Stable Cascade LORA training

24 Feb 202411:46

TLDRIn this fast-paced tutorial, the speaker guides viewers through the process of setting up and using the Stable Cascade LORA training with One Trainer. The video covers installation of the One Trainer, preparation of the dataset, loading presets, checking training settings, defining the concept, and finally, starting the training. The speaker emphasizes the importance of a stable internet connection for downloading large files like PyTorch and provides a workaround for those with unstable connections. The tutorial also includes tips on creating a Python virtual environment, manually installing PyTorch, and using the One Trainer UI to load presets and configure the training process. The speaker discusses the use of the fnet encoder, setting up the training dataset, and provides insights into the training options and settings. The video concludes with a look at the training results and the potential for future developments with the integration of new models like Stable Diffusion 3.


  • ๐Ÿ› ๏ธ Install One Trainer by following the instructions in the 'install.bat' file, which references the 'requirements.txt'.
  • ๐Ÿ“š Prepare your dataset by ensuring that image file names match their captions for consistency.
  • ๐Ÿ”— Manually install large files like PyTorch to avoid issues with unstable internet connections.
  • ๐Ÿ’พ Create a Python virtual environment to isolate the project dependencies from your system's Python.
  • ๐Ÿ“ Organize your project with a 'models' folder to store necessary files like the fnet encoder.
  • ๐Ÿ“ˆ Use the One Trainer UI to load presets for Stable Cascade and LORA training.
  • ๐Ÿ“‹ Define your concept in the UI by adding a concept and providing the path to your prepared dataset.
  • โš™๏ธ Review and adjust training settings such as SNR, gamma, learning rate, and epochs before starting.
  • ๐Ÿ” Enable or disable concept toggling and create additional configs for different concept groupings.
  • ๐Ÿš€ Start the training process and monitor progress through the UI and TensorBoard graphs.
  • ๐Ÿ“‰ Expect variations in training times and VRAM usage depending on your system's specifications.
  • ๐Ÿ”„ Be patient for updates and converters for the UI to fully support block weights in LORA training.

Q & A

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

    -The first step is to install the One Trainer by following the instructions in the 'install.bat' file and downloading the necessary requirements, such as CU 1118 for Python 3.10.

  • How does the speaker handle the installation of large files like PyTorch with a potentially unstable internet connection?

    -The speaker manually downloads the PyTorch file and places it in the desired installation folder, then uses the command line to install it without relying on an uninterrupted internet connection.

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

    -Creating a Python virtual environment isolates the project's dependencies from the system's Python installation, preventing conflicts and ensuring the correct versions of libraries are used.

  • Why is the fnet encoder required for Stable Cascade LORA training?

    -The fnet encoder is required as it is a part of the model architecture used in the training process, and it needs to be downloaded manually as it's not included in the main installation.

  • How does one prepare the dataset for training?

    -The dataset should consist of images with matching filenames and a caption that describes the image. The images and captions are organized in a folder, and the path to this folder is used in the training setup.

  • What is the significance of matching filenames and captions in the dataset?

    -Matching filenames and captions ensure that the training algorithm can correctly associate each image with its corresponding description, which is crucial for the model to learn and generate accurate outputs.

  • How does one add a concept to the training setup in One Trainer?

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

  • What additional configurations can be adjusted in the training settings?

    -Additional configurations include SNR gamma, offset noise weights, learning rate, and epochs. These can be adjusted through the UI for fine-tuning the training process.

  • What does the speaker suggest doing after the training is completed?

    -After training, the speaker suggests looking at the output results and the model's performance using TensorBoard to analyze the training process and the model's effectiveness.

  • How long did it take for the speaker to train a specific model, and what resources were required?

    -The speaker mentions that it took about 40 minutes to train a model with a batch size of 1 and a VRAM usage of approximately 12.8 gigabytes.

  • What is the speaker's view on the future of training models on Cascade?

    -The speaker anticipates that with the training process now available, more people will start training models on Cascade, leading to more releases and experimentations as they get to grips with the new tools and techniques.

  • What is the speaker's current status regarding access to Stable Diffusion 3?

    -The speaker has been granted access but has not yet received Stable Diffusion 3, suggesting that there might be a queue or certain criteria (like being part of a 'red team' or equivalent) to actually get access to it.



๐Ÿ˜€ Installing One Trainer and Preparing the Data Set

The video begins with a quick overview of the process, which includes installing a trainer, preparing a data set, loading presets, checking training settings, defining the concept, and starting the training. The speaker emphasizes the need to pause the video due to the fast pace and provides a step-by-step guide on how to install the trainer using a specific version of PyTorch, which is manually downloaded and installed to accommodate users with slow internet connections. The process involves cloning a repository, creating a Python virtual environment, and running a requirements text file. The video also touches on upgrading pip and deactivating the virtual environment before the batch file restarts it.


๐Ÿ“š Setting Up the Training Environment and Data

After installing One Trainer, the video moves on to preparing the data set. The speaker uses a test set from an old model with a small number of images and a consistent caption to demonstrate the process. The key is to ensure that file names match and captions describe the images. The viewer is guided on how to open One Trainer, load presets (specifically Cascade and Laura), and manually install the fnet encoder, which is required for training. The video also covers how to add concepts to the training data set, toggle them on and off, and make additional configurations for grouping concepts. It concludes with a reminder to read and understand the training options before starting the training process.


๐Ÿ”ง Training the Model and Discussing Variants of Lora

The video continues with the training process, explaining how to start training and monitor progress through TensorBoard. It mentions the output results and provides an example of a trained model's performance. The speaker discusses the use of Argus text-to-image for Cascade and experimenting with different configurations. They also touch on the challenges of finding a consistent version of Lora to use, as there are many variants. The video acknowledges that converters are being developed to address this issue. The speaker shares their anticipation for more training experiments and the upcoming release of Stable Diffusion 3, which they have access to but have not yet received. The video concludes with the speaker's intention to leave the current setup for testing and encourages viewers to experiment with Lora as more models become available.



๐Ÿ’กStable Cascade

Stable Cascade refers to a specific type of training process for generative models, likely related to image synthesis or style transfer. In the video, it is the main focus and the process that the user guides the audience through, starting from installation to training completion.


LORA stands for Low-Rank Adaptation, a technique used in machine learning to modify pretrained models with lower computational costs. In the context of the video, LORA is used in conjunction with Stable Cascade for efficient model training.

๐Ÿ’กOne Trainer

One Trainer is a software tool mentioned in the video used for managing the training of AI models. The script details the process of installing and using One Trainer to prepare for and execute the Stable Cascade LORA training.

๐Ÿ’กData Set

A data set is a collection of data used for analysis or machine learning. In the video, the data set is prepared and loaded into One Trainer for the purpose of training the AI model, which is a crucial step in the process.


Presets in the context of the video refer to pre-configured settings within the One Trainer software that are used to streamline the training process. The user loads specific presets for the Stable Cascade LORA training.


In the video, a concept refers to a specific category or idea that the AI model is being trained to recognize or generate. The user defines a concept by associating it with a data set for the training process.

๐Ÿ’กTraining Settings

Training settings are the parameters and configurations that dictate how the AI model is trained. The video discusses checking and adjusting these settings to ensure the training process aligns with the user's goals.

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 user creates a virtual environment to install dependencies for One Trainer without affecting the system's Python installation.


Pip is a package installer for Python. In the video, it is used to install the required packages listed in the 'requirements.txt' file for setting up the One Trainer environment.


PyTorch is an open-source machine learning library based on the Torch library. It is mentioned in the video as a large file that needs to be manually installed as part of setting up the training environment.


VRAM stands for Video Random Access Memory and refers to the memory used by graphics processing units (GPUs). In the video, the user mentions the amount of VRAM required for training the AI model, indicating the computational resources needed.

๐Ÿ’กStable Diffusion 3

Stable Diffusion 3 is a newer version of a generative model or software mentioned towards the end of the video. It signifies an upcoming advancement in the field that the user is looking forward to exploring.


The video provides a quick overview of installing One Trainer and setting up a Stable Cascade LORA training session.

The process begins with installing One Trainer and preparing the dataset.

Loading One Trainer presets and checking training settings are crucial steps.

Defining the concept, which refers to the dataset, is a key part of the training process.

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

Creating a Python virtual environment is recommended to prevent system Python interference.

The video explains how to clone the repository and set up the necessary files for training.

The importance of matching file names and captions for the dataset is emphasized for effective training.

The Stable Cascade LORA training utilizes the fnet encoder, which needs to be downloaded manually.

The video outlines how to add concepts to the training dataset within the One Trainer UI.

It is advised to read and understand the training options before starting the training process.

The video discusses the potential need to upgrade pip within the virtual environment.

Once training is complete, the model can be found within the One Trainer models directory.

The video shows the difference in output results with and without the LORA being loaded.

The training session's duration and VRAM usage are discussed, providing insights into system requirements.

The video touches on the use of Argus text to image for Cascade and experimenting with different packs for improved results.

The presenter shares their experience with training a rank 16 model and the resources it required.

The video concludes with a teaser for upcoming training experiments and the anticipation of Stable Diffusion 3.