Text Embeds Dataset

Image Embeds Dataset

Conditioning Image Embeds

For I2V models (Wan 2.1/2.2, Qwen Edit)

Conditioning Embed Cache Dir

Preserve backend cache

Keep data backend cache between runs

Update Interval (seconds)

How often to refresh dataset metadata during training

Encode VAE latents on demand

Encodes missing latents during training and saves them to the VAE cache.

Disable VAE cache writes

Encodes latents on demand for this dataset without writing generated latents to disk.

Clear VAE cache each epoch

Enable for random crop/aspect training to regenerate latents per epoch.

Delete unwanted images

Permanently deletes images that fail size or aspect ratio filters

Delete problematic images

Permanently deletes images that fail VAE encoding (corrupted files)