This is a Flux version of my TI embeddings:

For Stable Diffusion v1.5:

https://civarchive.com/models/135403/crystaliti

For Stable Diffusion XL:

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

This time, since the target is Flux, I have made a LoRA using CivitAI’s trainer. It uses mostly the same dataset as my SDXL Crystal TI, but with some reprocessing and basic captioning. Amazingly, given that LoRAs and TIs are very different, it works reasonably well. I hope to revisit this as I learn more about Flux... and if I can clear just a few things of the To-Do list!

I have tried to build in an “intensifier” trigger. Flux doesn’t seem to like triggers the way SDXL does, but it is possible to make a usable token that didn’t previously exist in Flux.1 DEV. For this LoRA, the intensifying trigger is “cs-cr1stal” without the quotes. And, of course, you can use any crystal-related real words to guide Flux as usual.

In this initial version, the LoRA+trigger will not necessarily do total conversion, but even if it doesn’t most of your image should still be Crystalised.

Below are a few grids showing some with-without images. They are XYZ grids made in Forge. For each grid of 4 images:

top left: cs-cr1stal + LoRA

bottom left: LoRA only

top right: cs-cr1stal only

bottom right: neither

Each grid uses a simple prompt such as “fox” or “king”, followed by the with/without variations.

Hopefully the grids will help you decide whether you want to try this LoRA or not.

The showcase images use more complex prompts (Tarot card descriptions + text) to push the LoRA a bit further. They’re all made in Forge and may use Hires.fix, but nothing else: no other LoRAs, ControlNet, detailer/FX plugins or post-processing.

[PUBLISHEDTOCIVITAIONLY]