MakeTrainingDataset - ComfyUI Built-in Node Documentation

This node prepares data for training by encoding images and text.

This node prepares data for training by encoding images and text. It takes a list of images and a corresponding list of text captions, then uses a VAE model to convert the images into latent representations and a CLIP model to convert the text into conditioning data. The resulting paired latents and conditioning are output as lists, ready for use in training workflows.

Inputs

Parameter	Description	Data Type	Required	Range
images	List of images to encode.	IMAGE	Yes	N/A
vae	VAE model for encoding images to latents.	VAE	Yes	N/A
clip	CLIP model for encoding text to conditioning.	CLIP	Yes	N/A
texts	List of text captions. Can be length n (matching images), 1 (repeated for all), or omitted (uses empty string).	STRING	No	0, 1, or n items (n = number of images)

Parameter Constraints:

• This node uses list inputs: images and texts are processed as lists, while vae and clip accept a single model each (the first item of the supplied list is used).
• The number of items in the texts list must be 0, 1, or exactly match the number of items in the images list. If it is 0 or omitted, an empty string is used for all images. If it is 1, that single text is repeated for all images. Any other length raises an error.
• The output latents and conditioning lists always contain the same number of items as the images list, so each latent is paired with the conditioning for its corresponding caption.

Outputs

Output Name	Description	Data Type
latents	List of latent dicts.	LATENT
conditioning	List of conditioning lists.	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): 244adc98810a874cfe42f834e89f96da300d883faeb5791dff19607c13d0c0db