ReservedNoName/sdxl-simple-icons-lora

Simple Icons LoRA for SDXL

A LoRA model trained on Simple Icons dataset for generating minimalist tech logos.

🔗 Links

• Hugging Face Model: https://huggingface.co/ReservedNoName/sdxl-simple-icons-lora
• GitHub Repository (Training code & tools): https://github.com/Octane0411/lora-playground
• Training Notebook: colab_training.ipynb

Model Details

• Base Model: Stable Diffusion XL 1.0
• Training Data: 187 Simple Icons logos
• Training Steps: 120 steps (5 epochs)
• Resolution: 384x384
• Rank: 8
• Learning Rate: 1e-4

Usage

from diffusers import DiffusionPipeline
import torch

# Load base model
pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16
)
pipe = pipe.to("cuda")

# Load LoRA weights
pipe.load_lora_weights("ReservedNoName/sdxl-simple-icons-lora")

# Generate image
prompt = "minimalist tech logo of React, geometric shape, flat design, single color"
image = pipe(prompt, num_inference_steps=50, guidance_scale=12.0).images[0]
image.save("output.png")

Trigger Words

Use these keywords for best results:

• minimalist tech logo
• geometric shape
• flat design
• single color

Example Prompts

• minimalist tech logo of Python, geometric shape, flat design, single color
• minimalist tech logo of GitHub, geometric shape, flat design, single color
• minimalist tech logo of Docker, geometric shape, flat design, single color

Limitations

This is an early version (v1) with limited training data. Results may vary. A v2 with improved parameters is in development.

Training Details

Trained using Hugging Face Diffusers library with the following parameters:

• Mixed Precision: BF16
• Optimizer: 8-bit Adam
• Gradient Checkpointing: Enabled
• Scheduler: Constant
• Training Time: ~32 minutes (120 steps)
• Final Loss: 5.11e-5

For complete training configuration and code, see the GitHub repository.

Known Issues

This is v1 with limited training data (187 samples). The model may:

• Generate overly simple/abstract shapes
• Have limited style variation
• Work best with tech-related prompts

A v2 with improved parameters is in development.

Development

Want to improve this model or train your own? Check out:

• Training notebook
• Data collection scripts
• Upload guide

License

This model is released under the same license as SDXL 1.0.