Sign In

Shaded: Retro Anime (SDXL)

31
188
463
12
Updated: Jun 19, 2024
styleanimevintagesexyretroart
Verified:
SafeTensor
Type
LoRA
Stats
188
463
Reviews
Published
Jun 9, 2024
Base Model
SDXL 1.0
Usage Tips
Clip Skip: 1
Strength: 0.8
Hash
AutoV2
4451188369

Shaded

This is a experimental LoRA

Key Points

  • Shaded was not trained on Booru tags

  • Shaded was trained on Natural Language captions, via a complex image captioning system I've been working on.

  • Despite not being trained on Booru tags, both natural language prompts and booru tags can be used — separately or in a hybrid approach.

    • For Booru tags to work, you will need a base model that was trained on Booru tags.

  • Subtitles and Text can be minimized by adding subtitle / subtitles or text to your negative prompt.

    • Some resolutions also tend to result in more or less subtitle outputs.

  • Most of the sample images were generated using sample captions from the dataset.

    • This was done to save time on inference.

    • For some prompts, I appended a few booru tags to test hybrid prompting.

  • Sample images are limited due to my PC running hot and overheating during inference.

    • This wasn't from the model. The GPU is reaching it's final days. (RIP)

  • Given the experimental nature of this LoRA. Feedback and Discoveries are welcome. Please provide your feedback and findings on the Discord thread linked bellow or, hear on Civitai.

    • Note: I get a lot of notifications on Civitai, so it may take a while for me to see your comment.


Usage

  • A LoRA weight of 0.8 is recommended. Though, this will vary depending on your base model.

    • If the style is too strong, you can lower the weight or use one of the earlier epochs (available on the huggingface repo linked bellow).

All Epochs can be found on 🤗Huggingface.


Limitations

  • Firstly, this is not the best Retro-Style anime LoRA for SDXL. At least, I don't believe it to be. Due to the focus on the captions, I admittedly ignored some of the data cleansing and post-processing steps that I commonly use when training a new model/LoRA. This is the reason for the follow limitations.

  • Unwanted Subtitles, Text, and Watermarks. (See Key Points on how to minimize this).

  • Booru tagging is dependent on your base model.

  • Some tokens seem to have a particularly large effect on the style.


Questions/Feedback/Updates

Visit my thread on the Unstable Diffusion Discord

Buy me a coffee ❤

https://ko-fi.com/ndimensional

All donations will be used to fund the creation of new Stable Diffusion fine-tunes and open-source AI tools.


Checkout my other models

SDXL

LoRA SDXL

SD1.5

LoRA SD1.5

https://tensor.art/u/738915824668190781

Discussion

Kotoshko's Avatar
Kotoshko
Celtic Creation Contest Participant

Thank the gods it's not a “Pony”, we'll try your model! 😉