ghost
Sign In

CrystaliTI

66
682
189
13
Updated: May 8, 2024
stylegemcrystal
Verified:
SafeTensor
Type
Embedding
Stats
185
189
Reviews
Published
May 8, 2024
Base Model
SD 1.5
Training
Steps: 8,000
Trigger Words
cs-cr1stal-18
Hash
AutoV2
96FCFC266F
default creator card background decoration
ODOR Badge
chromesun's Avatar
chromesun

UPDATE: 2024-MAY-07

A complete redo-from-start. If you liked the v17.08 one, keep it around because this v18 is very different.

Trigger: cs-cr1stal-18

This is the SD15 version of my SDXL crystal one at:

https://civitai.com/models/139230/crystalitixl

I’ve tried to make it give a similar effect as the SDXL one within the limits of SD15. Being an SD15 model, it’s harder to control. Generally you shouldn’t need word salad negative prompts - mostly I’ve just used the easynegative embedding and perhaps a couple of words to get SFW output.

For short prompts you should get a crystalline version of your subject. The effect should be a mix of large crystalline facets + small bling type crystals.

For longer prompts the embedding should still try to change the main subject in the prompt, and then partially affect the setting. It does like to scatter faceted gems around the place. In testing I found that in longer prompts the crystal effect could become more opaque and metallic, sometimes feathery.

The showcase images use some short and some medium length prompts. Long detailed prompts can give some great results but the hit rate wasn’t high for me.

Weighting from 0.8 <---> 1.2 worked fine but I found little need for weighting in automatic1111 webui / FORGE. It was easier to move the trigger right or left in the prompt. Not tested with ComfyUI.

Generally I’ve been using it with:

DPM++ 2M Karras

width x height: 512 x 512

30 <---> 50 steps

cfg from 7 <---> 10

Hires.fix:

4x_NMKD-Siax_200k or 4x-UltraSharp scaling of either 2 or 2.5, steps=15, denoise=0.4

Models:

mostly tested with:

avalonTruvision v4

https://civitai.com/models/13020?modelVersionId=256774

crystalClearPro v1

https://civitai.com/models/223302?modelVersionId=251930

faetastic v2

https://civitai.com/models/14065?modelVersionId=105796

vendoPhotorealistic 2 v1

https://civitai.com/models/83490?modelVersionId=88738

I feel the crystalClearPro output was slightly more consistent across different types of prompt but, hey, that’s purely subjective!

[PUBLISHEDTOCIVITAIONLY]

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

This is a TI embedding for Stable Diffusion v1.5 that adds a crystal and/or gem effect to your generated images - everything from spiky crystals to flat stained glass The crystal/gem effect is based on realistic jewels and crystal growths rather than cartoon/manga style.

The inspiration for this TI was the 1985 scifi novel “Sentenced to Prism” by Alan Dean Foster.

Good pun, better book :-)

https://en.wikipedia.org/wiki/Sentenced_to_Prism

Training was run on the base Stable Diffusion v1-5-pruned.ckpt [e1441589a6].

My personal preference is for surreal/fantasy images, so that’s what this TI is aimed at.

All of the showcase gens are done with

Vendo Photorealistic 2 v1

https://civitai.com/models/83490/vendo-photorealistic-2

or

Vendo Universal v1

https://civitai.com/models/119469/vendo-universal

I’ve tried this TI on a dozen or so other models and it seems to work reasonably with them all, but I prefer the Vendo ones at the moment.

This TI works with short and long prompts. Short prompts are quite overwhelming with total crystal conversion. Long prompts vary between adding crystalline effects globally, to just adding jewellery/gems here and there. For strength adjustment I usually just move the TI token forwards or backwards in my prompts. When I’ve tried emphasis syntax, values from about 0.85 to 1.15 seem to be OK.

The usual SD v1.5 issues apply: far-away faces are not great, hands are variable, eyes can be odd, limbs can be doubled/missing/merged. Using your favorite negatives will help, as will inpainting of course.

The only extra problem I’ve been getting with this TI is a tendancy to orange for shorter prompts / higher emphasis. Not sure why because the training images don’t seem to have too much orange in them. Whatever the reason, adding orange to the negative prompt on a case-by-case basis mostly sorts it out.