@GloberTrotzer it's def. made with clip skip 1 when most pony checkpoints specify the use of clip skip 2 .. i mean it's not like your lora doesn't work w/ pony checkpoints that use 2, It produces results, but somewhat off because the LoRA learned to work with textual embeddings with the final CLIP layer (1), but most Pony checkpoints with Clip Skip 2 learned on the one prior. Mixing both leads to the LoRAs fine-tuned weights not aligning perfectly with the text representations which often results in potential color shifts, altered compositions, reduced prompt accuracy, and overall less predictable generation results. If it was built with the properly aligned clip skip then the same dataset and every other training parameters left as is, it would have performed way way better. Anyway, like I said it's not that it doesn't work at all when trying to create realistic results with it - it does - but I'm guessing that if made with aligned text-encoder layers it would have been absolutely fantastic.

@GloberTrotzer cool cool - and yes if text-encoder layers are aligned between LoRAs and checkpoints, all things are better.. I can't say for sure if it will produce more realistic results because that's really the quality and type of the dataset but I would say it def. would produce better results because of TE layer alignment. At the very least you won't have a setting hurting the quality of the outputs anymore. Based on what I did produce, it was fairly good which leads me to think maybe with that one setting change it would be really good.

The other thing you might try to really pump his LoRA up, heh, is upscaling training images to the minimum resolution spec guidelines for Pony/SDXL. It's easy enough to do with your fav. SD platform and upscaler model - UltraSharp/UltraMix_*, foolhardy remacri, Superscale 1780000, Swin4 etc, results will vary depending on the state of the source image but it's test fast/fail fast effort so you'll quickly find your favorites when doing it.

Of your datasets used for the Pony, only a few are actually within specs - only 32% are. This probably would be the biggest net-gain or net-loss when creating realistic images - probably more so than TE layer alignment. The resolution buckets used in the LoRA were made with buckets 1 thru 16 (512x512) are well below the recommended resolution. Buckets 17-24 start approaching or exceeding 1024x1024 in at least one dimension.

Despite the above, it's still decent - I'm just thinking with a few tweaks it would have been chefs kiss :D

Automatic and Forge don't even take into account the clip skip when generating on XL. Same for Easy LoRA scripts, it doesn't even let you pick it. So, you are fine creator.

I did not know he had 200hourabon terraria, I love him more now