Train your own LORA model in 30 minutes LIKE A PRO!

Code & bird
9 Oct 202330:12

TLDRDiscover how to train your own LoRA model efficiently in this tutorial video. LoRA, standing for Low Rank Adaptation, enhances stable diffusion models, making it easier to generate consistent images with specific poses or styles. The tutorial covers preparing a dataset with around 15-35 images, setting up training using a specific notebook, and executing training on Google Colab. After training, you can export and share your LoRA model. The process requires minimal data and effort, allowing you to customize and improve image generation for specific concepts like character poses or artwork styles.


  • ๐Ÿš€ Train your own LORA (Low Rank Adaptation) model to fine-tune Stable Diffusion checkpoints for generating images with consistent characters, poses, or objects.
  • ๐ŸŽจ LORA technology helps in training Stable Diffusion on specific concepts like characters, poses, objects, and different artwork styles.
  • ๐Ÿ’ก After training, you can export and reuse or share your LORA model with others, contributing to a community like Civit AI.
  • ๐Ÿ“ธ Prepare a dataset of 15 to 35 varied pictures of your subject for LORA training, ensuring they are in different stances, poses, and conditions.
  • ๐Ÿ–ผ๏ธ Crop images to a square size, e.g., 512x512 pixels, and describe each picture with a specific tag and caption for the LORA model.
  • ๐Ÿ“‚ Organize your dataset in a specific directory structure, with repetitions and model names corresponding to your LORA model.
  • ๐Ÿ“š Find and use a suitable notebook for LORA training, like the one from user l q Ruf, and save a copy in your Google Drive for stability.
  • ๐Ÿ”ง Install necessary Python dependencies and connect your Google Drive to the notebook for saving your LORA model.
  • ๐Ÿ”„ Download the base Stable Diffusion model and optional VAE, and configure the local train directory to access your Google Drive.
  • ๐Ÿ› ๏ธ Configure the model, dataset, and LORA settings in the notebook, including custom tags, network parameters, and optimizer configurations.
  • ๐Ÿƒ Execute the training process, monitor the output, and ensure the paths are correctly configured to avoid errors.

Q & A

  • What is LORA and what problem does it solve in image generation?

    -LORA stands for Low Rank Adaptation. It addresses the challenge of generating images with consistent character poses or objects in Stable Diffusion, which can be difficult. LORA fine-tunes Stable Diffusion checkpoints to make training on specific concepts like character poses and artwork styles easier and more effective.

  • How does LORA simplify the process of training Stable Diffusion models?

    -LORA simplifies the training process by utilizing low-rank adaptation technology, which requires fewer images and less effort compared to other models. This makes it more accessible for users to create and fine-tune their own models with specific features or styles.

  • What are the requirements for preparing a dataset to train a LORA model?

    -To train a LORA model, you need 15 to 35 pictures of the subject in various poses and conditions. These images should be diverse and not repetitive. Each image should be cropped to a square size and accompanied by a text description to help the model learn from the dataset effectively.

  • How can one start training a LORA model after preparing the dataset?

    -After preparing the dataset, the next steps involve finding a suitable training notebook, like the one provided by the user 'l q Ruf', saving a copy of it, and running the necessary cells with the configured settings specific to your LORA model in an environment like Google Colab.

  • Why is it recommended to save a copy of the training notebook in your own Google Drive?

    -Saving a copy of the training notebook in your Google Drive ensures you have a stable working version. The original notebook could be updated or modified, potentially breaking functionality or changing configurations that might be crucial for your specific training setup.

  • What are the necessary steps involved in the actual training process of a LORA model?

    -The training process involves setting up the model configuration, defining the training dataset path, selecting the model and VAE to use, configuring the dataset specifics like image resolution and repetition, and finally, running the training cells in the notebook to start the model training.

  • What should you do if you encounter errors during the training process?

    -If errors occur during training, it's important to double-check all configurations and file paths. Errors often arise from misconfigurations or incorrect paths to datasets, models, or other necessary files.

  • What happens after the LORA training is completed?

    -Once LORA training is completed, the trained model files are saved in a specified output directory. These files can then be downloaded, and the model can be deployed or shared for generating images with the trained characteristics.

  • How can the trained LORA model be integrated into Stable Diffusion for generating images?

    -The trained LORA model can be uploaded into a Stable Diffusion environment, such as a web UI, where it can be selected and used to generate images by setting the model as the active one and adjusting parameters like the strength of the model's influence on the output.

  • What is the benefit of generating images using LORA in different artistic styles?

    -Using LORA to generate images in various artistic styles allows for greater creative flexibility and customization. It enables users to create unique and consistent visuals tailored to specific themes or styles, enhancing the overall aesthetic and thematic coherence of the generated images.



๐Ÿค– Introduction to Laura and Stable Diffusion Training

The video begins with an introduction to Laura, a model that utilizes low rank adaptation technology to fine-tune stable diffusion checkpoints, making it easier to generate images with consistent character poses, objects, or styles. The creator explains the process of training Laura with a dataset of images, highlighting its lower effort and resource requirements compared to other models. The video will demonstrate how to train Laura using a dataset of pictures of a parrot named Drari.


๐Ÿ“š Preparing the Data and Selecting the Right Notebook

The second paragraph focuses on the preparation of the data set and selecting the appropriate notebook for training Laura. The creator describes the need for 15 to 35 varied pictures of the subject and explains the process of cropping and naming the images. They also discuss the importance of saving the notebook to Google Drive to ensure a stable and working version. The creator provides a link to the notebook used for training and emphasizes the need to run specific cells of the notebook with the correct configuration for the Laura model.


๐Ÿ› ๏ธ Training the Laura Model and Uploading Data

This paragraph delves into the actual training process of the Laura model. The creator outlines the steps to configure the model, including setting up the project name, model path, and data set configurations. They explain the process of uploading the prepared data set to Google Drive and adjusting the settings for the training, such as the custom tag, network category, and optimizer configuration. The paragraph concludes with the start of the training process and a brief mention of the expected output.


๐ŸŽจ Testing the Trained Laura Model

The fourth paragraph describes the process of testing the trained Laura model. The creator explains how to upload the trained Laura model to the web UI and how to use it in conjunction with the stable diffusion web UI. They demonstrate the model's effectiveness by generating images using various prompts and adjusting the weight of the Laura model to achieve different results. The creator also explores the use of different samplers and their impact on the final images, showcasing the flexibility and potential of the trained Laura model.


๐Ÿ”„ Experimenting with Different Models and Styles

In this paragraph, the creator experiments with different models and styles to see how the trained Laura model can be applied. They test the model with the dream shaper and other models, noting the varying results and potential for improvement. The creator also discusses the use of the image to image feature to enhance and refine the generated images, demonstrating the iterative process of achieving a desired outcome. The paragraph highlights the potential for creativity and customization when using the trained Laura model with different styles and models.


๐Ÿ“ˆ Final Thoughts and Encouragement for Iterative Improvement

The final paragraph wraps up the video with a summary of the process and an encouragement for viewers to experiment and improve upon the trained Laura model. The creator shares their excitement about the results and potential of the model, urging viewers to share their tips and tricks for Laura training. They emphasize the value of iteration and refinement in achieving satisfying results, and invite feedback and engagement from the audience.




LORA stands for Low-Rank Adaptation, a technology used to fine-tune models like Stable Diffusion. In the context of the video, LORA is employed to train a model that can generate images with consistent character poses, objects, or styles. The process involves using a smaller amount of pictures and less effort compared to other models, making it accessible for individuals to create their own LORA models. The video demonstrates how to train a LORA model on the concept of a character, specifically a parrot named Drari, using various pictures and descriptions to achieve this goal.

๐Ÿ’กStable Diffusion

Stable Diffusion is a type of machine learning model used for generating images. It is mentioned in the video as the base model that the LORA technology is used to fine-tune. The script highlights a challenge with Stable Diffusion, which is generating images with consistent character poses or objects. The integration of LORA aims to address this issue by allowing users to train the model on specific concepts, such as a particular character or style, thereby improving the quality and consistency of the generated images.

๐Ÿ’กData Set

A data set, in the context of the video, refers to a collection of images and their corresponding descriptions used to train the LORA model. The data set must contain varied pictures of the subject in different poses, conditions, and backgrounds. Each image is also accompanied by a text description that tags the specific features or elements in the picture. The data set is crucial for the training process as it provides the model with the necessary information to learn and generate images that match the desired characteristics.

๐Ÿ’กCustom Tag

A custom tag, as used in the video, is a specific word or phrase that is associated with the LORA model being trained. This tag is used to trigger the model to generate images that correspond to the concept represented by the tag. For instance, in the video, 'Drari the Parrot' is the custom tag used to train the LORA model, and it serves as the activation word to generate images of Drari when used in conjunction with the Stable Diffusion model.

๐Ÿ’กGoogle Drive

Google Drive is a cloud storage service used in the video for saving and accessing files related to the LORA model training process. It is where the data set, the trained model, and the Jupyter notebook used for training are stored and shared. The use of Google Drive facilitates the process by allowing the user to save their work, access it from different devices, and share it with others if needed.

๐Ÿ’กJupyter Notebook

A Jupyter Notebook is an open-source web application that allows users to create and share documents containing live code, equations, visualizations, and narrative text. In the video, a Jupyter Notebook is used as the interface for training the LORA model. It contains the code and instructions necessary for the training process, and the user can run these instructions step by step to fine-tune the model according to their data set.


Training, in the context of the video, refers to the process of teaching the LORA model to recognize and generate images based on the data set provided. This involves fine-tuning the Stable Diffusion model with the help of the custom tag and the images in the data set. The training process is essential for creating a model that can generate consistent and accurate images of the specified subject.

๐Ÿ’กModel Export

Model export refers to the process of saving the trained LORA model in a format that can be reused or shared with others. After the training is complete, the model can be exported and uploaded to platforms like Civit AI for further use or distribution. This allows others to utilize the trained model without having to go through the entire training process themselves.

๐Ÿ’กImage Generation

Image generation is the outcome of the LORA model training process, where the model creates new images based on the learned characteristics from the data set. The generated images should reflect the subject, pose, or style that the model was trained on. The video demonstrates the use of the trained LORA model to generate images of Drari the Parrot, showcasing the effectiveness of the training and the potential for artistic creation.

๐Ÿ’กArtwork Styles

Artwork styles refer to the different visual aesthetics or artistic approaches that can be applied to the generated images. In the context of the video, the LORA model is trained not only on a specific subject (Drari the Parrot) but also on varying artwork styles to ensure that the generated images are diverse and visually appealing. The model's ability to adapt to different styles allows for a broader range of creative possibilities.


Learn how to train your own LORA model in a short span of 30 minutes.

LORA stands for Low Rank Adaptation, a technology used to fine-tune Stable Diffusion checkpoints.

Training LORA can help generate images with consistent character poses or objects in Stable Diffusion.

LORA models can be exported and reused or shared with others.

Creating your own LORA model requires a smaller amount of pictures and lower effort compared to other models.

The first step in LORA training is preparing a diverse dataset of 15 to 35 pictures of your subject.

Pictures should be in different stances, poses, and conditions, and should not be repetitive.

The training process involves using a specific notebook and following a series of steps in Google Colab.

Select a base model like Stable Diffusion for your LORA training.

Customize the training by setting specific tags and descriptions for each image.

Upload your prepared dataset to Google Drive and use it in the training notebook.

Configure the training settings such as the model name, path, and network parameters.

Monitor the training progress and logs to ensure correct configuration and identify any errors.

After training, upload your LORA model to a web UI and test it with different prompts and styles.

Experiment with different samplers and settings to achieve the desired image quality and style.

Use LORA models with various styles and models to create unique and interesting images.

Iterate and refine the generated images using image-to-image techniques for improved results.