<h3 id="a-lora-for">A LoRA for</h3><h3 id="women's-shoes-and-boots-of-every-kind:">women's shoes and boots of every kind:</h3><h3 id="pumps-sandals-mules-wedges-boots-up-to-the-thigh-and-ornate-jeweled-fantasy-shoes.-use-the-trigger-word-alchshoes-right-before-the-shoe-type.">pumps, sandals, mules, wedges, boots up to the thigh, and ornate jeweled / fantasy shoes. Use the trigger word <code>alchshoes</code> right before the shoe type.</h3><h2 id="alchshoes-women's-footwear-lora-for-krea-2">alchshoes — Women's Footwear LoRA for Krea 2</h2>A LoRA for women's shoes and boots of every kind: pumps, sandals, mules, wedges, boots up to the thigh, and ornate jeweled / fantasy shoes. Use the trigger word <code>alchshoes</code> right before the shoe type.What it does well<ul><li>Realistic, well-shaped heels: stilettos, block heels, platforms, wedges.</li><li>Jeweled and fantasy shoes: crystals, chains, filigree, sculpted ornaments. This is where it shows its strongest effect over the base model.</li><li>Strappy and lace-up sandals, wedges, tall boots.</li><li>Glossy patent leather, satin, suede and metallic finishes.</li></ul>What it does NOT do (honest notes)<ul><li>It is not a miracle. The base model already draws classic shoes well. This LoRA mainly pushes style and detail.</li><li>It leans toward sandals. The dataset contains many open-toe shoes. With closed classic shoes (closed mules, pumps, peep-toes) it may add straps or open the shoe up. For those, lower the strength to 0.4–0.5 or turn it off, and always write "closed toe" in the prompt.</li><li>Sometimes it makes feet too veiny or bony at high strength.</li><li>Small shoes in wide full-body shots get little detail. Use a legs shot when the shoe matters.</li><li>It does not reproduce specific brands or designer models.</li></ul><h3 id="usage">Usage</h3>Trigger and prompt structure. Follow the caption format the LoRA was trained on:<pre><code>a pair of [color] [material] alchshoes [shoe type] with [heel] and [toe/detail]
</code></pre>Example:<pre><code>Full-body shot of a woman standing in a sunlit city street, navy knee-length dress, a pair of black patent leather alchshoes pointed pumps with stiletto heels, closed pointed toe, no straps
</code></pre>Shoe types it was trained on. Write the class name right after <code>alchshoes</code>.<ul><li>Pumps &amp; mules: pointed pumps, peep-toe pumps, slingback pumps, mary jane shoes, heeled mules</li><li>Heeled sandals: ankle-strap heeled sandals, lace-up heeled sandals, thigh-high lace-up sandals, platform sandals, wedge sandals</li><li>Flats: flat sandals, slides, ballet flats, loafers, sneakers, espadrilles</li><li>Ornate: jeweled fantasy shoes</li><li>Boots: ankle boots, heeled knee-high boots, over-the-knee boots, thigh-high boots, lace-up heeled boots, lace-up combat boots, platform boots, sock boots, western boots, biker boots, riding boots, slouch boots, rain boots</li></ul>Recommended settingsStrength0.8 (0.9–1.0 for jeweled/fantasy, 0.4–0.5 for closed classic shoes)FormatPortrait, about 1.5–2 MP (e.g. 1088×1936). Landscape tends to produce two women side by side. Very high resolutions (6 MP) create texture artifacts on the floor.Framing"Full-body shot" for the look; "Legs shot" or "Legs shot from the hips down" for shoe detail.Tips<ul><li>Don't write "sexy": describe what makes the shoe sexy (very high stiletto heels, thin straps, glossy patent leather, platform sole).</li><li>For fantasy designs, use <code>featuring</code> and place the ornament on the shoe: "featuring a sculpted gold dragon coiling around the heel". This works much better than "with a dragon".</li><li>Use one theme per shoe.</li><li>For lace-only sandals, write <code>made only of thin round cord laces … bare skin visible between every lace</code>, otherwise the model adds wide bands.</li><li>For tall lace-ups and boots, add <code>bare legs</code> and a short dress or a high slit so the shoe isn't hidden.</li></ul><h3 id="training-details">Training details</h3><ul><li>Base: Krea 2. Tested on a Krea 2 finetune, 12 steps.</li><li>Dataset: 405 real photos, no renders or illustrations.<ul><li>The photos are mostly worn shoes, with some product shots.</li><li>Captions use a fixed class vocabulary.</li></ul></li><li>Training: musubi-tuner, LoRA rank 32 / alpha 32, LR 1e-4, AdamW8bit, 3750 steps, resolution 768.</li><li>Released file: the final checkpoint. It was compared at a fixed seed against the 2000 / 2500 / 3000 / 3500 checkpoints and the no-LoRA output. The final had the cleanest shapes and feet.</li></ul>Feedback and example images are welcome, especially failures: they help the next version. 🙏

Files with SHA256: 77f509b1b0865cc440e2d4523271a3bf737f759ea457758a8752fced2ebc8b69