dhatchin/divine-studio-lora

Divine Studio LoRA

A custom LoRA (Low-Rank Adaptation) trained on Stable Diffusion v1.5 for the Divine Studio project.

This repository contains the trained LoRA adapter weights that can be loaded on top of a compatible Stable Diffusion v1.5 base model.

Model Details

Property	Details
Model type	LoRA adapter
Base model	Stable Diffusion v1.5
Training method	LoRA
Training resolution	512 × 512
Training steps	400
Training batch size	1
Gradient accumulation steps	1
Learning rate	1e-4
Learning rate scheduler	Constant
Warmup steps	0
Random seed	0
Mixed precision	FP16
Trigger word	divine_style
Weight format	Safetensors
Repository	dhatchin/divine-studio-lora

Base Model

This LoRA was trained using:

runwayml/stable-diffusion-v1-5

The LoRA is an adapter and does not contain the complete Stable Diffusion base model.

To use the adapter, load it together with a compatible Stable Diffusion v1.5 base model.

Please review the base model's license and usage terms before using the LoRA.

LoRA Weights

The trained adapter is provided as:

pytorch_lora_weights.safetensors

The weights are stored in Safetensors format and are intended to be loaded with the compatible Stable Diffusion v1.5 base model.

Trigger Word

The trigger word used during training was:

divine_style

The trigger word can be included in prompts when using the LoRA.

Using the trigger word helps indicate the learned concept associated with the training process.

Usage with Diffusers

The LoRA can be loaded using the Diffusers library.

Installation

Install the required packages:

'pip install torch pip install diffusers pip install transformers pip install accelerate pip install safetensors'

Load the LoRA 'import torch from diffusers import StableDiffusionPipeline base_model = "runwayml/stable-diffusion-v1-5" pipe = StableDiffusionPipeline.from_pretrained( base_model, torch_dtype=torch.float16 ).to("cuda") pipe.load_lora_weights( "dhatchin/divine-studio-lora", weight_name="pytorch_lora_weights.safetensors" ) prompt = "divine_style, a beautiful traditional Indian fashion design" image = pipe(prompt).images[0] image.save("divine_studio_output.png")'

Example A basic example prompt using the trained trigger word:

'divine_style, a beautiful traditional Indian fashion design'

You can extend the prompt with additional descriptions depending on the desired output.

For example:

'divine_style, elegant traditional Indian fashion, detailed fabric, professional studio lighting, high quality'

The divine_style trigger word was used as the instance prompt during training.

Training Details

The LoRA was trained using the Diffusers DreamBooth LoRA training script:

'diffusers/examples/dreambooth/train_dreambooth_lora.py'

The training command used the following configuration:

'Base Model: runwayml/stable-diffusion-v1-5 Training Images: 20 Training Resolution: 512 × 512 Train Batch Size: 1 Gradient Accumulation Steps: 1 Checkpointing Steps: 100 Learning Rate: 1e-4 Learning Rate Scheduler: constant Learning Rate Warmup Steps: 0 Maximum Training Steps: 400 Random Seed: 0 Mixed Precision: fp16'

The training completed for 400 steps.

The final training output was saved as:

'pytorch_lora_weights.safetensors'

Training Dataset

The training images were stored in Google Drive during training:

'/content/drive/MyDrive/divine_training_images'

The notebook detected 20 training images in the directory.

The training dataset itself is not included in this repository.

The dataset consisted of design images used to train the Divine Studio LoRA.

Users should only use training images for which they have the appropriate rights and permissions.

Training Process

The overall training workflow was:

1.Mount Google Drive. 2.Load the training images from the configured training directory. 3.Verify the available training images. 4.Install the required Diffusers and training dependencies. 5.Authenticate with Hugging Face. 6.Load Stable Diffusion v1.5 as the pretrained base model. 7.Train a LoRA adapter using the Diffusers DreamBooth LoRA training script. 8.Use divine_style as the instance prompt. 9.Train at a resolution of 512 × 512. 10.Train for a maximum of 400 steps. 11.Save the trained adapter as a Safetensors file. 12.Verify the generated LoRA weights. 13.Upload the trained weights to the Hugging Face Hub. 14.Verify that the uploaded model can be downloaded from the repository.

Training Output

The final LoRA file generated by the training process was:

'/content/divine-studio-lora/pytorch_lora_weights.safetensors'

The file was then uploaded to the Hugging Face repository as:

'pytorch_lora_weights.safetensors'

Evaluation

No formal quantitative evaluation metrics were calculated in the training notebook.

The notebook verifies the existence of the trained LoRA weights and performs repository/download verification after uploading the model.

Therefore, this repository does not report numerical metrics such as:

1.Accuracy 2.FID 3.CLIP Score 4.Precision 5.Recall 6.F1 Score

No such metrics should be interpreted as having been measured for this model.

Repository Structure

The Hugging Face repository contains:

divine-studio-lora/ ├── .gitattributes ├── README.md └── pytorch_lora_weights.safetensors

pytorch_lora_weights.safetensors - Contains the trained LoRA adapter weights.

README.md - Contains the model card, model information, training configuration, usage instructions, limitations, and citation information.

.gitattributes - Repository configuration used by the Hugging Face Hub for file handling.

Requirements

The example inference code requires Python packages including:

'torch diffusers transformers accelerate safetensors'

Install them with:

'pip install torch diffusers transformers accelerate safetensors'

A CUDA-compatible GPU is recommended for practical FP16 image generation.

Loading the LoRA

The LoRA adapter can be loaded directly from the Hugging Face Hub:

'dhatchin/divine-studio-lora'

The adapter file is:

'pytorch_lora_weights.safetensors'

Example:

'pipe.load_lora_weights( "dhatchin/divine-studio-lora", weight_name="pytorch_lora_weights.safetensors" )'

Limitations

1.This repository contains a LoRA adapter, not a complete Stable Diffusion model. 2.The adapter requires a compatible Stable Diffusion v1.5 base model. 3.The LoRA was trained using a relatively small dataset of 20 images. 4.The learned visual characteristics are dependent on the training dataset. 5.Generated results can vary depending on the prompt and inference configuration. 6.The training notebook does not provide formal quantitative evaluation metrics. 7.The original training dataset is not included in this repository. 8.The divine_style trigger word was used during training and is recommended when attempting to reproduce the learned concept. 9.Results may differ when using different inference settings, model versions, or LoRA weights.

License

The metadata for this repository specifies:

'CreativeML OpenRAIL-M'

The underlying Stable Diffusion v1.5 base model has its own license and usage terms.

Users should review the applicable license and usage conditions of the base model before using the LoRA.

The training dataset may also have separate copyright, licensing, or usage restrictions.

Before redistributing or commercially using the model or generated content, users should ensure that their intended use complies with the applicable licenses and rights associated with the base model, LoRA, training data, and generated content.

Citation

If you use the Divine Studio LoRA in a project, you may cite it as:

'@misc{divine_studio_lora, author = {S S Dhatchin}, title = {Divine Studio LoRA}, year = {2026}, publisher = {Hugging Face}, howpublished = {https://huggingface.co/dhatchin/divine-studio-lora} }'

Stable Diffusion Reference

The underlying Stable Diffusion architecture is based on latent diffusion models.

'@InProceedings{Rombach_2022_CVPR, author = {Rombach, Robin and Blattmann, Andreas, Lorenz, Dominik and Esser, Patrick and Ommer, Björn}, title = {High-Resolution Image Synthesis With Latent Diffusion Models}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, month = {June}, year = {2022}, pages = {10684--10695} }'

Hugging Face Repository

The trained LoRA is available on the Hugging Face Hub:

'https://huggingface.co/dhatchin/divine-studio-lora'

Repository ID:

dhatchin/divine-studio-lora