Jatin1999/patent-style-lora

Patent-Drawing Style LoRA (SDXL)

Turns any object into a black-and-white US patent figure: black lines on white, line-based shading, standard views.

Trigger word: ptnt_style. Put it first, then the view and the object:

ptnt_style, perspective view of a quadcopter drone

Recommended LoRA strength: 0.7. Negative prompt that helps: color, gray background, photo, photorealistic, 3D render, blurry, text, watermark

Usage

import torch
from diffusers import AutoPipelineForText2Image

pipe = AutoPipelineForText2Image.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
pipe.load_lora_weights("Jatin1999/patent-style-lora")
image = pipe("ptnt_style, side view of an electric scooter",
             negative_prompt="color, photo, 3D render, text").images[0]

Training

• Base: SDXL base 1.0, with the madebyollin/sdxl-vae-fp16-fix VAE
• Data: 24 drawing sheets from US design patents (via the DeepPatent dataset), one figure each, labels removed, padded to 1024×1024. Patent numbers are listed in manifest.csv in the project repo.
• LoRA rank 16 (alpha 16) on the UNet and both text encoders, LR 1e-4 constant, AdamW 8-bit, batch 1, bf16, noise offset 0.1
• Steps: 1,500 of a 3,000-step run (later checkpoints overfit)
• Hardware: one NVIDIA L4 (Google Colab)

Evaluation

10 subjects × 3 seeds per condition, fixed sampler, steps, CFG and negative prompt. Unseen gadgets (drone, smartwatch, VR headset, scooter, robot arm, foldable phone):

Condition	White border ↑	Gray share ↓	CLIP subject ↑
Prompt only ("patent drawing … line art")	0.440	0.241	0.301
LoRA @ 0.7	0.998	0.038	0.316

The LoRA's gain is the page: pure white, crisp black lines, no beige paper, with the subject preserved.

Limitations

• Generated figures are illustrations, not valid patent drawings, and carry no legal meaning.
• Text and reference numbers are not reliable; the training data had labels removed on purpose.
• Mechanical plausibility is not guaranteed: parts may not connect sensibly.

Data credits

US patent drawings are generally not subject to copyright (USPTO). Images were sourced through DeepPatent (Kucer et al., WACV 2022).