Settings

Here is a list of all the setting that one can edit in the extension

Create model tab

Name
The name of the model to create.

Import Model from Huggingface Hub
Import a model from Huggingface.co instead of using a local checkpoint. Hub model MUST contain diffusion weights.
Source Checkpoint
The source checkpoint to extract for training.

Extract EMA Weights
If EMA weights are saved in a mode. these will be extracted instead of the full Unet. Probably not necessary for training or fine-tuning.

Scheduler
Model scheduler to use. Only applies to models before 2.0.

Settings model tab

WIP

Concepts

WIP

Saving

Custom Model Name
A custom name to use when saving .ckpt and .pt files. Subdirectories will also be named this.

Half Model
Enable this to generate model with fp16 precision. Results in a smaller checkpoint with minimal loss in quality.

Save Checkpoint to Subdirectory
When enable. checkpoints will be saved to a subdirectory in the selected checkpoints folder.


Generate A .ckpt file when saving during training.
When enable. a checkpoint will be generated at the specified epoch intervals while training is active. This also controls manual generation using the 'save weights' button while training is active.

Generate A .ckpt file when training completes.
When enable. a checkpoint will be generated when training completes successfully.

Generate A .ckpt file when training is cancelled.
When enable. a checkpoint will be generated when training is cancelled by the user.

Save separate diffusers snapshots when saving during training.
When enable. a unique snapshot of the diffusion weights will be saved at each specified epoch interval. This uses more HDD space (A LOT. but allows resuming from trainin. including the optimizer state.

Save separate diffusers snapshots when training completes.
When enable. a unique snapshot of the diffusion weights will be saved when training completes. This uses more HDD spac. but allows resuming from training including the optimizer state.

Save separate diffusers snapshots when training is cancelled.
When enable. a unique snapshot of the diffusion weights will be saved when training is canceled. This uses more HDD spac. but allows resuming from training including the optimizer state.

Generate

Generate Class Images
Create classification images using training settings without training.

Preview Prompts
Generate a JSON representation of prompt data used for training.

Generate Graph
Generate graphs from training logs showing learning rate and loss averages over the course of training.

Graph Smoothing Steps
How many timesteps to smooth graph data over. A lower value means a more jagged graph with more informatio. higher value will make things prettier but slightly less accurate.

Generate Sample Images
Generate sample images using the currently saved diffusers model.

Sample Prompt
The prompt to use to generate a sample imag.

Sample Negative Prompt
A negative prompt to use when generating preview images.

Sample Seed
The seed to use when generating samples. Set to -1 to use a random seed every time.

Number of Samples to Generate
How many samples to generate. Sample Steps
The number of steps to use when generating the sample image

Sample CFG Scale
The Classifier-Free Guidance Scale to use for preview images.