##### V2 is here! #####
Changes in v2:
Larger dataset
better training resolution
buckets!
better cations
new trigger words for better separation (see below)
v2 has a new trigger system. I wanted to separate the elegant and classic strappy lingerie from the harder, leather and BDSM-inspired lingerie.
Global trigger (still the same): strappy_lingerie
elegant_strappy_lingerie creates refined strappy lingerie with satin elastic straps, sheer mesh, lace, cage bras, underwire strap bras, high-waisted strap panties, garter straps, thigh straps, small O-ring connectors, buckle sliders, chokers and body-framing geometric strap layouts.
fetish_strappy_lingerie creates harder fetish-inspired harness lingerie with leather, patent leather, glossy straps, wide harness straps, heavy buckle hardware, collars, wrist cuffs, thigh cuffs, chains, rivets, large O-ring connectors, open-cup harness bras and industrial body-framing constructions.
Works best with clear material prompts such as satin elastic, matte elastic, sheer mesh, lace, leather or patent leather. Strong for classic lingerie, thigh straps, garter constructions, chokers/collars, ring hardware and upper-body detail shots. Full-body wrap straps, foot harnesses and hand harness gloves can work but are not fully reliable and should be prompted very explicitly.
Dataset and training data:
172 images
5 steps per image
10 epochs
bucks with 1152px max size, no auto scaling
cosine scheduler with 0.2 warmup and 0.7 decay
Captioning was done with OpenAI
alpha == dim: 32
##### V1 Description #####
This LoRA is trained on various strappy lingerie and bondage-inspired outfits, including leather and latex harness sets with metal rings, buckles, and chains, cage bras, corsets, and garter belts with attached thigh straps. It also covers mesh and lace bodysuits with strap patterns, halter and choker designs, cupless bras, and harnesses that wrap around the bust, waist, hips, and thighs. Materials include leather, latex, satin, lace, mesh, and elastic straps in colors such as black, red, white, and metallic gold. Accessories like chokers, cuffs, and chain details are common, and the style ranges from fetish-inspired harness lingerie to elegant strappy fashion sets.
The LoRA was trained on the default flux1-dev model.
Dataset and training data:
136 images
9 steps per image
10 epochs
cosine scheduler with 0.2 warmup and 0.8 decay
Captioning was done with joycap-batch
If you use it with other loras, for example a character lora, try to lower the weight - I had good results at around 0.3 - 0.5.
Flux knew the concept of straps already pretty well, but the difference between a non-lora and a lora image is still huge, especially for the details like cuffs, chains and strap patterns.
A word of "warning": while the dataset was more or less SFW (A lot of skin and pasties), it is pretty easy to generate NSFW images with a NSFW model.
The showcase shows: 10 images with very specific prompts, 5 images with a relatively open and free prompt and 5 images with a NSFW model (I didn't put any effort into getting the nipples right because it's about the possibility of the outfits and not about perfect nipples).
Description
FAQ
Comments (3)
This LoRA doesn't really work because whatever software was used to train it made the CLIP-L text encoder (TE1) modifications in the LoRA too aggressive, causing the T5 encoder to allocate massive memory buffers. With SwarmUI on my 4090 I OOM using this LoRA. You can fix it by scaling the TE1 weights down by 30% with a simple python script. Also drops the size from 39 MB to 15 MB and seems to work quite well from what I can see.
Huh? You go OOM on a 4090? I am using this on a 3060 with 12GB RAM without any issues. Maybe it's something in connection with your workflow or SwarmUI, never tried that, I am only using ComfyUI. Maybe I will try SwarmUI to see what happens
I did a little research. I think what’s happening here is more related to how SwarmUI handles TE1 weights rather than an issue with the LoRA itself. Some UIs (like SwarmUI) tend to keep the full text encoder loaded in VRAM, which can cause memory spikes, while others (like ComfyUI) free it right after tokenization. That's why I don't have any OOM problems, so it seems to depend on the interface rather than the model. Still, your tip about scaling down TE1 weights is really helpful for broader compatibility — thanks for sharing it!