I have noticed that Flux sometimes struggles with Hotpants and generates normal pants instead. So I have decided to create this LoRA. Hotpants Collection is a LoRA focused on short, fitted hotpants in different cuts, lengths, and materials.Main trigger:hotpants_collectionThe LoRA was trained to cover several length categories, so you can control the overall cut more directly in your prompts:<ul><li>regular hotpants</li><li>short hotpants</li><li>very short hotpants</li><li>micro hotpants</li></ul>It also supports different rises and styles, for example:<ul><li>high-waisted hotpants</li><li>mid-rise hotpants</li><li>low-rise hotpants</li></ul>Supported materials and variations include:<ul><li>denim hotpants</li><li>faux-leather hotpants</li><li>leather-look hotpants</li><li>cotton hotpants</li><li>sporty stretch hotpants</li><li>ribbed knit hotpants</li><li>sequined hotpants</li><li>tweed / tailored hotpants</li></ul>Visible design details can also be prompted, such as:<ul><li>rolled hems</li><li>frayed hems</li><li>side lacing</li><li>front buttons</li><li>visible pockets</li><li>belt loops</li><li>glossy finish</li><li>matte finish</li></ul>This LoRA works best when the prompt clearly describes the hotpants cut and length instead of using only generic words like “short shorts”.Examples:<ul><li>hotpants_collection, high-waisted regular blue denim hotpants with rolled hems</li><li>hotpants_collection, low-rise short white cotton hotpants with front buttons</li><li>hotpants_collection, high-waisted very short black sporty stretch hotpants</li><li>hotpants_collection, high-waisted micro glossy black faux-leather hotpants</li></ul>Important note about framing:The dataset contains many waist-to-thigh and lower-body focused images. Because of that, the LoRA can sometimes encourage tighter framing. If you want a clean full-body image, explicitly describe the framing in your prompt, for example:<ul><li>full-body photo</li><li>her entire head, both legs, and both feet fully visible</li></ul>It also helps to describe the face and hair when you want uncropped portraits or full-body shots, for example:<ul><li>looking directly at the camera</li><li>calm confident expression</li><li>lips slightly parted</li><li>long wavy hair</li></ul>Another important note: Some combinations of material, length category, and wording may occasionally result in normal long pants instead of hotpants. This is difficult to predict reliably, but in most cases it helps to generate a few more images or adjust the prompt slightly.During testing, I only encountered 2 failures with long pants out of 40 generated images. One potentially difficult combination is “regular denim hotpants”, because Flux tends to interpret “regular denim” as regular full-length jeans.I decided not to force this behavior too aggressively during training, because that would require training strongly against the base model and could make the LoRA unnecessarily strong or less flexible.This is, just like my other clothing LoRAs, not a “single outfit LoRA”. Instead, I trained the design language of hotpants made out of different fabrics and lengths. Would love to see what you create with it.The LoRA was trained on the default flux1-dev model for maximum compatibility.Dataset and training data:<ul><li>177 images</li><li>5 steps per image</li><li>10 epochs</li><li>bucks with 1152px max size, no auto scaling</li><li>cosine scheduler with 0.2 warmup and 0.7 decay</li><li>Captioning was done with OpenAI</li><li>alpha == dim: 32</li></ul>If you use it with other loras, for example a character lora, try to lower the weight - I had good results at around 0.4-0.5.As always: I did not polish any of the images in the showcase, they all came out like this. I did 2 images per prompt and picked the better looking one.

Files with SHA256: 06f2015298abae43e5c7ca60a85eeba2cdf067fa3b373e633a196cf1933f536c