Myne Factory Base Model

The foundation of our models

Technical Details

Model Training

MyneFactoryBase was trained using ~18000 high scored samples from Yande.re and ~5000 high scored samples from Konachan. File captions were generated using 3 iterations of WD1.4 tagger to ensure maximum identification of objects within the training data. A second captioning run was done using one tagger with a reduced threshold to produce shorter captions for later use. Adam optimizer was used with a manually set maximum learning rate and cosine decay. Training was done on an RTX 4090 with a batch size of 4, utilizing DDIM sample scheduler and DDPM noise scheduler with mix precision.

Text Encoder Training

Text Encoder was trained for 50% of the training durations, freezing and unfreezing every 10ep. During the final 20ep of finetuning, the TE was frozen.

Block Merge

At the ep20 milestone, a block merge was done with BasilMix. However, it was evident that the merged weights were being trained out quickly, and the weights had entirely shifted back to the training data by the end of the training. Ultimately, the decision was made to not use a block merge for the final release.

For more detailed technical information on the training process and model architecture, please refer to this document.

Authors: 金Goldkoron, tsmkirby, Juusoz

Visit our Discord community if you have any questions.

Prompt Format

It is recommended to use booru styled tags to for the prompts.

Example: woman, decorated horns, long robes, fog, long curly hair, freckles, solo, masterpiece, reflective, depth of field, caustics, detailed night, forest, leaves, moonlight, eyes, orange hair, green eyes, vines

Example: 1girl, solo, skirt, book, glasses, long hair, looking at viewer, bookshelf, jacket, plaid skirt, school uniform, long sleeves, parted lips, semi-rimless eyewear, bangs, blush, holding, blazer, indoors, sweater, under-rim eyewear, red-framed eyewear, holding book, brown eyes, library, sitting

The tags were generated with WD14 tagger for the dataset.

The model has also been fine tuned to be better at handling shorter prompts.

Recommended Settings

This model performs best with the following settings:

• Image Size

  1024x576 for wide 16:9, 768x768 for square, and ??0x1024 for portrait

  Feel free to experiment with higher resolutions, Juusoz made all the examples at higher than recommended resolutions

• Vae

  vae-ft-mse-840000-ema-pruned.ckpt

• Sampler

  DPM++ SDE Karras (preferred)

  2S Karras

  Karras samplers tend to create more dynamic and interesting generations

  Euler A

  Results tends to look smoother and more Airbrushed

• Steps

  30 minimum and +70 can give nice results

• Skip Clip:

  Clip 1

  Clip 2 and 4 are valid for experimentation and we recommend trying it for more variation.

• CFG

  9-12

• Not required, but these tags improve the quality of the image:

  Prompt: best quality, masterpiece

  Negative Prompt: lowres, bad ???????, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry

Socials

Website | Discord | Patreon