Hello @dr_dark_flames ,

It is possible that the LoRAs favor one character, thus if they are set too strongly, you will struggle to generate multiple characters. In addition, generating wider images will typically better generate multiple characters, versus generating narrower, tall images. Another approach might be to generate without the LoRAs in txt2img to get multiple characters, then apply the LoRAs via img2img, if it appears the LoRAs are causing it not to generate multiple characters. Also, make sure to use proper group tags, as outlined at https://danbooru.donmai.us/wiki_pages/tag_group%3Agroups.

Hello @xyf ,

The pruned versions of this model are fp16.

Hey @746977787AF ,

These are pruned and fp16, with the VAE replaced with a non-distorted one (merging distorts the VAE in the model, so normally you'd have to use an external one, or replace the one in the model). Pruning a model and making it fp16 does not remove content or reduce quality, only components that aren't needed are removed, and weights trimmed to fp16. All popular UIs load and process at fp16 by default, you would have to manually set it to load and process at fp32 (which will increase vram utilization significantly and slow processing somewhat), and you wouldn't see improvements in quality, but might notice small differences in image content.

As for the style, it can vary depending on the prompt. Typically the style of the examples you see here is what people seem to prefer over what the examples on HF are, but the model is still capable of both.

Hello @kawaiinekogomeowz ,

I use Automatic1111's UI. InvokeAI as some nice polish and features in certain areas, but Automatic1111 is typically the most feature rich and tends to keep up with new developments in the space, either directly or via community extensions.