SplatoonXL

Generates colourful squids and octos.

Usage notes

• This model will help SDXL-based checkpoints draw:

  • in a style reminiscent of Splatoon concept and promotional 2D artwork.

  • concepts from Splatoon, including Inklings and Octolings.

• Recommended checkpoints:

  • v3.0: Use a Pony V6 checkpoint, such as Pony Diffusion V6 XL, or AutismMix.

  • v2.0/v1.0: Use an anime checkpoint, like kohakuXL, blue_pencilXL, CounterfeitXL, or Yamer's Anime/Unstable Diffusers.

• Guidance:

  • Prompting for keywords relating to place, ambience, or an activity can help with drawing scenery around characters.

  • Negative prompting is needed to reduce limb deformities and unwanted clutter.

  • Use your favourite prompting patterns or these suggestions.

    • v3.0: Prefix positive prompt with "score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up" following Pony guidelines.

    • v3.0/v2.0/v1.0: I found "photorealistic, 3d model, bad, worse, worst, ugly, bad anatomy, blurry, close-up, disembodied limb" in the negative prompt (taken from kohakuXL page) to consistently give cleaner images.

• Have fun!

Release notes

v3.0 (Pony V6)

• Retrained against the Pony Diffusion V6 XL checkpoint.

• Training dataset is identical to v2.0.

• Adjusted training parameters.

v2.0 (SDXL 1.0)

• Primary focus of changes were to improve generated Octoling accuracy, and general control over Inkling/Octoling features.

• Cleaned up tagging.

• Added more images to dataset.

• Tuned training parameters.

v1.0 (SDXL 1.0)

• Initial release.

Model information

This model was trained against official Nintendo Splatoon artwork. Art was selected by hand to match a particular visual style (highly saturated flat-shaded art).

See the "About this version" card for information about the dataset and training parameters. Images, tags, and training configuration are provided in an optional zip file along the model.

Attribution

This work was inspired by the following linked model by IWannaGoBack, which targets a similar niche for SD 1.5 based networks.

https://civitai.com/models/85425/splatoon-style

Thanks to IWannaGoBack. I enjoyed playing with your model, and it inspired me to dive into the deep rabbit hole of LoRA training.

Public domain notice and disclaimers

The author of this generative AI model (model) and curator of its training database (database) releases rights and disclaims liabilities that pertain to the model/database themselves or usage thereof, under the mark of the CC0 1.0 legal code.

http://creativecommons.org/publicdomain/zero/1.0

This is a fan work that derives from intellectual property owned by Nintendo Co., Ltd. It is not endorsed by Nintendo.

描述:

Training run 28, epoch 000014

Dataset 16

Images

• Count: 77 images

• Sources:

  • Danbooru (search terms: splatoon official_artwork)

  • Inkipedia

  • Splatoon Base

  • Splatoon 3 artbook scans

• Scanning methodology

  • Flatbed scans at 600 dpi

  • Descreening: https://github.com/6o6o/fft-descreen

  • Despeckling: Unity (1x) upscaling through R-ESRGAN 4x+ Anime6B

  • Dedusting: manual painting in an image editor.

• Treatment:

  • Most source images were either unmodified, or simply cropped to isolate individuals or mask logos/text.

  • Special effort was invested in obtaining additional images of individual Octolings.

    • Some images were obtained from Splatoon 3 artbook scans.

    • Some images were obtained by lassoing individuals from scenes, and repainting obscured areas in an image editor. Quality and resolution were enhanced by 4x upscaling through R-ESRGAN 4x+ Anime6B.

• Flip augmentation: enabled

Tagging

• Methodology: Tags were initially scraped alongside images, or tagged by WD14 (wd-v1-4-swinv2-tagger-v2). Tags were edited by model author for correctness and consistency.

• Caption shuffling: All tags are randomly shuffled. No trigger words were used.

Training

• Dataset replicates/epoch: 10

• Epochs: 14

• Batch size: 2

• Checkpoint: sd_xl_base_1.0.safetensors

• Network configuration: LyCORIS/LoCon (dim=16, alpha=8, conv_dim=8, conv_alpha=1)

• LR: 0.0001 for both U-Net and text encoder. Algorithm was "cosine_with_restarts" with 5% warmup.

• Optimizer: AdamW

训练词语: