Stable Cascade LORA training

24 Feb 202411:46

TLDRThe video provides a fast-paced tutorial on training a Stable Cascade LORA model using One Trainer. It begins with installing One Trainer, preparing the dataset, and loading presets. The process includes defining the concept, which is the dataset, and checking training settings. The speaker emphasizes the importance of matching file names and captions, and manually installing large files like PyTorch to avoid internet connectivity issues. The video also covers the installation of the fnet encoder, setting up the training dataset, and adjusting additional options before starting the training. The training results are analyzed, and the speaker shares insights on optimizing training and the potential impact of new updates like Stable Diffusion 3. The tutorial concludes with a teaser for future training experiments and releases from Cascade.


  • ๐Ÿ’ป **Installation Process**: The video begins with the installation of One Trainer, which includes setting up a Python virtual environment and manually installing PyTorch.
  • ๐Ÿ“š **Data Preparation**: The presenter discusses preparing a dataset with consistent images and captions for training the model.
  • ๐Ÿ”„ **Cloning and Environment Setup**: The process involves cloning the One Trainer repository and setting up a Python virtual environment to avoid system conflicts.
  • ๐Ÿ“ˆ **Training Settings**: Before starting the training, it's important to check and adjust the training settings, including learning rate, SNR gamma, and noise weights.
  • ๐Ÿ“ **Defining Concepts**: Concepts are defined in the training software, which involves adding a concept and specifying the path to the dataset.
  • ๐Ÿ”— **Downloading Models**: Certain models, like the fnet encoder, need to be downloaded manually and placed in the correct folder within the One Trainer directory.
  • ๐Ÿ”ง **Configuration Options**: The video highlights additional configuration options available in the software, such as toggling on and off different concepts and making additional configs.
  • ๐Ÿš€ **Starting Training**: After setting up the environment and defining the concept, the training can be started, and progress can be monitored through TensorBoard.
  • ๐Ÿ“Š **Monitoring Progress**: It's recommended to read through the options and settings before starting the training to understand the process and expected outcomes.
  • ๐Ÿ” **Results and Evaluation**: Once training is complete, the model's performance can be evaluated, and the trained model can be found within the One Trainer's models directory.
  • ๐Ÿ”ง **UI and Model Loading**: The presenter mentions issues with the UI not fully respecting certain blocks and the need for waiting until converters are available for better results.
  • โฑ๏ธ **Training Time and Resources**: The video provides an example of a trained model, noting the training time and VRAM usage, which can vary depending on the system.

Q & A

  • What is the first step in the Stable Cascade LORA training process?

    -The first step is to install the One Trainer, which includes setting up the environment and installing necessary packages like PyTorch.

  • How does one download and install PyTorch manually to avoid internet connection issues?

    -You can download the PyTorch file manually, place it in a desired folder, and then use the command line to follow the instructions for installation.

  • What is the purpose of creating a Python virtual environment during the installation process?

    -Creating a Python virtual environment isolates the project's dependencies from the system Python, ensuring that the project has a consistent and controlled environment.

  • What is the role of the 'requirements.txt' file in the installation process?

    -The 'requirements.txt' file lists all the dependencies needed for the project. It is used to install these dependencies using pip, ensuring that the project has all the necessary packages.

  • How does one prepare the dataset for Stable Cascade LORA training?

    -You need to have files and images matching, with the file names matching and the caption describing the image. The dataset should be organized and placed in a specific path that will be used during the training.

  • What is the significance of the fnet encoder in the Stable Cascade LORA training?

    -The fnet encoder is a required component for the training process. It needs to be downloaded manually and placed in the 'models' folder within the One Trainer directory.

  • How does one define a concept in the One Trainer UI?

    -In the One Trainer UI, you go to the 'Concepts' tab, click 'Add Concept', select the folder containing your dataset, and then provide the path to the dataset.

  • What are the additional options that can be configured before starting the training?

    -Additional options include SNR gamma, offset noise weights, and learning rate. These can be found and adjusted in the settings before starting the training.

  • How long does it typically take to train a Stable Cascade LORA model?

    -The training time can vary based on the complexity of the model and the hardware used. In the script, it is mentioned that a model took about 40 minutes to train.

  • What is the impact of using different LORA weights in the training process?

    -Different LORA weights can affect the quality of the generated images. The script mentions experimenting with the Inspire pack and block weights to achieve slightly better results.

  • What is the current status of the LORA model in the One Trainer UI?

    -As of the time of the script, the LORA model is trained with the keys, but the UI is not fully rendering them. There is ongoing work to improve the UI to properly use block weights.

  • What is the next step or upcoming development mentioned in the script?

    -The script mentions the arrival of NSD3 (Stable Diffusion 3) and the anticipation of more training experiments and releases from Cascade as more people start training LORA models.



๐Ÿš€ Installing One Trainer and Preparing the Dataset

The video begins with a quick overview of the process, which includes installing a trainer, preparing a dataset, loading presets, checking training settings, defining the concept, and starting the training. The speaker emphasizes the need to occasionally pause the video due to the fast pace. They guide viewers on how to install One Trainer, mentioning the use of a specific version of CUDA and PyTorch, and provide a step-by-step method to clone the repository and set up a Python virtual environment. The speaker also discusses manual installation of PyTorch to avoid issues with a poor internet connection and advises on upgrading pip. The final steps involve preparing the dataset with matching file names and captions, and using One Trainer's UI to load presets and configure the training environment.


๐Ÿ“š Configuring Training Settings and Starting the Training

After loading the preset, the video moves on to setting up the training environment. The speaker provides a link to download the required fnet encoder manually and guides on where to place it within the One Trainer's model folder. They then explain how to add the training dataset by defining a concept in the UI, specifying the path to the dataset. The video also covers additional training options, such as SNR gamma, offset noise weights, and learning rate, which can be adjusted according to preference. The speaker advises reading through the options before starting the training and suggests saving preferred settings for future use. Once everything is set, the training can be initiated, and progress can be monitored through TensorBoard graphs. The video concludes with a discussion on the results and the expected model performance, including VRAM usage and training time.


๐Ÿ” Dealing with Multiple Versions of Lura and Upcoming Developments

The final paragraph addresses the challenges of finding a consistent version of Lura to use, as there are many variations available. The speaker mentions that converters are being developed to address key discrepancies. They also discuss the potential for improved image quality once the UI fully supports block weights. The speaker shares their experience with training Luras on the Cascade model and hints at more releases and training experiments in the future. They also touch on the recent developments with stable diffusion 3, expressing anticipation for its release and the impact it may have on training and image generation.



๐Ÿ’กStable Cascade

Stable Cascade refers to a specific type of model training in the field of artificial intelligence, particularly in the context of image generation. In the video, it is the main focus of the training process that the speaker is guiding the audience through. It involves using a dataset to train a model to generate images that are consistent and stable in terms of the content they depict.


LORA stands for Low-Rank Adaptation, which is a technique used in the training of neural networks. It involves modifying the weights of a pre-trained model in a low-rank manner to adapt it to a new task. In the video, LORA is used in conjunction with Stable Cascade training to refine the model's performance.

๐Ÿ’กOne Trainer

One Trainer is a software tool mentioned in the video that is used for training AI models. It provides a user interface for managing the training process, including setting up datasets, adjusting training parameters, and monitoring the progress of the model as it learns.


A dataset is a collection of data that is used for training machine learning models. In the context of the video, the speaker discusses preparing a dataset consisting of images with matching captions to train the AI in recognizing and generating specific concepts.


Presets in the context of the video refer to pre-configured settings within the One Trainer software that can be loaded to apply specific configurations for training models. The speaker mentions loading presets for Cascade and LORA training.


In the video, a concept refers to a specific idea or theme that the AI model is being trained to recognize and generate images for. The speaker defines a concept by associating a dataset with a particular theme or subject.

๐Ÿ’กTraining Settings

Training settings are the parameters and configurations that define how a machine learning model will be trained. The video script discusses checking these settings to ensure they are correctly set up for the Stable Cascade LORA training process.

๐Ÿ’กVirtual Environment

A virtual environment in the context of the video is a tool used in software development to create isolated Python environments. The speaker mentions creating a Python virtual environment to install dependencies for the One Trainer without affecting the system's default Python setup.



Pip is a package installer for Python that allows users to install and manage software packages. In the video, the speaker talks about using pip to install the required packages as listed in the 'requirements.txt' file for setting up the One Trainer.


VRAM stands for Video Random Access Memory and refers to the memory used by a graphics processing unit (GPU) to store image data. The video mentions the amount of VRAM required for training the AI model, indicating the computational resources needed for the process.

๐Ÿ’กStable Diffusion 3

Stable Diffusion 3 is a newer version of a model or software mentioned towards the end of the video. It signifies an upcoming advancement or update in the AI field that the speaker has access to but has not yet fully utilized.


The video provides a quick overview of the Stable Cascade LORA training process using One Trainer.

Installation of One Trainer is the first step, followed by preparing the dataset.

Loading presets and checking training settings are crucial before starting the training.

Defining the concept, which refers to the dataset, is a key step in the training process.

The video demonstrates how to manually install large files like PyTorch to avoid internet connectivity issues.

Creating a Python virtual environment ensures that system Python is unaffected.

The fnet encoder is required for Stable Cascade and needs to be downloaded manually.

The video shows how to add a concept to the training dataset in One Trainer.

The training settings can be adjusted and saved for future use.

The training process is monitored through TensorBoard for real-time progress.

Once training is complete, the model is saved within the One Trainer's models directory.

The video compares training results with and without the LORA being loaded.

The training process is resource-intensive, requiring a significant amount of VRAM.

The presenter discusses the use of different text-to-image models and their impact on training outcomes.

The video mentions the upcoming release of Stable Diffusion 3 and its potential impact on training.

The presenter plans to conduct more training experiments to better understand the LORA model.

The video concludes with an invitation to try out the presented LORA model for testing purposes.