UmaMusume All In One LoRA版本SDXL-alpha (ID: 212365)

This is AIO LoRA of Uma Musumes.

All of our works is free. If you'd like to support our team, feel free to buy me a coffee. ?

One of our team member have admission into a graduate school. So, we don't have much times. Here is some plan for this

TL;DR

• SDXL-pony-uaf

  • Based on PonyXL-V6, for its variant models

  • Dataset: ULTIMA-uaf

• SDXL-animagine

  • Based on Animagine, for its variant models

  • Dataset: ULTIMA-uaf

• SD1.5-uaf

  • Based on SD1.5 animefull, for sd1.5 variant models. e.g. Counterfeit-V3.0

  • Dataset: ULTIMA-uaf

• Deprecated version

  • SDXL-alpha

    • Based on SDXL1.0, for booru-based models. e.g. AnimagineXL,

    • Dataset: ULTIMA

  • SDXL-pony

    • Based on PonyXL-V6, for pony variant models

    • Dataset: ULTIMA

  • SD1.5

    • Based on SD1.5 animefull, for sd1.5 variant models. e.g. Counterfeit-V3.0

    • Dataset: ULTIMA

PLAN

1. We will refine ULTIMA Dataset for characters who have insufficient images. Done.

   1. Refining done!

2. When the main scenario ends, we will supply the costumes or characters released during the period.

   1. 4 months are too short. We focus on refining dataset now. Done!

3. If there is a new Stable Diffusion model like SDXL, we will make LoRA for this depending on budget.

   1. I wait SD-3 will be released.

4. As we said earlier, the above plan can be changed at any time because we don't have much time. We hope this is sustainable project.

NOTICE

• Some SDXL model, like Kohaku-XL Epsilon, can generate various Uma Musumes!

• For now, SDXL version is unstable and alpha version. I made this for a test.

• This is AIO LoRA of almost characters of Uma Musume : Pretty Derby.

  • 131 characters in Uma Musume: Pretty Derby

  • About 300+ costumes. For each character, average 2+ costumes.

  • 8 official clothes for all characters.

• You can find all possible trigger words in here.

• Depending on model, Some characters are not able to generate.

  • For example, Ikuno Dictus - Counterfeit V3.0

• The LoRA is optimized for Hires.fix.

Settings

• Please see ABOUT THIS VERSION in right side of page ----->

other training settings is included in metadata of LoRA file

License

• This model is released under "Derivative work guidelines for umamusume".

• And "Fair-AI public license 1.0-SD" License

描述:

This is alpha version of SDXL AIO LoRA of Uma musume

Recommended options

• LoRA weight 1.0

• Trigger words

  • [character name] \(umamusume\) or [character name]

  • For example,

    • manhattan cafe \(umamusume\)

    • hayakawa tazuna

SDXL version

• trained on sd-scripts by kohya_ss and LyCORIS by KohakuBlueleaf

  • Thank you all a lot!

• Base model : SDXL 1.0

• Hardware : 4x RTX TITAN

• Training time : 37 hours

• Dataset : ULTIMA Dataset

• Resolution : 768x1024 with aspect ratio bucketing

• Dim, Alpha : 32

  • with dynamic pruning rank=24

  • Network dropout = 0.1

• Optimizer : AdamW, weight decay=0.1

• Steps : 12,024

• Total Batch size : 48

  • 4 GPU x 3 batch size x 4 Gradient Accumulation steps

• Learning rate : warmup to 1e-4 for 340 steps and then kept constantLR scheduler

  • LR scheduler is constant_with_warmup.

Settings

• For Stable Diffusion XL

  • Use a model derived from Stable Diffusion XL 0.9 or 1.0

  • DPM++ 2M SDE Karras

  • about 30 steps, CFG scale ~6.5

  • SDXL vae

• CLIP skip = 1 or 2. Use whatever you want.

• Use Hires.fix to get higher quality image

  • Upscaler

    • Latent, Latent (nearest-exact), Latent (bicubic antialiased) or other Latent series.

  • Denoising strength 0.50~0.65

If you want to use SDXL-alpha LoRA with AnimagineXL, use weight around 0.6

训练词语: PLEASE SEE DOCUMENTS