Loading
Have a ComfyUI workflow? Get a readable Workflow Doc in your browser.Make a workflow readable.Try it now ↗
A TI (Textual Inversion) embedding (uses 8 tokens in your prompt) to make embroidered quilting images. SDXL seems to have had a reasonable amount of crafting images related to quilting, embroidery, sewing and fabric...
Model source
Source description
2024-JUL-03
A TI (Textual Inversion) embedding (uses 8 tokens in your prompt) to make embroidered quilting images. 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 a TI to leverage the existing content in a more consistent way. It 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.
The trigger word is cs-qu1lt-1 (note the “1” instead of “i”), 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!
Things to be aware of:
Sources
1 sourceModel artifacts
1 artifactSource excerpts
1 excerptSource context: 498 downloads · Type TextualInversion · Base model SDXL 1.0
The TI usually makes a (fabric) border around the image even though there are no such borders in the training images. This is typical of real quilted art and must be to do with the images SDXL was originally trained on.
Simplistic prompts should give mostly quilting with some embroidery, more complex prompts should give more stitching/embroidery.
Even when the rest of the scene is quilted, people (faces/hands/bare skin) can be photographic. I guess that’s a factor of SDXL being trained on so many photoreal images.
Sometimes the TI will make a 3D scene with a mix of quilted/unquilted elements rather than a flattish quilted image.
By the nature of the target technique, images are simplified compared to prompting for photorealism.
If you prompt with just the TI trigger, you should get a random quilted/embroidered landscape - 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-qu1lt-1, summer meadow with hills in the background, sunny, puffy clouds
cs-qu1lt-1, Victorian drawing room, fireplace, mirror, Christmas
cs-qu1lt-1, 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 quilted scene to scenes with fabric/sewn elements. Some of these were really interesting and I got quite distracted during testing :-)
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 with this TI:
Magie_Noire v4
Crystal Clear XL
Clarity XL
ZavyChromaXL v8
Note that some checkpoints respond well to fabric crafting terms without additional TIs or LoRAs. A good example is @Marielle 's Magie_Noire series:
I’ve done some examples on Magie_Noire ver3 without using this TI:
[PUBLISHEDTOCIVITAIONLY]
