Spider Gwen / Gwen Stacy

The model is intended to help produce vibrant illustration/3d like images of Spider Gwen / Gwen Stacy.

You're more likely to get the unmasked character using the term Gwen Stacy and the masked character with Spider Gwen. If you're trying to get a character without the outfit at all then don't use either and just describe the clothing without the "who".

Version 1.1 leans more towards crisp, contrasting, eye popping illustrations and moves away from the 3d effect, though lower weights can achieve the latter as well.

I'll often use Chroma as a target stage for LoRA from which I'll then produce a dataset for the eventual target, in this case Z-Image, but there are times when I'll target Illustrious for this instead, especially when deriving my set entirely from that one stage. In this particular case I chose a refine of Illustrious.

When I posted this LoRA I still didn't know the origin of the Illustrious model I've been using for my datasets but that's been resolved and, though I still have the uploaded version, I've pointed to the resource in order to give credit, see iLustMix (v3) in the resources. Also, there's other versions of that mix and I'll be testing a few of them, though v9 seems to have a very different shift and I'm kind of liking it.

Whenever I'm testing the Danbooru or e621 set of trigger terms for interesting characters, outfits or actions I'll go to this first, which has been identified as "iLustMix" and now provided in the resources,to make sure it's in Illustrious before pulling my hair out.

Choosing a themed set of images to produce an Illustrious LoRA helped me guide the results towards the goal. Then the dataset was produced using Illustrious to generate the images for the Z-Image target, of which 218 were used. Letting the seed determine pose for a nice random action set and then running through deliberate poses, facial expressions, orientation, masked and unmasked and giving a good amount of clothing variety aught to allow for a good enough base to explore your creativity.

The images were then cropped for focus and then any remaining artifacts were scrubbed. I trained 3 sets with 20 epochs, bringing us just past 6k steps per, with various rank and alpha, settling on 32/1 for its superior detail. From that I generated a few thousand images to determine flexibility and went back to the drawing board to add the remaining images, from 174 to 218, landing on 33 epochs for 7194 steps, note that the trainer (OneTrainer) uses zero indexing so the resulting 7193 is actually 7194 steps.

Too much info? Why did you keep reading? :P~

-

Gwen Stacy's first appearance in media was in Marvel Comics' The Amazing Spider-Man #31, published in December 1965. It's interesting to know this because I've been a Spider-Man fan forever and didn't know about Gwen until that Spider-Verse movie came out, which I mention next.

The character portrayed by this LoRA intends to bring back the esthetic of this fictitious character before the release of Spider-Man: Beyond the Spider-Verse", though with a more modern esthetic and 3d shine.

To generate eye popping images you'll want to add a lot of detail and even describe the things that may appear obvious to you, detailing their colors, feel, texture.

Often times I will use the caption I trained with in order to bring back the essence of the image that it was trained from but this time the images were manually prompted so, instead, I grabbed a bunch of images, from the net, of Spider-Gwen (Stacy) in different actions and scenes then joy-captioned those to produce detailed descriptions for inference, then adjusting those captions as needed to produce prompts for the examples.

-

v1.1 brings us to step 10027, sourcing a refined set of images from the same batch in order to steer the model into illustrations vs 3d.