A Lora for the creation of images with Clear or Translucent Acrylic High Heels.
Different parts of the shoe can be transparent or opaque. Customizable parts include The Platform, The Vamp, The Heel, The Ankle Strap.
Different types of High Heel styles can be customized with transparency. Block Heels, Stiletto Heels, Wedge Heels, Pumps, Pointed Toe, Mule Heels.
The Lora will try to render a realistic image, if this lora is overpowering your style I have found that lowering the strength of the lora helps. I have found that keeping it somewhere between 0.2 and 0.5 allows any other styles to come through.
Creative Commons licensed Images were sourced from Wikimedia Commons (There is some wild stuff there), and Flickr.
Description
FAQ
Comments (7)
This lora will change the style of painting
Try reducing the strength of the Lora, and it should hopefully fix that.
any plans for anima update?
tyvm for sharing data
Your approach is highly intriguing. With 777 images ranging from foot focus to full-body shots and various poses, I believe that training over multiple epochs with a low repeat setting will yield excellent model generalization and minimize artifacts.
I'd love to share my workflow for this same shoe style for comparison: I ran all my images through SeedVR and Flux2Klein9B to upscale skin and material details, and then annotated them using a combination of 'custom universal tags' alongside VLM-generated natural language captions tailored to those tags. However, after curating and preparing about 200 images this way, the workload became a major bottleneck.
Looking at how successfully you handled a larger dataset, I’m thinking about combining our two mindsets into a hybrid strategy:
1.Detail-focused subset: Take the 200 selected images, process them with my workflow, and set a higher repeat count to lock in the intricate material details.
2.Generalization subset: Follow your approach by gathering a large volume of similar shoe images, batch upscale them with tools like SeedVR, and assign them a lower repeat count to maintain dataset diversity.
3.Mix: Mix these two subsets together for training.
This way, the model might get the best of both worlds: the robust generalization from your massive dataset style, and the crystal-clear texture from the upscaled workflow. What are your thoughts on this hybrid approach?