Art Official

I feel it too, that we've been here before, but when gathering material to use for training the LoRA it's always a different experience, and yea, it will end up as a dud sometimes, but this one didn't.

While it may appear to be similar to some others I've done let's break down the process and results. While scraping images for something to inspire me I had a light bulb moment, why not dig into my old generated images that I've saved over the couple of years that we've been able to do these things, and see what I can find.

I didn't remember that SDXL had eye popping beauty and, while digging through my old generated images, I found a set that gave me the inspiration I needed, but it didn't make sense that they even existed, nor do I recall making/saving them, and that they are SDXL?

Anyway, I had accumulated thousands of images that I liked, and usually I only save those that are the most moving, and these were no exception. In fact, they were so unique that it took about 1k images generated to hit one that had similar appeal, though I had amassed hundreds in this particular folder, and so this must be the reason why SDXL impressed me so, when looking back, I simply had been saving the most outstanding of those I had generated.

This LoRA comes from a set of colorful cyber something or other female subjects in mostly sci fi scenarios. They were great already but something I did differently than my typical process was, instead of creating a Chroma LoRA first, in order to generate much clearer images with good prompt tickling, I tossed them into Qwen Image Edit (2511) and spiced them up that way. It was faster to do it this way, of course, and the results were far superior.

From that process I created 78 images of 2 genre in 2 variations each, resulting in 312 images total. The subjects I targeted were cyberpunk+cyborg with "normal" and dark themes and synthetics like toy esthetic with rubberized or silicone surfaces for both "normal" and dark themes as well.

I created a pipeline to swap environments each time to generate substantially different images, picking text from 200 previously captioned images for this process, as well as changing poses in a similar fashion, offering color and theme adjustments depending on the specific target I was working with.