Check out fast inference at frosting.ai who helped sponsor this model!

Recommended Settings

We recommend using the following settings with our model for best results.

Resolution: 1024x1024
Steps: 8-12
CFG Guidance: 2.5
Sampler: Euler or Euler a

The model utilises the low guidance and steps due to it being an SDXL Lightning model. It's advised to keep your step count as well as your guidance low, whilst you may increase the step count if you choose to do so, there is a clear point of diminishing returns, therefore we recommend you to follow our recommended values here.

Model Info

Bananastrike XL is a successor to CompassMix XL Lightning. Its primary goal was to improve upon several issues related to Compass.

In our testing, it has shown to have particular improvements in.

Overall reduction of noise issue compared to CompassMix.
Improvements to background fidelity, less 'compressed' looking images.
Improved saturation balance, reducing overblown images.
Greatly improves issue where it looks like painterly characters are superimposed onto photographs.

The model still maintains much of the hallmarks of CompassMix XL Lighting, such as a high-level of natural language understanding, prompt coherence and a wide-range of art styles.

The model has also retained much of the knowledge from base SDXL, leaving an enormous array of concepts that the model understands.

Bananastrike is a very close relative of CompassMix, the same seed with the same prompt tends to provide very similar outputs, albeit it with a different style.

Strengths & Weaknesses

During our evaluations of the model by ourself and the community, we find the following strengths and weaknesses in the model.

Strengths

High-level natural language understanding with the ability to handle multiple concepts.
A good level of tag understanding, allowing for a mixture of prompt styling.
A wide range of art styles and following of art styles.
Great prompt coherence
Ability to do humans and anthros well, with a wide range of supported species.

Weaknesses

Noisy looking images, especially with photo backgrounds and plain backgrounds
Scenes with duo and higher tend to have various anatomical problems.
The model tends to prefer shorter captions compared to longer more detailed captions.

We aim to improve upon these issues in future releases of our models.

Prompting & Tips

Natural language usage is recommended as the model and SDXL in particular understands it. However, you are also welcome to use tagging as you please.

This model is based off of SeaArt, it's advised to use their quality tags. They provide an example of this from their HuggingFace page.

The model supports various levels of quality hints, from "masterpiece" to "worst quality". Be aware that "masterpiece" and "best quality" may lean towards nsfw content.

They provide other prompting tips on their HuggingFace page that may be applicable to the model, you can find that page below.

https://huggingface.co/SeaArtLab/SeaArt-Furry-XL-1.0

Credits

This model was developed by the Horizon Team, a team dedicated to creating high-quality furry AI models.

Lodestone Rock
Bananapuncakes
Thessalo

We also thank SeaArtLab for training the original SDXL furry model that was the basis of this merge.

Extra Funded Provided By:

Minotoro @ frosting.ai
Mobbun
Spectre

Finally, we'd like to express our heartfelt gratitude to our wonderful Patrons who made this model possible:

You are awesome! Patron:

Nexus
Nothingburger1234
TheGreatSako

Serious Patron:

SunBee
Simon
AndriFox
Mobbun
Kadah
Zenko

Patron:

Joseph
Krazark
kcanis

We are currently experiencing issues with our Patreon, if you'd like to discuss ways to donate, you can contact us at our Discord server here.

https://discord.gg/3GZHQTEEJq

You can give a direct PayPal donation via our PayPal link here.

https://www.paypal.com/paypalme/lodestonerock

You can reach us at our Discord server!

https://discord.gg/3GZHQTEEJq

We've also got a link at SubscribeStar.

https://subscribestar.com/lodestone-rock

Enjoy the release everyone! Take care.