Version 2.

Removed 10 training images and captioned the remaining 30 (simple short captions by hand, not BLIPped or whatever). Submitted to the CivitAI trainer again. I think this version is better in general. The biggest issue that’s annoying me is that sometimes the LoRA doesn’t “lock on” to a prompt. So sometimes you get a photo instead of a quilting/embroidery result... but with some embroidered elements such as a character’s clothes, or furniture, or things in background.

All in all I am happier with version 2. The LoRAs are small enough that you can keep both.

Originally I created the Embroidery SDXL TI:

https://civarchive.com/models/554798/embroidered-quilting-ti-xl

and now this LoRA for generating ideas for Real Life quilting designs.

At the moment I think the SDXL TI does a better job, and is much faster of course. Also cheaper in CivitAI trainer! But this LoRA does benefit from certain things that Flux does better, e.g. hands. Flux isn’t perfect (especially with its AI/3D/rendered look), but it does generally give good prompt adherence.

Another thing I’ve noticed about Flux is that human proportions are sometimes (often?) non-real. It seems to use artistic “heroic”, “monumental” and “naive” proportions on occasion. Hands and eyes/faces sometimes seem too big. All this is something Real Life artists do for effect of course, but I am not comfortable with an AI/ML model doing it by default or at random.

Enough rambling! I hope you find this LoRA interesting/useful. Please try the SDXL TI as well if you haven’t already.

Comment below or Chat me if you have something to say about this LoRA. Image gens posted on this page are always welcome :-)

= = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = =

Thanks to CivitAI staff for featuring this model on the homepage :-)

Thank you to downloaders and reactors for version 1. This LoRA (v1) does work, although often quite weakly depending on which software you are using (Comfy, Forge, etc). It will probably work well in CivitAI's generator since it was trained in the CivitAI generator. But in Forge I'm not happy.

This version was made with 40 uncaptioned images. From what CivitAI says in its guidance, uncaptioned ought to work well with Flux training. However, the same guidance suggests perhaps 20 images is enough. So perhaps I've used too many.

I expected this LoRA to be as "strong" as my decorative leather one:

https://civarchive.com/models/660817/decorative-leather-and-mixed-media-flux-lora

but it's not. I did (hand) caption the Decorative Leather training images, so I think I will need do the same for this Quilting LoRA and redo the training. I do not enjoy captioning 🙄, but the automated tools I've tried for other models did not meet my needs.

Long way of saying there will be a new version. Soonish 😀

Oh, while I remember, the initial 3 512x512 images in the showcase are the automated ones that the CivitAI trainer produces. I must create some new ones!

= = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = =

A CivitAI-trained Flux.1.dev LoRA that might - or might not - mimic my Embroidered Quilting TI XL

https://civarchive.com/models/554798/embroidered-quilting-ti-xl

Initial version. It might need the trigger cs-qu1lt. Or not.

This LoRA should do a world-morph of your prompt ---> a quilted art piece.

[PUBLISHEDTOCIVITAIONLY]