How to Use LoRA for Fine-Tuning Stable Diffusion

How to Use LoRA for Fine-Tuning Stable Diffusion

Unleash your creativity with LoRA Stable Diffusion models. Follow our blog to create your own LoRA training model.

Key Highlights

  • LoRA is a more efficient way to fine-tune the Stable Diffusion models.
  • LoRA models (1MB-6MB) are easier to share and distribute.
  • You can easily find many LoRA models on Civitai and Hugging Face.
  • Novita AI features APIs of LoRA training models for developers like you to create your own LoRA models.
  • LoRA is better than Dreambooth and textual inversion in fine-tuning Stable Diffusion.


The LoRA Stable Diffusion model is an innovative approach within the realm of AI-generated content, leveraging the concept of Low-Rank Adaptation to fine-tune existing Stable Diffusion models, allowing for the efficient customization of AI-generated images to specific styles, characters, or artistic preferences.

In this blog, we’ll give you a comprehensive introduction to LoRA models and teach you how to find and use LoRA to train Stable Diffusion models. Moreover, we’ll provide a customized way for you to create your own LoRA training models with Novita AI APIs. Let’s delve into the world of LoRA Stable Diffusion now!

Overview of LoRA (Low-Rank Adaption)

LoRA (Low-Rank Adaptation) is a training technique for fine-tuning Stable Diffusion models.

What is LoRA?

LoRA stands for “Low-Rank Adaptation,” which is a novel technique introduced by Microsoft researchers to deal with the problem of fine-tuning large-language models (LLMs). 

While, in the context of Stable Diffusion, it is a technique to fine-tune Stable Diffusion models efficiently. LoRA Stable Diffusion models are compact and allow for the application of subtle yet impactful modifications to standard checkpoint models, making them significantly smaller, usually 10 to 100 times smaller than the original models.

How does LoRA Work?

LoRA is a technique that makes Stable Diffusion models more adaptable without the need to retrain the whole model. Instead, it adds a few new weights to key parts of the model that handle how images and text interact. This change is made more efficient by using a method called rank decomposition, which simplifies the model’s weight structure. The result is a smaller, customized model that keeps its original abilities but can also focus on specific tasks. This model is then used alongside the original model checkpoint to create images that match particular styles or ideas.

Key Points of LoRA Models

Before using LoRA models, it’s necessary to learn about why you should use them and the different types of them.

Features of LoRA Models

  • Support for inpainting.
  • Out-of-the-box multi-vector pivotal tuning inversion.
  • Efficiency: LoRA requires less computational power and memory, making it more efficient for training large models.
  • Targeted Adjustments: It focuses on modifying ross-attention layers rather than the entire model, leading to more precise control over the model’s behavior.
  • Compatibility: LoRA models are designed to work alongside the original model checkpoints, ensuring compatibility and ease of use.
  • Ease of Sharing: The compact size of LoRA models (1MB-6MB) makes them easy to share and distribute.

Examples of LoRA Models

Based on the different functions of the LoRA models, we can divide them into the following types:

  • Character LoRA: Character LoRA models are fine-tuned to encapsulate the distinctive visual aspects, physique, and facial expressions of particular characters from cartoons, video games, and other entertainment platforms. They are instrumental in producing fan art, enhancing game development processes, and serving the needs of the animation and illustration industries.
  • Style LoRA: Style LoRA models are fine-tuned to emulate the artistic techniques or signature styles of certain artists, enabling the creation of images that reflect those specific aesthetics. They are frequently utilized to imbue a reference image with a unique visual flair, aligning it with a chosen artistic style.

Where to Find and Use LoRA Models

Download LoRA Model Files on Civitai or Hugging Face

Civitai is a site hosting a large collection of LoRA models, and Hugging Face is also a famous source of LoRA libraries. You can easily find many LoRA models on these two sites, searching them by visiting their website.

Use LoRA Model in AUTOMATIC1111 Stable Diffusion GUI

  • Install the LoRA Models: All you need to do is put the model files into the following folder.
  • Use the LoRA in the Prompt: Use the following syntax in the prompt or the negative prompt to add a LoRA with weight.
<lora: name: weight>

Train Your Own LoRA Model for Fine-Tuning with Novita AI

In Stable Diffusion, there are many models that are similar in architecture and pipeline but different in output. So, the LoRA model for fine-tuning Stable Diffusion is various, too. Moreover, LoRA isn’t a full model, it’s an add-on. That’s to say, you can create your own LoRA model, “teach” and train it any concepts continuously according to your needs.

If you want to create your own LoRA, you can integrate APIs into your own program. Novita AI is such a platform that offers developers like you both APIs to create a unique LoRA model and a playground to train your LoRA models. Since all training will be computationally intensive, be sure you have a machine with GPU to continue.

Find APIs About LoRA Training

  • Step 1: Visit the Novita AI website and create an account.
  • Step 2: Navigate to “API” and find the APIs of LoRA training under the “Training” tab. Novita AI provides both “LoRA for subject training” and “LoRA for style training”, which you can use to train both character LoRA and style LoRA.
  • Step 3: Start training your LoRA model.

Start LoRA Model Training

After getting the Novita AI API key, you can start to train your LoRA model. Let’s take “LoRA for subject training” as an example to show you how to train your model.

  • Step 1: Upload the images for model training. Get the image upload URL and put it into the following folder. Each task supports uploading up to 50 images. In order to have a good final effect, the images uploaded should meet some basic conditions, such as: “portrait in the center”, “no watermark”, “clear picture”, etc.
curl -X PUT -T “{{filepath}}” “{{upload_url}}”
  • Step 2: Configure parameters. There are four types of parameters for model training: Model info parameters, dataset parameters, components parameters, and expert parameters. According to your needs configure the parameters.
  • Step 3: Start Training Task. Novita AI provides a user-friendly training playground for you to train your model. Simply input the Novita AI API key and follow the steps in the following pic, then you can review the training result.
  • Step 4: Get training status. Input the command below to obtain the progress of model training and the status of model deployment after training. 
curl — location — request GET ‘training task_id’ \ — header ‘Authorization: Bearer {{Key}}’
  • Step 5: Start Using the Model: When the response shows “task_status: SUCCESS” and “model_status: SERVING”, you can start using the training model.

Other Methods of Training Stable Diffusion

Dreambooth and Textual Inversion

In addition to LoRA, Dreambooth, and textual inversion are other popular methods to fine-tune Stable Diffusion. However, compared to Dreambooth results in large model files (2–7 GBs), LoRA is much smaller and more efficient; compared to textual inversion tiny (about 100 KBs) results, LoRA can be used for general-purpose fine-tuning, adapting to new domains or datasets.

Training Tips

  • GPU Requirement: Training LoRA models is computationally demanding, so ensure you have a machine with a capable GPU to handle the workload.
  • Model Compatibility: LoRA models should be compatible with the base Stable Diffusion model they are intended to augment. Check compatibility to avoid integration issues.
  • Parameter Tuning: Fine-tune the parameters of the LoRA model carefully, as they can significantly affect the output, including the weight of the LoRA adjustments and any other specific hyperparameters.


In conclusion, the LoRA Stable Diffusion model represents a significant advancement in the field of AI-generated art and content creation. Utilize Novita AI APIs of LoRA models to train your own model to fine-tune the Stable Diffusion models. Overall, the LoRA Stable Diffusion model is a powerful tool that, when used responsibly, can greatly enhance the capabilities of AI in artistic and creative applications.

Novita AI, the one-stop platform for limitless creativity that gives you access to 100+ APIs. From image generation and language processing to audio enhancement and video manipulation, cheap pay-as-you-go, it frees you from GPU maintenance hassles while building your own products. Try it for free.
Recommended reading
  1. Add LoRA Stable Diffusion: A Comprehensive Guide
  2. LoRA AI art generator: Create art with ease
  3. Lora AI-Generated Images — The Future of Art