Sharing another personal project—this time featuring Wanda Maximoff. I wanted to create one of Wanda herself, not her as the Scarlet Witch in the default outfit. Finding good source images for this was like hunting unicorns.

This LoRA turned out to be the most finicky, tricky, pain-in-the-butt project I’ve ever attempted. I hand-picked 70 curated images, edited each one in Photoshop to make the dataset at least somewhat usable, and trained for 5 epochs at 10 steps each. The source material was low quality, and I did what I could to clean it up.

After baking, this LoRA became temperamental and nearly ended up in the bin three times. I only kept pushing through because of the many days I’d already spent on it. Testing was a bit of a nightmare—it got a little "schizo" at times—but it seems to be somewhat tamed now.

It works on most, if not all, Pony models as well as most SDXL models—though each model produces different results, so please experiment. If the LoRA strength is set too high, parts of her default outfit may ghost through; you can counter this with negative prompts. Some models do well with a strength of 1.0, while others may require up to 1.5, however most success so far have been with strength 1.1. There's no hard and fast rule here—play around and see what works.

Since I’ve put so much time and effort into this, an update is inevitable—so watch this space. Feel free to share your creations and any tips you discover along the way. I’m still experimenting too, but I’ve run out of time for now.

For V1:

Trigger: wanda

CFG: 3-6 (feel free to experiment)

Sampler:

Euler @ 10-20 steps, 15 seems to be the sweet spot

DPM++2M @ 10-20 steps, again 15 seems to hit the mark

Euler A @ 15 steps seems to work as well

Important: loRA strength of 1.0 - 1.5, though try and stick to 1.1