LomoXL

Similar to LomoStyle for SD1.5; LomoXL is trained on a dataset consisting of images taken with analog tech.

Features

High-resolution images: Trained on 1024x1024 images (default for XL) manually captioned and cropped for optimal outputs.
Photorealism: While still lacking (compared to SD1.5), everything that went into this model was intended to create more authentic photorealistic images. This includes data selection, processing, captioning, and training parameters.
Full parameter fine-tune: This is a full parameter fine-tune of the original SDXL base 1.0 model. Both the U-Net and both text-encoders were tuned during training.

Limitations

LomoXL unfortunately, inherits some of the limitations from the original SDXL model. While improvements to photorealism were made in LomoXL, it still does not compare to what we get with SD1.5 models. Notably, skin texture and nudity. NSFW generations are possible but it's not ideal. To address this, I recommend using additional networks (such as LoRA's and Textual Inversions) as a temporary fix until SDXL base models have been trained enough to overcome this hurdle.

Usage

VAE - https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

+ Prompt: Natural language, with the addition of two special tags:

lomostyle, a prefix you can put at the start of your prompt.
in lomostyle, a suffix that can be added at the end of your prompt.

Note: These tokens are not required.

For more information regarding other special tokens (tags) that can be used with LomoXL, please refer to Lomostyle.

- Prompt: This one I'm not sure about yet. I mostly used the following negative prompt:

kid, child, 3D render, cartoon, watermark, text, signature, distorted, bad, exaggerated limbs, fake

Although, feel free to experiment with other negative tokens. Generally, less is more with diffusion models.

Questions/Feedback/Updates?

Visit my thread on the Unstable Diffusion Discord

Future updates are already in the works to bring SDXL up to — and beyond SD1.5 standards. If you want to contribute, DM me at the discord thread linked above.