CivArchive
    Clear Single Strap Heels 透明一字带高跟鞋 - KR2
    NSFW
    Preview 149406449
    Preview 149406432
    Preview 149406431
    Preview 149406439
    Preview 149405550
    Preview 149405556
    Preview 149405557
    Preview 149406438
    Preview 149406445
    Preview 149406435
    Preview 149406448
    Preview 149406433
    Preview 149406441
    Preview 149406442
    Preview 149406440
    Preview 149406436
    Preview 149406452
    Preview 149405532
    Preview 149406434
    Preview 149406451


    𝗞𝗿𝗲𝗮𝟮 | 𝗠𝗢𝗦𝗦 | 𝗔𝗻𝗶𝗺𝗮:

    • Steps: 3,828 | 1,912 | 2,868

    • Effective BS: 3 | 6 | 6

    • Samples Seen: 11,484 | 11,472 | 17,208

    • LR: 3e-04 | 2.5e-04 | 3e-05 (scaled for BS; trained in OneTrainer with Alpha=1, roughly equivalent to ~1e-04 @ BS 1 in standard Alpha=Rank training tools)

    • Dataset: 1,098 images (see Krea 2 page for download) —— weighted repeats, AI-upscaled & raw photo mix, not a 20-image, inbred, overfit AI-slop batch of garbage.

    As a general community consensus, 𝗞𝗿𝗲𝗮 𝟮 offers noticeably more stable human and foot anatomy out of the box, whereas 𝗠𝗢𝗦𝗦 𝗜𝗺𝗮𝗴𝗲 trades some structural rigidity for superior prompt freedom and compositional flexibility powered by its larger Qwen3.5-9B text encoder.

    Since both Krea 2 and MOSS already have a solid baseline understanding of generic clear single-strap heels, the goal of this training wasn't to force a brand-new visual concept from scratch, but rather to reinforce and lock in a very specific real-world silhouette commonly seen in street-snap photography and auto-show fancams. Specifically, this LoRA sharpens the signature rear-foot geometry: the vertical heel spine loop, the transparent PVC heel counter, and the wrap-around single ankle strap. To achieve this without sacrificing realism, the dataset mixes AI-upscaled images (where I spent extensive time manually inpainting the rear heel details while keeping the rest of the frame untouched) with hand-picked, unedited raw captures featuring naturally sharp rear heel counters to counterbalance any artifacts introduced by AI upscaling.

    Note on Anima: My hybrid captioning strategy trigger + prefix tags + VLM-generated natural language prose worked here. Because Anima is heavily tuned on Danbooru-style tags, the prefix tags help anchor the shoe features reliably. However, since Anima relies on a compact 0.6B text encoder (Qwen3-0.6B)—and is primarily an anime/illustration-oriented backbone—its ability to parse dense, fine-grained photorealistic prose is naturally more limited compared to larger encoders. Overall, the results on Anima are just ok.


    Description

    FAQ

    LORA
    Krea 2

    Details

    Downloads
    154
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/18/2026
    Updated
    11/20/2026
    Deleted
    -
    Trigger Words:
    css_1, clear single-strap heeled sandals, transparent single front strap, clear PVC heel counter, vertical black leather heel spine, ankle strap passing through heel loop,

    Files

    dataset_ClearSingleStrapSandals_KR2.zip

    Size:
    294.46 MB
    SHA256:

    Mirrors