This is a TEST to see how training works on site with Wasabi's Eclipse XL.

Archived FOR NOW until I can test it better, trust me on this - I want to make sure I can fully test this - the Generator is not playing ball tonight!

https://civarchive.com/models/486131?modelVersionId=540597

Information from Wasabi & Co about prompting:

How to use:

Recommended starter prompt:

Positive prompt [prompt like any other tag based model]:

masterpiece, best, great, ...

Negative prompt (NO NEED for a long negative):

worst, worse, average, signature, watermark

recommended resolutions and prompting method:

(768, 1280), 3:5 ratio
(768, 1344), 4:7 ratio
(832, 1216), 13:19 ratio
(896, 1152), 7:9 ratio
(960, 1088), 15:17 ratio
(1024, 1024), square ratio

We recommend CFG between 5-8, sampling steps above 20 (we use 36 steps), and CLIP SKIP 1 (Pony Says to use 2, but Clip Skip is disabled for XL training on kohya so it doesn't make sense to use 2, people are probably gaslighting themselves from SD 1.5 experience).

(This is from : https://civarchive.com/models/486131?modelVersionId=540597 )

More information can be found about their model here:

https://civarchive.com/articles/5508/eclipsexl-in-depth-details-about-the-model

The images in question you see at first are INFERENCE from when I trained it on site.
I have not the time nor funds RIGHT NOW TODAY to test it offsite, so I am going to attempt to test it ON SITE. If you don't see anything of quality at first, this is not a knock on Eclipse XL: the generator is being a tad slow, I will be updating pics as it allows.

This is NOT LIKELY to work with base PonyXL nor SDXL, because it was trained ON Eclipse XL.

If this works, I'll replace this information with a better model card and sponsor details.

Because this is a LORA TEST + Loras being Finetunes off a base model, this is not meant to be uplifted elsewhere.