Visualize Latents

Documentation

• Class name: VisualizeLatents
• Category: utils
• Output node: False

The VisualizeLatents node transforms latent space representations into visualizable images, adjusting their mean and standard deviation for better visual interpretation. It organizes the visualized latents into a grid layout based on the square root of the number of channels, facilitating the examination of latent space features.

Input types

Required

• latent
  • The 'latent' input is a dictionary containing the latent space samples to be visualized. It is crucial for generating the visual representation of the latent space, affecting the output image's appearance and layout.
  • Comfy dtype: LATENT
  • Python dtype: Dict[str, torch.Tensor]

Output types

• image
  • Comfy dtype: IMAGE
  • The output is an image tensor representing the visualized latent space, adjusted for better visual interpretation and organized in a grid layout.
  • Python dtype: torch.Tensor

Usage tips

• Infra type: GPU
• Common nodes: unknown

Source code

class VisualizeLatents:
    @classmethod
    def INPUT_TYPES(s):
        return {"required": {"latent": ("LATENT", ),}}

    RETURN_TYPES = ("IMAGE",)
    FUNCTION = "visualize"

    CATEGORY = "utils"

    def visualize(self, latent):
        latents = latent["samples"]
        batch, channels, height, width = latents.size()

        latents = latents - latents.mean()
        latents = latents / latents.std()
        latents = latents / 10 + 0.5

        scale = int(channels ** 0.5)
        vis = torch.zeros(batch, height * scale, width * scale)

        for i in range(channels):
            start_h  = (i % scale) * height
            end_h    = start_h + height
            start_w  = (i // scale) * width
            end_w    = start_w + width

            vis[:, start_h:end_h, start_w:end_w] = latents[:, i]

        return (vis.unsqueeze(-1).repeat(1, 1, 1, 3),)