ClayCanvas - Artistic Style LoRA for Stable Diffusion 1.5

A powerful LoRA (Low-Rank Adaptation) fine-tune of Stable Diffusion 1.5 trained on 12,554 real artworks from the OpenBrush dataset. This model learns authentic artistic styles including oil painting, watercolor, line art, pencil drawing, charcoal, pastel, and more.

Model Details

Component	Details
Base Model	Stable Diffusion 1.5 (runwayml/stable-diffusion-v1-5)
Base Parameters	1.07 Billion (UNet: 860M + Text Encoder: 123M + VAE: 83M)
LoRA Trainable Parameters	6.4 Million (0.7% of base)
LoRA Rank	32
LoRA Alpha	32
Target Modules	to_q, to_k, to_v, to_out.0
Training Steps	10,000
Dataset	OpenBrush (12,554 real artworks, 30GB)
Resolution	256x256
Framework	Diffusers + PEFT + Accelerate
Date	September 2026

Training Configuration

Parameter	Value
Optimizer	AdamW
Learning Rate	1e-4
Scheduler	Cosine
Warmup Steps	50
Batch Size	1
Gradient Accumulation	1
Mixed Precision	None (CPU)
Loss Function	MSE

Styles Learned

The model was trained on diverse artistic styles:

• Oil Painting - Rich textures, brush strokes, vibrant colors
• Watercolor - Soft washes, flowing colors, transparency
• Pencil Drawing - Detailed sketches, shading, cross-hatching
• Line Art - Clean outlines, minimal, black and white
• Crayon - Waxy texture, vibrant colors, childlike
• Charcoal - Dramatic lighting, smudged textures
• Pastel - Soft colors, powdery texture
• Ink - Bold lines, high contrast
• Impressionism - Light effects, visible brushstrokes

Usage

Option 1: Python with Diffusers

from diffusers import StableDiffusionPipeline
import torch

# Load base model
pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    torch_dtype=torch.float16
)
pipe.to("cuda")

# Load ClayCanvas LoRA
pipe.load_lora_weights("satu1234/ClayCanvas")

# Generate image
image = pipe(
    "a tiger in a forest, oil painting",
    num_inference_steps=30,
    guidance_scale=7.5
).images[0]
image.save("output.png")

Option 2: CPU (Slower)

from diffusers import StableDiffusionPipeline
import torch

pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    torch_dtype=torch.float32
)
pipe.load_lora_weights("satu1234/ClayCanvas")
pipe.to("cpu")

image = pipe(
    "a Chinese city with pagoda temples, watercolor painting",
    num_inference_steps=100,
    guidance_scale=7.5
).images[0]
image.save("chinese_city.png")

Option 3: Automatic1111 WebUI / Forge

1. Download lora_weights.safetensors from this repo
2. Place in models/LoRA/ folder
3. Use in prompt: <lora:ClayCanvas:1.0>
4. Add style keywords to your prompt

Option 4: ComfyUI

1. Download lora_weights.safetensors
2. Place in models/loras/ folder
3. Add "Load LoRA" node with strength 0.8-1.0

Prompt Guide

Style Keywords

Add these to your prompt for different styles:

Style	Prompt Keywords
Oil Painting	oil painting, oil on canvas, rich textures, brush strokes
Watercolor	watercolor, watercolor painting, soft washes, flowing colors
Pencil Drawing	pencil drawing, sketch, detailed shading, graphite
Line Art	line art, clean lines, outline, minimal
Crayon	crayon drawing, waxy texture, vibrant colors
Charcoal	charcoal drawing, dramatic lighting, smudged
Pastel	pastel colors, soft texture, powdery
Ink	ink drawing, bold lines, high contrast

Example Prompts

# Oil Painting Landscape
"a mountain landscape at sunset, oil painting, rich textures, vibrant colors, masterpiece"

# Watercolor Portrait
"a woman's face in profile, watercolor painting, soft colors, flowing, artistic"

# Line Art Architecture
"a modern city skyline, line art, clean lines, minimal, black and white"

# Mixed Style
"a forest scene, watercolor background, oil painting on trees, line art details"

# Chinese City
"a beautiful Chinese city with traditional pagoda temples, cherry blossoms, oil painting style, detailed, masterpiece"

# Realistic Scenery with Artistic Touch
"a realistic mountain landscape, watercolor sky, oil painting foreground, line art details"

Recommended Settings

Parameter	Recommended Value
Inference Steps	30-100 (higher = better quality)
Guidance Scale	7.0-8.5 (higher = more prompt adherence)
Resolution	512x512 (or 768x768 with high VRAM)
Sampler	Euler a, DPM++ 2M Karras
LoRA Strength	0.7-1.0

Training Progress

• Dataset: 12,554 real artworks from OpenBrush
• Training Steps: 10,000
• Current Status: Training in progress
• Final Loss: ~0.18

How It Works

1. Base Model: Stable Diffusion 1.5 (1.07B parameters) generates images from text
2. LoRA Adapter: Our trained weights (6.4M parameters) modify the base model to apply artistic styles
3. Result: You get artistic images without retraining the full model

Comparison

Method	Parameters	Storage	Training Time
Full Fine-tune	1.07B	2.5GB	Days
LoRA (this model)	6.4M	25MB	Hours

License

CC0 (Public Domain) - Based on OpenBrush dataset which is CC0 licensed.

Citation

@misc{claycanvas2026,
  title={ClayCanvas: Artistic Style LoRA for Stable Diffusion 1.5},
  author={Satyam},
  year={2026},
  howpublished={\url{https://huggingface.co/satu1234/ClayCanvas}},
  note={Trained on OpenBrush dataset with 12,554 artworks}
}

Acknowledgments

• OpenBrush Dataset - 75K+ public domain artworks
• Stable Diffusion 1.5 by RunwayML
• Diffusers library
• PEFT for LoRA implementation

Links

• Model Page: https://huggingface.co/satu1234/ClayCanvas
• Base Model: https://huggingface.co/runwayml/stable-diffusion-v1-5
• Dataset: https://huggingface.co/datasets/jaddai/openbrush
• Documentation: https://huggingface.co/docs/diffusers

---

Made with ❤️ by Satyam