Test LoRAs for Waifu Diffusion v1.3

These LoRAs are experimental LoRAs for WD1.3 to produce high resolution or different aspect ratio images.

Model Description

They have fine-tuned from the original WD1.3 or a model merged with LoRA in this repository by thousands of unselected AI illustrations by various authors and models published on the Internet. Each networks has been fine-tuned with a learning rate of 6.0e-5 for 5 epochs on about 5-8k images at batch size 8, using Aspect Ratio Bucketing with a maximum resolution of 768x768. Fine tuning performed by RTX3090 at fp16 with AdamW8bit optimizer and took 2-3 hours for each network.

LoRA Name	Base model	images	note
hires_test_a	WD1.3	~5k
hires_test_b	WD1.3	~7k
hires_test_c	WD1.3 + 1.0 * hires_test_a	~8k	recommended for use with hires_test_a
hires_test_d	WD1.3	~5k	dim 128, U-net only
smooth_test_a	WD1.3 + 2.0 * hires_test_a	~7k
smooth_test_b	WD1.3 + 2.0 * hires_test_a	~7k	different seed

There is probably no overlap between the three image sets (5k, 7k, 8k).

Usage

The LoRA are mainly classified into two types: for high-resolution and for smoothing. First, please apply high resolution LoRA at the preferred ratio: 1-2 is recommended for ~768x768, and the higher the resolution, the more weight is recommended. In some cases, especially when weights are large, adverse effects may be observed. In such cases, please consider applying a leveling LoRA.

License

This model is open access and available to all, with a CreativeML OpenRAIL-M license further specifying rights and usage. The CreativeML OpenRAIL License specifies:

You can't use the model to deliberately produce nor share illegal or harmful outputs or content
The authors claims no rights on the outputs you generate, you are free to use them and are accountable for their use which must not go against the provisions set in the license
You may re-distribute the weights and use the model commercially and/or as a service. If you do, please be aware you have to include the same use restrictions as the ones in the license and share a copy of the CreativeML OpenRAIL-M to all your users (please read the license entirely and carefully) Please read the full license here

Acknowledgements

These LoRAs build on the two excellent works: SD1.4, developed by CompVis Researchers, and WD1.3, developed by Anthony Mercurio, Salt, and Cafe.