# Reynolds Oil Style LoRA Models

A collection of LoRA (Low-Rank Adaptation) models trained to generate images in the distinctive Reynolds oil painting style using Stable Diffusion 3.5 Large.

## Models Overview

This repository contains four different model variants, each trained with different captioning strategies and optimizers:

| Model Version | Captioning Strategy | Optimizer | Description |
|---------------|-------------------|-----------|-------------|
| **v0b**  | PREFIX  | PRODIGY   | Optimized with Prodigy optimizer using prefix-style captions |
| **v0c**  | PREFIX  | ADAMW8BIT | Optimized with AdamW 8-bit using prefix-style captions |
| **v0d**  | CONTEXT | PRODIGY   | Optimized with Prodigy optimizer using contextual captions |
| **v0f**  | CONTEXT | ADAMW8BIT | Optimized with AdamW 8-bit using contextual captions |

## Dataset

All models were trained on the same high-quality dataset:
- **Dataset Source**: [mushroomfleet/reynolds-oil-style](https://huggingface.co/datasets/mushroomfleet/reynolds-oil-style)
- **Base Model**: Stable Diffusion 3.5 Large
- **Training Type**: LoRA (Low-Rank Adaptation)

## Usage

### Basic Usage with Diffusers

```python
from diffusers import StableDiffusion3Pipeline
import torch

# Load the base model
pipe = StableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3.5-large",
    torch_dtype=torch.float16
)

# Load the LoRA weights (replace with your chosen variant)
pipe.load_lora_weights("mushroomfleet/reynolds-oil-style", weight_name="reynolds-oil-style-b")

# Generate an image
prompt = "a majestic landscape in the style of Reynolds oil painting"
image = pipe(prompt, num_inference_steps=28, guidance_scale=7.5).images[0]
image.save("reynolds_style_output.png")
```

### ComfyUI Usage

1. Download the desired model files (.safetensors)
2. Place them in your `ComfyUI/models/loras/` directory
3. Use the "Load LoRA" node in your workflow
4. Set the strength between 0.7-1.2 for optimal results

## Model Differences

### Captioning Strategies

- **PREFIX captions**: Captions begin with style descriptors, helping the model associate the style trigger more directly
- **CONTEXT captions**: Captions integrate style information contextually within the description

### Optimizers

- **PRODIGY**: An adaptive learning rate optimizer that can provide better convergence for some datasets
- **ADAMW8BIT**: Memory-efficient version of AdamW optimizer, good for stable training

## Recommended Settings

- **Strength**: 0.8-1.2
- **CFG Scale**: 7.0-9.0
- **Steps**: 25-35
- **Sampler**: DPM++ 2M Karras or Euler A

## Trigger Words

Use these terms in your prompts to activate the style:
- "R3YN0LDS oil painting style"


## Example Prompts

```
a portrait of a noble gentleman, Reynolds oil painting style, classical lighting, rich colors

a dramatic landscape with stormy skies, oil painting, academic style, detailed brushwork

elegant lady in 18th century dress, Reynolds style portrait, soft lighting, refined details
```

## Training Details

- **Base Model**: Stable Diffusion 3.5 Large
- **Training Resolution**: [Add resolution if known]
- **Batch Size**: [Add if known]
- **Learning Rate**: [Add if known]
- **Training Steps**: [Add if known]

## License

[Add appropriate license information]

## Credits

- **Dataset**: Created and curated by mushroomfleet
- **Training**: Conducted using advanced LoRA techniques on SD3.5 Large
- **Style**: Inspired by the classical oil painting techniques of Joshua Reynolds

## Citation

If you use these models in your research or creative work, please consider citing:

```
@misc{reynolds-oil-style-lora,
  title={Reynolds Oil Style LoRA Models for Stable Diffusion 3.5},
  author={mushroomfleet},
  year={2025},
  url={https://huggingface.co/mushroomfleet/reynolds-oil-style}
}
```

---

*For questions, issues, or feedback, please open an issue in this repository or contact the author.*