Model identifier from huggingface.co/models

Default: "stabilityai/stable-diffusion-2-1"

The prompt you use to describe your training images, in the format: `a [identifier] [class noun]`, where the `[identifier]` should be a rare token. Relatively short sequences with 1-3 letters work the best (e.g. `sks`, `xjy`). `[class noun]` is a coarse class descriptor of the subject (e.g. cat, dog, watch, etc.). For example, your `instance_prompt` can be: `a sks dog`, or with some extra description `a photo of a sks dog`. The trained model will learn to bind a unique identifier with your specific subject in the `instance_data`.

The prompt or description of the coarse class of your training images, in the format of `a [class noun]`, optionally with some extra description. `class_prompt` is used to alleviate overfitting to your customised images (the trained model should still keep the learnt prior so that it can still generate different dogs when the `[identifier]` is not in the prompt). Corresponding to the examples of the `instant_prompt` above, the `class_prompt` can be `a dog` or `a photo of a dog`.

Flag to add prior preservation loss.

Default: true

The weight of prior preservation loss.

Default: 1

A ZIP file containing your training images (JPG, PNG, etc. size not restricted). These images contain your 'subject' that you want the trained model to embed in the output domain for later generating customized scenes beyond the training images. For best results, use images without noise or unrelated objects in the background.

An optional ZIP file containing the training data of class images. This corresponds to `class_prompt` above, also with the purpose of keeping the model generalizable. By default, the pretrained stable-diffusion model will generate N images (determined by the `num_class_images` you set) based on the `class_prompt` provided. But to save time or to have your preferred specific set of `class_data`, you can also provide them in a ZIP file.

Minimal class images for prior preservation loss. If not enough images are provided in class_data, additional images will be sampled with class_prompt.

Default: 50

A seed for reproducible training.

Default: 1337

The resolution for input images. All the images in the train/validation dataset will be resized to this resolution.

Default: 768

Whether to center crop the input images to the resolution. If not set, the images will be randomly cropped. The images will be resized to the resolution first before cropping.

Default: false

Whether to train the text encoder. If set, the text encoder should be float32 precision.

Default: true

Batch size (per device) for the training dataloader.

Default: 1

Batch size (per device) for sampling images.

Default: 4

Default: 1

Total number of training steps to perform. If provided, overrides num_train_epochs.

Default: 2000

Number of updates steps to accumulate before performing a backward/update pass.

Default: 1

Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.

Default: false

Initial learning rate (after the potential warmup period) to use.

Default: 0.000005

Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.

Default: false

The scheduler type to use

Default: "constant"

Number of steps for the warmup in the lr scheduler.

Default: 0

Number of hard resets of the lr in cosine_with_restarts scheduler.

Default: 1

Power factor of the polynomial scheduler.

Default: 1

Whether or not to use 8-bit Adam from bitsandbytes.

Default: false

The beta1 parameter for the Adam optimizer.

Default: 0.9

The beta2 parameter for the Adam optimizer.

Default: 0.999

Weight decay to use

Default: 0.01

Epsilon value for the Adam optimizer

Default: 1e-8

Max gradient norm.

Default: 1

Whether or not to horizontally flip training images 50 percent of the time.

Default: false

json of samples to generate: [{"name": "sample_name", "input": {"prompt": "a sks dog", "num_samples": 4, "save_guidance_scale": 7.5, "save_infer_steps": 50}}]