI on 120 (60 originals + 60 mirrored) 768x768 resolution with manually enhanced BLIP2 captions.

Training values (all default except resolution, targetSteps, maxTrainEpochs, numRepeats): { "unetLR": 0.0005, "clipSkip": 1, "loraType": "lora", "keepTokens": 0, "networkDim": 32, "numRepeats": 10, "resolution": 768, "lrScheduler": "cosine_with_restarts", "minSnrGamma": 5, "targetSteps": 1600, "enableBucket": true, "networkAlpha": 16, "optimizerArgs": "weight_decay=0.1", "optimizerType": "AdamW8Bit", "textEncoderLR": 0.00005, "maxTrainEpochs": 8, "shuffleCaption": true, "trainBatchSize": 6, "flipAugmentation": false, "lrSchedulerNumCycles": 3 }

Usage:

Best LoRA values: from 0,5 to 0,9

Higher and lower values yet not enough tested.

Trigger word: AIDA_ColGruBioMec

Very often I use trigger word in constructions like: "artwork by AIDA_ColGruBioMec" OR "drawing in style of AIDA_ColGruBioMec" OR "close up portrait in style of AIDA_ColGruBioMec"

It's OK to use trigger word several times in one prompt.

Also LoRA should understand:

1. Several tips of backgrounds/environments:

   1.1. "no background" (for full picture covered with mechanical parts),

   1.2. "black background, simple background, isolated on black",

   1.3. "black background, grunge background",

   1.4. "white background, simple background, isolated on white",

   1.5. "sky background, dirty blue background, grunge background" (in this case you can replace "blue" with any color and sometimes remove "sky background").

2. Several tips of styles inside this styles (instead of "head" or "face" in all this phrases could be any other object, body part or character):

   2.1. "head turning into engine equipped with wires and gears" (here main keyword is "gears"),

   2.2. "head turning into engine entangled by tubes and equipped with air intakes covered by grilles" (here main keyword is "intakes"),

   2.3. "head is flying in the sky turning into alien spaceship, massive structure entangled by wires and tubes" (here main keywords are "flying", "alien spaceship" and "massive structure"),

   2.4. "face with spiky implants instead of hair, edges of face disappearing turning into wires, edges of face disappearing turning into wires and tubes, edges of face disappearing turning into ribbons and paper cuts, face disappearing turning into wires" (this long formula gives subject surrounded by implants),

   2.5. "an anime painting of a head with a hair blowing in the wind" (here main keywords are "anime" and "hair", this will give less mechanical and more grunge in style plus anime eyes and anime hair).

In some cases recommended in negative prompt: blurry, hair, anime.