When working with these datatypes, you will need to know about the torch.Tensor class. Complete documentation is here, or an introduction to the key concepts required for Comfy here.

If your node has a single output which is a tensor, remember to return (image,) not (image)

Most of the concepts below are illustrated in the example code snippets.

​
Images

An IMAGE is a torch.Tensor with shape [B,H,W,C], C=3. If you are going to save or load images, you will need to convert to and from PIL.Image format - see the code snippets below! Note that some pytorch operations offer (or expect) [B,C,H,W], known as ‘channel first’, for reasons of computational efficiency. Just be careful.

​
Working with PIL.Image

If you want to load and save images, you’ll want to use PIL:

from PIL import Image, ImageOps

​
Masks

A MASK is a torch.Tensor with shape [B,H,W]. In many contexts, masks have binary values (0 or 1), which are used to indicate which pixels should undergo specific operations. In some cases values between 0 and 1 are used indicate an extent of masking, (for instance, to alter transparency, adjust filters, or composite layers).

​
Masks from the Load Image Node

The LoadImage node uses an image’s alpha channel (the “A” in “RGBA”) to create MASKs. The values from the alpha channel are normalized to the range [0,1] (torch.float32) and then inverted. The LoadImage node always produces a MASK output when loading an image. Many images (like JPEGs) don’t have an alpha channel. In these cases, LoadImage creates a default mask with the shape [1, 64, 64].

​
Understanding Mask Shapes

In libraries like numpy, PIL, and many others, single-channel images (like masks) are typically represented as 2D arrays, shape [H,W]. This means the C (channel) dimension is implicit, and thus unlike IMAGE types, batches of MASKs have only three dimensions: [B, H, W]. It is not uncommon to encounter a mask which has had the B dimension implicitly squeezed, giving a tensor [H,W].

To use a MASK, you will often have to match shapes by unsqueezing to produce a shape [B,H,W,C] with C=1 To unsqueezing the C dimension, so you should unsqueeze(-1), to unsqueeze B, you unsqueeze(0). If your node receives a MASK as input, you would be wise to always check len(mask.shape).

​
Latents

A LATENT is a dict; the latent sample is referenced by the key samples and has shape [B,C,H,W], with C=4.

LATENT is channel first, IMAGE is channel last