# v3nu5-cop - SD3.5 Large LoRA Models

## Model Description

This repository contains four LoRA (Low-Rank Adaptation) models trained on Stable Diffusion 3.5 Large, designed for generating images with specific styling and characteristics. Each model variant explores different caption preprocessing and optimization approaches.

## Model Variants

| Model | Caption Type | Optimizer | File |
|-------|-------------|-----------|------|
| v0b | PREFIX | PRODIGY | `sd35L-v3nu5-cop-v0b.safetensors` |
| v0c | PREFIX | ADAMW8BIT | `sd35L-v3nu5-cop-v0c.safetensors` |
| v0d | CONTEXT | PRODIGY | `sd35L-v3nu5-cop-v0d.safetensors` |
| v0f | CONTEXT | ADAMW8BIT | `sd35L-v3nu5-cop-v0f.safetensors` |

### Training Details

- **Base Model**: Stable Diffusion 3.5 Large
- **Training Dataset**: [mushroomfleet/venus-cop](https://huggingface.co/datasets/mushroomfleet/venus-cop)
- **Model Type**: LoRA (Low-Rank Adaptation)
- **Model Size**: ~147 MB per variant

### Caption Types
- **PREFIX**: Captions are structured with key descriptors at the beginning
- **CONTEXT**: Captions provide contextual scene descriptions

### Optimizers
- **PRODIGY**: Advanced adaptive learning rate optimizer
- **ADAMW8BIT**: Memory-efficient 8-bit AdamW optimizer

## Usage

### ComfyUI
1. Download the desired `.safetensors` file
2. Place it in your `ComfyUI/models/loras/` directory
3. Load the LoRA in your workflow with appropriate strength (recommended: 0.6-1.0)

### Automatic1111/Forge
1. Download the desired `.safetensors` file
2. Place it in your `stable-diffusion-webui/models/Lora/` directory
3. Use in prompts with: `<lora:sd35L-v3nu5-cop-v0X:0.8>` (replace X with variant)

### Diffusers
```python
from diffusers import StableDiffusion3Pipeline
import torch

# Load base model
pipe = StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3.5-large")

# Load LoRA weights
pipe.load_lora_weights("mushroomfleet/v3nu5-cop", weight_name="sd35L-v3nu5-cop-v0b.safetensors")

# Generate image
prompt = "your prompt here"
image = pipe(prompt, num_inference_steps=28, guidance_scale=4.5).images[0]
```

## Recommended Settings

- **Strength**: 0.6-1.0
- **CFG Scale**: 4.5-7.0
- **Steps**: 28-35
- **Sampler**: DPM++ 2M or Euler

## Model Comparison

Each variant offers different characteristics:
- **v0b (PREFIX + PRODIGY)**: Strong adherence to structured prompting
- **v0c (PREFIX + ADAMW8BIT)**: Balanced approach with prefix structure
- **v0d (CONTEXT + PRODIGY)**: Natural scene understanding
- **v0f (CONTEXT + ADAMW8BIT)**: Contextual generation with efficiency

## License

Please refer to the base Stable Diffusion 3.5 Large license terms.

## Training Dataset

These models were trained using the [venus-cop dataset](https://huggingface.co/datasets/mushroomfleet/venus-cop). Please refer to the dataset page for more information about the training data.

## Citation

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

```bibtex
@misc{v3nu5-cop-lora,
  title={v3nu5-cop: SD3.5 Large LoRA Models},
  author={mushroomfleet},
  year={2025},
  howpublished={\url{https://huggingface.co/mushroomfleet/v3nu5-cop}}
}
```

## Contact

For questions or issues, please open an issue in this repository or contact through Hugging Face.