Murder on the Dancefloor boots

Step into the legend. These are the murderboots—the iconic, show-stopping boots immortalized in Sophie Ellis-Bextor’s anthem "Murder on the Dancefloor." Designed for the bold, the fearless, and the ones who own the night, these peep-toe leather stunners are more than just footwear—they’re a statement. With their daring cutouts and high-gloss finish, they’re built to turn heads, spark envy, and make every step feel like a power move. Whether you’re commanding the dancefloor or ruling the streets, these boots are your ultimate weapon of style. Slay in every stride. 💃✨

How to Use This LoRA

Activation & Trigger Word:

The main trigger word is: murderboots

Controlling the Appearance:

Generally you don't need much more than writing murderboots instead to boots. Stating the nail polish color for the visible toes is working fine.

When your generated image is showing closed murderboots instead of the correct peep-toe boots, it can help to explicitly state the toes, especially with their nail polish color.

Sometimes the model mixes up where to place the buckles. They should be on the outside and not on the inside. When that is happening you should generate more images with different seeds (or increase the batch size) and then you'll usually also get fine images. Or fix it with a bit of inpainting.
The same holds for other shoes that are visible in the same image. Although it was made sure that the murderboots concept doesn't bleed, it can happen sometimes. In those cases just regenerate or fix it with a bit of inpainting.

Training details

Qwen:

• LoKR with factor=60 (Attention=60, FeedForward=30), linear_dim=10000, linear_alpha=1, use_scalar and full_matrix

• 8288 steps (2072 real steps * batch size 1 * gradient accumulation 4)

• training time: 7.25 hours on one L40S, including validation image generation

• regularization images were used

• each image had two captions for versatility

FLUX.2[klein] 9B:

• LoKR with factor=16 (Attention=8, FeedForward=16), linear_dim=10000, linear_alpha=1, use_scalar and full_matrix

• 14328 steps (3582 real steps * batch size 2 * gradient accumulation 2)

• training time: 7.5 hours on one 5090, including validation image generation

• regularization images were used

• each image had two captions for versatility