Corset Shirt - Versatile Fashion LoRA

There are some pieces of clothing that just get it. They have that certain something—the power to transform a simple outfit into a statement, and a good day into a great one. This is one of those pieces. It’s designed with the timeless, elegant structure of a corset, so it gives you that incredible silhouette, but it has all the softness and ease of your favorite shirt. It's that rare, perfect balance of strength and grace.

The real magic, you'll find, is how it effortlessly adapts to your story. It’s the kind of top you'll reach for when you want to feel powerful in a business meeting, effortlessly chic at brunch, or subtly alluring on a dinner date. It never screams for attention, but it always holds it, quietly and confidently. More than just a piece of clothing, it becomes a way to express a mood or an attitude, making every look feel like a true and unique reflection of you.

How to Use This LoRA

Hint: When you are looking for prompt inspirations look at the Flux.1[dev] version, as it has many sample images.

This LoRA is designed for flexibility and ease of use. It responds well to descriptive prompts, allowing you to control the material, color, fit, and styling with precision.

Activation & Trigger Word:

The main trigger word is: corset2shirt

Controlling the Appearance:

Simply describe what you want. The model understands a wide range of materials and colors.

• For Material: black cotton corset2shirt, blue denim corset2shirt, white broderie anglaise corset2shirt, leather-look corset2shirt, ...

• Controlling the Styling: "tucked into" or "worn over" will (mostly) work

• The model is trained on the lacing detail. To emphasize it, add phrases like "showing the intricate back lacing", or "with ribbon lacing down the back"

Training details

All training was conducted with SimpleTuner.

Krea 2:

• When you want to use it with ComfyUI, please use https://github.com/Comfy-Org/ComfyUI/pull/16857 (and give that PR a 👍 so that it gets pulled quickly);
  using it with diffusers should work out of the box.

• LoKR with factor=28 (Krea2Attention=28, Krea2SwiGLU=28), linear_dim=10000, linear_alpha=1, use_scalar and full_matrix as well as DoRA (dora_wd=true)

• 22388 steps (11194 real steps * batch size 2; gradient accumulation 2)

• Trained with Krea 2 Raw, to be used with Krea 2 Turbo

• training time: 16.5 hours on one 5090, including validation image generation

• regularization images were used

• masked training with the faces masked away

• each image had multiple captions for versatility

Flux.1[dev]:

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

• 8100 steps (2025 real steps * batch size 4)

• training time: 5.5 hours on one L40S, including validation image generation

• regularization images were used

• masked training with the faces masked away

• each image had two captions for versatility