Enter a unique name for your dataset

Dataset ID *

This identifier must be unique across all datasets in your configuration.

Treat this as a regularisation dataset

Regularisation datasets help preserve the base model's identity. You must also include at least one normal image or video dataset.

Add a primary image or video dataset first. Once one is queued, you can enable regularisation for supporting datasets.

Create new standalone config

Unchecked (default): Add datasets to your current environment's config.
Checked: Create a completely new config file (includes required text embed datasets).

Dataset Type

Image Dataset

For training with still images

Video Dataset

For training with video sequences Model not marked video-capable; continue with caution.

Audio Dataset

For training with audio/speech

Where is your dataset stored?

Configure required settings

Cache Dataset Options

Create separate dedicated datasets for caching embeddings. This allows you to reuse cached data across multiple training datasets.

Create separate text embeddings cache dataset

Creates a dedicated text_embeds dataset that can be shared across multiple image datasets.

Create separate VAE latent cache dataset

Creates a dedicated image_embeds dataset for VAE latent caching. Useful when training multiple models with the same images.

Configure Text Embeddings Cache

This dedicated dataset will store cached text embeddings that can be reused across multiple training sessions.

Text Embeds Dataset ID *

Unique identifier for this text embeddings cache dataset

Cache Directory *

Directory where text embeddings will be stored

Tip: All image datasets in your configuration will automatically use this shared text embeddings cache.

Safe to Share: Multiple datasets can safely use the same text embed cache directory. Common captions will be deduplicated, reducing disk space usage.

Configure VAE Latent Cache Backend

The image_embeds dataset provides a storage backend for VAE latents. Each image dataset can point to its own cache directory within this backend.

VAE Cache Dataset ID *

Unique identifier for this VAE cache backend dataset

Select Storage Backend

Local Storage

Store VAE cache on local filesystem

S3-Compatible

Store VAE cache in S3-compatible storage

S3-Compatible Storage Configuration

S3 Bucket Name *

Name of the S3 bucket to store VAE cache

S3 Endpoint URL

S3-compatible endpoint URL (e.g., AWS S3, Cloudflare R2, MinIO, etc.)

Access Key ID

S3 access key (optional if using IAM role)

Secret Access Key

S3 secret key (optional if using IAM role)

Region Name

Region name (optional, leave empty to use default)

VAE Cache Directory * VAE Cache Prefix *

Path where VAE latents will be stored for your image dataset. This will be stored in the image dataset config, not the image_embeds dataset.

S3 prefix where VAE latents will be stored for your image dataset. This will be stored in the image dataset config, not the image_embeds dataset.

Note: Multiple image datasets can share the same image_embeds backend by pointing to different .

Warning: Each dataset must use a unique VAE cache . Sharing the same cache location across datasets will cause data corruption.

Configure image resolution settings

Resolution Type

pixel_area (recommended) maintains aspect ratios • pixel uses fixed edge length • area targets direct megapixels

pixel_area (Recommended) pixel area

Base Resolution

Higher resolutions require more VRAM but produce better quality. 1024 is recommended for most GPUs.

Or enter custom resolution:

Maximum Image Size

Clamp before downsampling (match resolution type)

Target Downsample Size

Secondary clamp for very large assets

Using default resolution: 1024px, pixel_area

Configure image cropping behavior

Enable Image Cropping

When disabled, images will only be resized without cropping

Crop Style

Choose where crops are taken from in the image

Random

Center

Corner

Face

Crop Aspect Ratio

Choose how aspect ratios are handled during cropping

Square

Preserve

Closest

Random

Aspect Ratio Buckets

Define aspect ratio buckets (e.g., 0.75 = 3:4, 1.0 = 1:1, 1.33 = 4:3)

Configure caption settings

Caption Strategy

How captions should be loaded for each image

Parquet Configuration

Parquet/JSONL File Path *

Path to the parquet or JSONL metadata file

Filename Column *

Column containing image IDs/filenames

Caption Column *

Column containing image captions

Width Column (optional)

Column with image width (speeds up loading)

Height Column (optional)

Column with image height (speeds up loading)

Fallback Caption Column (optional)

Fallback column when primary caption is empty

Filename includes extension

Prepend Instance Prompt

Add a consistent trigger word or phrase to the beginning of all captions (e.g., "a photo of xyz")

Instance Prompt *

Auto-generate conditioning data for ControlNet

ControlNet training detected! Would you like to automatically generate conditioning data during training?

Select Conditioning Generator

Canny Parameters

Low Threshold

High Threshold

Review Your Dataset Configuration

Backend Type:

Dataset Type:

Configuration:

Instance Data Path:

Metadata Backend:

Caption Strategy:

VAE Cache Dir:

Resolution:

Size:

Type:

Cropping:

Captions:

Strategy:

Additional Settings:

Probability:

Repeats: