See v1.0 for more examples.

仅作学习目的。

DO NOT POST YOUR ???? WORKS HERE.

Notes on usage:

1. LoRA weights around 1.0 should work fine when this model is the only one applied on.

2. Try different combinations, such as (portrait | full body) + (from side | from behind | from above)

3. I'm using the DPM++2M Karras sampler for 30 steps to generate my showcases. No hires-fix is required to generate these photos.

What is the purpose of this model?

This model is a test case on the two parameters network_dim and network_alpha, my previous models are all trained with network_dim=network_alpha=32, and requires lot of fine-tuning, but this one is trained with network_dim=network_alpha=128, and nothing else changed.

It's much, much more robust when generating full body photos from different angles, and more sensitive to the prompts!

Conclusions:

Though the file size is much larger, it does perfom a lot better.

TODO:

Go back to train my previous models with these params to see how it performs.