adarrshDev/streetwear-v5-calibrated-master-lora

👕 Streetwear Decal SDXL LoRA (Release v5.0 - Calibrated Master Edition)

A calibrated, production-grade SDXL 1.0 LoRA fine-tuned on the Model v5.0 Calibrated Master Dataset (492 samples: 428 organic texture-rich decals + 64 weighted sovereign jewels) on pure #000000 pitch-black canvas.

🎯 Key Architectural Innovations

1. Calibrated Scaling (Rank 32, Alpha 16): By setting $\alpha/r = 0.5$ (Alpha 16, Rank 32), we prevent the contrast-burn and gradient flattening seen in over-fitted LoRAs, preserving SDXL Base's native textural depth and micro-details.
2. Optimal Epoch Budget (11.4 Epochs): 1,400 gradient accumulation steps across 492 images (5,600 image passes) perfectly strikes the balance between stylistic guidance and base model generalization, completely avoiding the 150-epoch overfit trap.
3. Rectangular Badge Purge: Every sample with an organic bounding-box fill ratio $> 0.82$ (full-bleed book covers, poster blocks, and woven rectangular patches) was surgically purged.
4. Physical Medium Caption Conditioning: Purged the flattening "sharp clean vector silhouette, studio product lighting" suffix. All captions describe physical mediums: archival woodblock grain, copperplate stippling, neo-traditional tattoo acrylic inks, and 3D high-density foam relief.

---

💻 Usage with Diffusers

import torch
from diffusers import StableDiffusionXLPipeline, DPMSolverMultistepScheduler

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    variant="fp16",
    use_safetensors=True
).to("cuda")

pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True)
pipe.load_lora_weights("adarrshDev/streetwear-v5-calibrated-master-lora", weight_name="sdxl_v5_calibrated_streetwear_decal_t4.safetensors")

prompt = "TOK_DECAL, TOK_BOTANICAL_ETCHING, vintage anatomical butterfly sketch, copperplate botanical etching, fine sepia line-art, biological specimen diagram with halo rings, isolated graphic decal on pure black background, zero text, zero garment, zero human"
negative_prompt = "clothing mockup, t-shirt mockup, photorealistic wrinkles, low quality, blurry, text watermark"

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=28,
    guidance_scale=6.0
).images[0]

image.save("botanical_butterfly_decal.png")