Ponygirl pony-play gear for Qwen-Image 2.1. A <a target="_blank" rel="ugc" href="https://civarchive.com/models/2997311">Krea 2 version</a> is posted separately.<h3 id="what-it-draws">What it draws</h3><ul><li>Leather or polished steel bridles with blinders and a bit gag</li><li>Reins running from the bit rings</li><li>Pony harnesses of leather straps across the chest and waist</li><li>Hoof boots</li><li>Collars and posture collars</li><li>Arms behind the back or in an armbinder</li><li>Being harnessed to pull a two-wheeled sulky</li></ul>Trigger word: <code>ponygirl</code><h3 id="how-to-prompt">How to prompt</h3>Write full sentences rather than tag lists. The training captions name the gear in this order, so reuse it:<blockquote>Ponygirl is standing ... She is nude. Visible pony gear includes a black leather bridle, blinders, a bit gag, reins, a pony harness of leather straps across her chest and waist, and hoof boots.</blockquote>Then describe hair and body, the camera, and the lighting.For full-body shots, say how far the camera is, or the feet tend to be cropped:<blockquote>The camera is at eye level and far enough back to show her whole body from the front, from the top of her head down to her hoof boots, with space above her head and below her feet.</blockquote>For the sulky, say that she stands between the shafts and the seat is empty, or she may end up sitting in it:<blockquote>She stands upright between its two long shafts, in front of the sulky, leaning forward to pull it ... The empty two-wheeled sulky is behind her, and its seat is empty.</blockquote>Recommended strength: 0.8 (0.6 to 1.0).The comparison images show this LoRA next to the Krea 2 version with the same prompt and the same seed. Every prompt is written under each image and stored in the image metadata.<h3 id="known-issues">Known issues</h3><ul><li>Hoof boots often come out as lace-up boots with hoof-shaped soles.</li><li>Armbinders are sometimes drawn on one arm only, or with the arms hidden.</li></ul>

Ponygirl - pony play gear (Qwen 2.1)

Qwen-Image 2.1 version, trained with ai-toolkit. Recommended settings <ul> <li>Strength 0.8</li> <li>25 steps, CFG 4, sampler er_sde, scheduler beta</li> <li>Short negative prompt: <code>extra legs, extra limbs, extra fingers, fused fingers, deformed hands</code>. For sulky shots add <code>sitting, seated, sitting in the sulky, riding, wheelchair, driver</code>.</li> <li>Works at 2048 px on the long side (1632x2048, 1152x2048, 2048x1376).</li> </ul> Training <ul> <li>1,701 captioned images, natural-language captions</li> <li>3,000 steps, batch 1, LoRA rank 32 / alpha 32</li> <li>AdamW 8-bit, learning rate 1e-4, bf16</li> <li>Buckets 768 and 1024, caption dropout 0.05</li> <li>Text encoder not trained</li> </ul> Changes from v2: a larger dataset with more sulky, armbinder and bit-gag close-up images.

Files with SHA256: 2e9cc2e2d9f3b0d07b33b646919589b070126ad0a90c1bf7994c324353ea9109