LoadTrainingDataset - ComfyUI Built-in Node Documentation

This node loads an encoded training dataset (latents and conditioning) that was previously saved to disk.

This node loads an encoded training dataset (latents and conditioning) that was previously saved to disk. It reads all shard_*.pkl data shard files from a selected dataset folder in the datasets directory and returns the combined latent vectors and conditioning data for use in training workflows.

Inputs

Parameter	Description	Data Type	Required	Range
folder_name	Saved dataset to load, from the datasets directory.	COMBO	Yes	One option per dataset folder found in the datasets directory

Note: The folder_name options are built automatically by scanning the datasets directory. A subfolder is listed as a dataset when it contains a metadata.json file or at least one .safetensors file (the scan does not descend into a folder that matches). The selected dataset folder is searched across all configured dataset root directories, and the folder name must resolve to a subfolder inside one of those roots. The node reads all files named shard_*.pkl in the selected folder, in sorted order, and raises an error if no shard files are found or if the folder cannot be located.

Outputs

Output Name	Description	Data Type
latents	List of latent dicts (output list), where each dict contains a "samples" key with a tensor.	LATENT
conditioning	List of conditioning lists (output list), where each inner list contains conditioning data for the corresponding sample.	CONDITIONING

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

---

Source fingerprint (SHA-256): 9f914b27f067460f6f3b54f3f2a7bb793c65b99c85e8aa14ab64894be26bd816