A LoRA adapter for SDXL that generates realistic handwritten signatures from a name prompt. Trained on a cleaned/processed signature dataset with text prompts; intended for research & non-commercial use.

Usage (Diffusers)

python

import torch
from diffusers import DiffusionPipeline, EulerAncestralDiscreteScheduler

base = "stabilityai/stable-diffusion-xl-base-1.0"
lora = "sl4shuur/sdxl-handwriting-signaturegen-lora"   # this repo id

pipe = DiffusionPipeline.from_pretrained(base, torch_dtype=torch.float16, use_safetensors=True).to("cuda")
pipe.load_lora_weights(lora)
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)

negative = "extra words, extra text, blurry, artifacts, extra lines, scribble, low quality, gray/black background, double signature"
name = "B. Navya"
prompt = f"signaturegen style: handwritten name {name} on white background"

image = pipe(
    prompt=prompt,
    negative_prompt=negative,
    num_inference_steps=50,
    guidance_scale=7.0,
    height=512, width=512,
    generator=torch.Generator(device='cuda').manual_seed(42)
).images[0]

image.save("sample.png")

Weights for this model are available in Safetensors format.

Prompting tips

Use a clear name inside the prompt.
Keep it single-line vs multi-line if that’s what you trained on.
Recommended: 50 steps, guidance 6.5–8.0, seed fixed for reproducibility.

Trigger words

You should use signaturegen style: to trigger the image generation.

Files

pytorch_lora_weights.safetensors — LoRA adapter weights.

Training & Data

Base model: stabilityai/stable-diffusion-xl-base-1.0
Dataset: sl4shuur/handwriting-signatures-dataset
Preprocessing: cleaning, cropping, centering, resize to 256×256; manual curation.

Training script

!accelerate launch /content/diffusers/examples/text_to_image/train_text_to_image_lora_sdxl.py \
  --pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" \
  --train_data_dir="sl4shuur/handwriting-signatures-dataset" \
  --caption_column="text" \
  --resolution=256 \
  --mixed_precision="fp16" \
  --allow_tf32 \
  --gradient_checkpointing \
  --use_8bit_adam \
  --enable_xformers_memory_efficient_attention \
  --dataloader_num_workers=4 \
  --learning_rate=1e-4 --lr_scheduler="cosine" --lr_warmup_steps=0 \
  --rank=4 \
  --max_train_steps=2000 \
  --checkpointing_steps=500 \
  --train_batch_size=2 --gradient_accumulation_steps=4 \
  --seed=69 \
  --output_dir="/content/drive/Shareddrives/SignatureGenDataset/lora_signature_model_OUT" \
  --hub_model_id="sl4shuur/sdxl-handwriting-signaturegen-lora" \
  --push_to_hub

Intended use & limitations

Research / non-commercial.
Generates stylized signatures from names; not for identity verification or impersonation.
Can fail on very long names or non-Latin scripts (out-of-distribution).
See original Stable Diffusion XL Base 1.0 page for additional limitations and biases.

License

CC-BY-NC-4.0