<h2 id="lingerie-that-is-actually-fastened.">Lingerie that is actually fastened.</h2>Base models get lingerie hardware right about a third of the time. The rest of the time they draw it as decoration — a suspender strap ending in a clip that floats above the stocking, the welt edge running straight past behind it, the strap slack, the clip gripping nothing. "detailed, highly detailed, 8k" does not fix it, because the model was never taught what a closed clip does to the fabric it is holding.This adapter teaches the mechanism: the welt edge pulled up into the clip, the pad pressing the band, the strap taut. Same for boning channels that actually shape a panel, busks that close a front, and rings that straps are genuinely threaded through.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/4b1d016b-b49c-4e50-a5e6-b6d95a6cefab/original=true/4b1d016b-b49c-4e50-a5e6-b6d95a6cefab.jpeg" /><h2 id="measured-not-claimed">Measured, not claimed</h2>24 clip junctions across 12 image pairs — the same prompt, seed, sampler and resolution, with and without the adapter, so it is the only variable. Every pair was judged blind by a reader who was not told which side carried it, at 8–15x native zoom, against a written test: the clip frame must sit at the welt edge, the welt must deform into a cusp falling away on both sides, and the strap must be taut. Proximity alone counts as a failure.<ul><li>Base stack, no adapter — 9 of 24 junctions correct (38%). 1 pair won.</li><li>Plus lngcon @ 0.8 — 14 of 24 (58%). 4 pairs won, 7 tied.</li></ul>That is the whole claim, and it is deliberately modest: roughly 38% → 58% at the framing tested — hips and upper thighs, the clip about 30 px across. A LoRA that turned 0% into 100% would be a nicer story; it would not be this one. Gallery image 2 is the with/without pair, same prompt and seed, clip junction enlarged on each half. About one junction in three still needs a re-roll, so look at the clip at native zoom before you keep a frame.Trigger word: <code>lngcon</code> Strength 0.8 is the tested default and the value behind every example image. Range 0.6–1.0: use 0.6 if you are already stacking two heavy LoRAs, 1.0 for difficult seeds at the cost of slightly flatter fabric texture. From 1.2 it starts accreting hardware you did not ask for.Base Krea 2 — trained on Krea-2-Raw, validated on Krea-2-Turbo. 1024x1280 (trained on 768x1024 buckets); holds at 3:4 and 2:3. Turbo: 12 steps, CFG 1.5, sampler er_sde, scheduler beta. Raw: 52 steps, CFG 3.5, euler / simple. UNet-only adapter with no text-encoder weights, so leave Clip Skip at 1 and load it last in the chain. Every example image is this adapter at 0.8 on top of two realism LoRAs at 0.6.<h2 id="how-to-prompt-the-opposite-of-a-style-lora">How to prompt — the opposite of a style LoRA</h2>Write a paragraph, not tags. Training captions are natural-language descriptions averaging ~130 words. <code>lingerie, garter, detailed, masterpiece</code> gets you very little.Name the mechanism, not the garment. Do not write "wearing a garter belt" and hope. Write what the hardware is doing:<blockquote>…each elastic strap runs down the thigh to a small metal clip whose rubber button is pressed through the metal loop against the wide dark welt of a sheer black stocking, the welt edge drawn up into the clip and the strap taut.</blockquote>That sentence is the difference between a clip that grips and a clip that floats.<h2 id="known-issues">Known issues</h2><ul><li>Close to inert where the base is already right. Corset boning and bra cups render correctly without it on a good realism stack. The clip-to-welt junction is where it earns its place.</li><li>It adds hardware you did not ask for on strap harnesses — extra gold or steel sliders at strap junctions. Often the correct adjustable topology; sometimes just more metal. Drop to 0.6.</li><li>It does not fix counting. Ask for five busk hooks and you may get six. Base-model property, untouched.</li><li>Not a style LoRA and not a body LoRA. It does not change lighting, grading, faces or bodies. The dataset is 155 different women.</li><li>Short prompts underuse it badly.</li><li>Sheer means sheer. Lace and mesh briefs frequently render fully transparent — opacity is not what this was trained on. Say it explicitly: "with a fully opaque gusset". Mind the word order: "a sheer lace bra with fully opaque cups" renders a sheer bra, because a contradiction inside one noun phrase resolves by position, not by emphasis — remove the contradicting word rather than correcting it afterwards. The clause is reliable on briefs and unreliable on cups.</li><li>Name every garment you want, including the briefs — then check it rendered. Describe a suspender belt and stockings without naming a brief and you will frequently get nothing between them. Naming is necessary, not sufficient.</li></ul><h2 id="training">Training</h2>Krea-2-Raw, rank 32 / alpha 32, UNet only, 768x1024 buckets, bf16, flowmatch, AdamW, lr 1e-4, ai-toolkit. 155 studio catalogue frames, captions hand-audited at native zoom: 95 of them made a fastening or mechanism claim, and 45 of those 95 were wrong and were corrected or dropped — mostly a caption calling a clip closed where the pixels show it hanging. Training on the uncorrected text would have taught the exact defect this adapter exists to fix.Trained to 2000 steps; this release is step 1000. Binding is achieved by 750–1000 and holds; from 1250 the model starts dragging in dataset hardware it was not asked for. The final step is not the artifact.

Lingerie Construction (lngcon) — Krea 2

<pre><code>Teaches Krea 2 what fastened lingerie hardware does to fabric — suspender
clips that grip the stocking welt, boning that shapes a panel, rings that
straps are genuinely threaded through. Measured blind against the same stack
without it: 58% of clip junctions correct vs 38%, blind-judged at native zoom.
A modest, measured improvement — not a total fix. Trigger lngcon, strength
0.8, written prompts not tags.</code></pre>

Files with SHA256: a8c6242b486faa325aa875ccc3d97096f97247afac17262e7aa948f222938eb9