Strappy Heels

This LoRA aims to produce better open-toe strappy high heels with visible toes and more stable strap construction. It supports ankle-strap heels, criss-cross heels, T-strap heels, multi-strap heels, lace-up heels, wrap-around heels, cage-style heels, platform strappy heels, and straps extending around the ankles or lower calves.

You can prompt different heel types such as stiletto heels, block heels, chunky heels, sculpted heels, and platform heels. Decorative details like buckles, rhinestones, pearls, studs, bows, tassels, chains, metallic finishes, patent leather, satin, suede, and glossy materials can also be described directly in the prompt.

The trigger word is better_strappy_heels. I recommend using it when you want the LoRA to actively shape the shoes, especially for complex strap designs. For subtle support, natural language descriptions of strappy high heels may already help, but the trigger gives more consistent results.

Important: For difficult visible-toe shots, this LoRA can optionally be combined with Pretty Sandal Heels at a low strength (0.3 was a good spot for me). Better Strappy Heels focuses on strap structure, lace-up and wrap-around designs, cage heels, T-straps, platforms, and decorative strappy heel variations, while Pretty Sandal Heels can help polish toe anatomy and open-toe sandal details.

Dataset and training data:

• 224 images

• 10 steps per image

• 10 epochs

• bucks with 1152px max size, no auto scaling

• cosine scheduler with 0.2 warmup and 0.7 decay

• Captioning was done with OpenAI

• alpha == dim: 32

Closed Heels

This LoRA aims to produce better closed-toe pumps and heels with different shapes and colors in different angles. You can prompt for the toe-type, e.g. pointed-toe, rounded-toe and square-toe. Possible heels are pumps, heeled (ankle) boots, Mary Jane heels/pumps and slingback heels/pumps, all also as ankle-strap variants. For the heel you also have different options like stiletto heel, chunky heel, block heel, stacked heel or sculpted heel (Probably the weakest option so don't rely on that). You probably don't need to include the trigger if you want this LoRA only to support the generation, the captioning was done with natural language, so the effect should be there.

Dataset and training data:

• 268 images

• 5 steps per image

• 10 epochs

• bucks with 1152px max size, no auto scaling

• cosine scheduler with 0.2 warmup and 0.7 decay

• Captioning was done with OpenAI

• alpha == dim: 32

Would love if you post your creations here with it.

The LoRA was trained on the default flux1-dev model for maximum compatibility.

If you use it with other loras, for example a character lora, try to lower the weight - I had good results at around 0.4.

As always: I did not polish any of the images in the showcase, they all came out like this.