2024-JUL-10
A TI (Textual Inversion) embedding to make decorative leather images with mixed media elements.
This one uses 8 tokens in your prompt. SDXL seems to have had a reasonable amount of crafting images related to quilting, embroidery, sewing and fabric art in its training set, but I’ve struggled to find consistent wording to create new images. So I thought I’d try making some TIs to leverage the existing content in a more consistent way. This TI should work with base SDXL and any checkpoints that are not too far away from the base.
Initial version, trained on base SDXL v1.0. This TI uses 8 vectors and is 33KB in size.
The trigger word is cs-l3ath3r, but you can change that simply by renaming the .safetensors file. If you do, try not to use a real word that SDXL already knows!
Simple prompts should give 2D applique-style results, complex prompts should give 3D diorama-style images.
Things to be aware of:
* By the nature of the target technique, images are simplified compared to prompting for photorealism.
* The TI often presents the image with a leather border or on a leather cushion even though there are very few such borders in the training images. I think this must be to do with the images SDXL was originally trained on.
* Even when the rest of the scene is leathery, people (faces/hands/bare skin) can be photoreal or plastic. I guess that’s a factor of SDXL being trained on so many photoreal images.
* Even with “signature,logo” in the negative prompt (and no sigs/logos in the training images), the TI often adds a sig at bottom right. Grrr.
As well as embroidered/stitched leather, suede and other fabrics, the TI was trained with:
rivets, sequins, rhinestones, flatback gems, embroidery jewels
and various sorts of beads: glass, pearl, faceted, plated, seed, bugle, Delica.
If you prompt with just the TI trigger, you should get a random tree/woman/flower - some will be scenes, others will be more like patterns. For more directed prompting I found that this form usually works:
the TI trigger, scene description, secondary descriptions
Examples would be things like:
cs-l3ath3r, unicorn, galloping
cs-l3ath3r, sheep, grassy meadow, wildflowers
cs-l3ath3r, sailing dinghy on a lake, distant mountains, stormy sky
Moving the TI token rightwards in the prompt (or reducing its weight) causes a shift from leathery scenes to scenes with some leather elements.
Testing and showcase images done in Forge version: f0.0.17v1.8.0rc-latest-276-g29be1da7.
The showcase images are from 10 simple prompts and 10 complex prompts.
Sampler: testing worked well with...
DPM++ 2M Karras 25<--->50 steps, generally I used 30 or 40 in testing
CFG scale: 5<--->10, generally I used 6 or 7 in testing
Mostly used Hires.fix at either 1.25 or 1.5 to increase the detail a bit.
Hires.fix steps 15<--->25, 4x_NMKD_Siax_200k or your favorite, denoising 0.4
I didn't use any other adjusters/controlnet/i2i/post-processing for the showcase images.
During testing I kept the negative prompt as simple as possible, e.g.:
closed eyes,signature,logo
Checkpoint models that worked well when I tested this TI:
Clarity XL
https://civarchive.com/models/471585/clarity-xl
Magie_Noire v4
https://civarchive.com/models/505656?modelVersionId=612138
Crystal Clear XL
https://civarchive.com/models/122822/crystal-clear-xl
ZavyChromaXL v8
https://civarchive.com/models/119229?modelVersionId=563988
Note that some checkpoints respond well to fabric crafting terms without additional TIs or LoRAs. A good example is @Marielle’s Magie_Noire series:
https://civarchive.com/models/505656
[PUBLISHEDTOCIVITAIONLY]
Description
Decorative leather and mixed media - Initial version
This is a TI so you can change the trigger word simply by renaming the .safetensors file.
[PUBLISHEDTOCIVITAIONLY]
FAQ
Comments (10)
Какая чудесная история, впрочем, как и все. Какой полёт фантазии и воображения! Это очень здорово!!!
Thank you, I am glad you like them 😀
It's really nice to discover this resource, very useful and it gives great results, thx.
Thank you - I really appreciate you and all the Daily Challengers taking a hard look at the model. I've been blown away by some of the entries today... very creative. And as someone pointed out on the challenge page, it doesn't cost extra in the generator 'cos it's an embedding. It (and the other crafting ones) was originally made to help out some crafter friends come up with test ideas quickly.
I saw your comment on the Challenge post. I posted my wet leather cow here too.
Thank you! I appreciate you making the effort. Love the cow too 👍
In my experience the water marks and signatures (like are mentioned in this model description) are part of CLIP alignment and are directly related to any wording that indicates that the prompt is not for a real image. This is a very common misconception people have with prompting on Civitai. Never use any keywords like: "realistic, photorealistic, realism, photorealism, photo, camera, or anything related to photography or videography. The syntax for the output from diffusion is specifically the "image". This is how CLIP was trained and the only word that does not come with additional context for errors. If the model errors into any form of traditional or digital media, simply add "real" to the prompt.
There is no free lunch here. "Real" means in the real world happening right now. There are QKV alignment ethics and morals that come into scope slightly differently when the prompt is for "real". Cultural norms come within more of the scope of how you do or do not define them as opposed to the model inferring cultural norms from some other assumed scope associated with some type of traditional media.
There are no errors in diffusion. That is crazy bold for me to say, but if you explore CLIP alignment enough, it all exists for a reason.
Thanks for the detailed comment. I can't say much about the theory - what I wrote here was simply from empirical testing and guesswork. Once I got a version of the embedding that produced what looked like photos of real artworks such as I've seen in galleries and craft fairs, I stopped thinking about it and just used it. The goal was to make a model that I could use to produce images for greetings cards quickly and easily. When sigs or watermarks appear, 10 seconds with a clone brush fixes that. 🙂 I'll bear in mind what you've said, and next time I get a batch with loads of sigs/watermarks I'll try adjusting my prompt.
chromesun Is Clarity XL the main model you used with this or have since?
DudeWTF It's one of my preferred models for general purpose SDXL imaging, along with CCXL and Jug. The embedding is trained on base SDXL of course.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.