Skip to main content
Latent space nodes work with compressed image representations before they’re decoded to pixels. Operations in latent space are faster and more memory-efficient than pixel-space operations.

​
Latent Creation

​
EmptyLatentImage

Creates a new batch of empty latent images to be denoised. Category: latent
​
width
int
default:"512"
Width of the latent in pixels (will be divided by 8 for actual latent dimensions). Range: 16 to 16384, step 8
​
height
int
default:"512"
Height of the latent in pixels. Range: 16 to 16384, step 8
​
batch_size
int
default:"1"
Number of latent images in the batch. Range: 1 to 4096
Returns:
  • LATENT: Empty latent image batch
Description: Creates blank latent tensors for text-to-image generation. The actual latent dimensions are width/8 × height/8 because latents are 8x compressed.

​
Encoding & Decoding

​
VAEEncode

Converts a pixel-space image to latent space. Category: latent
​
pixels
IMAGE
required
The image to encode
​
vae
VAE
required
The VAE model for encoding
Returns:
  • LATENT: The encoded latent representation
Description: Compresses an image into latent space. Used for img2img workflows or when you need to manipulate an image in latent space.

​
VAEDecode

Converts a latent image back to pixel space. Category: latent
​
samples
LATENT
required
The latent to decode
​
vae
VAE
required
The VAE model for decoding
Returns:
  • IMAGE: The decoded pixel-space image
Description: Decodes latent images back into pixel space. This is typically the final step before saving an image.

​
VAEEncodeTiled

Encodes large images using a tiled approach to save memory. Category: _for_testing
​
pixels
IMAGE
required
The image to encode
​
vae
VAE
required
The VAE model
​
tile_size
int
default:"512"
Size of each tile. Range: 64 to 4096, step 64
​
overlap
int
default:"64"
Overlap between tiles to avoid seams. Range: 0 to 4096, step 32
​
temporal_size
int
default:"64"
Frames to encode at once (for video VAEs). Range: 8 to 4096, step 4
​
temporal_overlap
int
default:"8"
Frame overlap (for video VAEs). Range: 4 to 4096, step 4
Returns:
  • LATENT: The encoded latent

​
VAEDecodeTiled

Decodes large latents using tiles to save memory. Category: _for_testing
​
samples
LATENT
required
The latent to decode
​
vae
VAE
required
The VAE model
​
tile_size
int
default:"512"
Tile size in pixels. Range: 64 to 4096, step 32
​
overlap
int
default:"64"
Overlap between tiles. Range: 0 to 4096, step 32
​
temporal_size
int
default:"64"
Frames per tile (video VAEs)
​
temporal_overlap
int
default:"8"
Frame overlap (video VAEs)
Returns:
  • IMAGE: The decoded image

​
VAEEncodeForInpaint

Encodes an image with a mask for inpainting. Category: latent/inpaint
​
pixels
IMAGE
required
The image to encode
​
vae
VAE
required
The VAE model
​
mask
MASK
required
The inpainting mask
​
grow_mask_by
int
default:"6"
Pixels to expand the mask by for seamless blending. Range: 0 to 64
Returns:
  • LATENT: Encoded latent with noise mask

​
Latent Transformations

​
LatentUpscale

Upscales a latent image to a specific size. Category: latent
​
samples
LATENT
required
The latent to upscale
​
upscale_method
string
required
Interpolation method. Options: nearest-exact, bilinear, area, bicubic, bislerp
​
width
int
default:"512"
Target width in pixels. Range: 0 to 16384, step 8
​
height
int
default:"512"
Target height in pixels. Range: 0 to 16384, step 8
​
crop
string
required
Crop method. Options: disabled, center
Returns:
  • LATENT: Upscaled latent
Description: Upscales in latent space before decoding. More efficient than upscaling pixels. Use with KSampler at low denoise (0.4-0.6) for best results.

​
LatentUpscaleBy

Upscales a latent by a scale factor. Category: latent
​
samples
LATENT
required
The latent to upscale
​
upscale_method
string
required
Interpolation method
​
scale_by
float
default:"1.5"
Scale multiplier. Range: 0.01 to 8.0
Returns:
  • LATENT: Upscaled latent

​
LatentRotate

Rotates a latent image in 90-degree increments. Category: latent/transform
​
samples
LATENT
required
The latent to rotate
​
rotation
string
required
Rotation amount. Options: none, 90 degrees, 180 degrees, 270 degrees
Returns:
  • LATENT: Rotated latent

​
LatentFlip

Flips a latent image horizontally or vertically. Category: latent/transform
​
samples
LATENT
required
The latent to flip
​
flip_method
string
required
Flip direction. Options: x-axis: vertically, y-axis: horizontally
Returns:
  • LATENT: Flipped latent

​
LatentCrop

Crops a latent image to a specific region. Category: latent/transform
​
samples
LATENT
required
The latent to crop
​
width
int
default:"512"
Crop width. Range: 64 to 16384, step 8
​
height
int
default:"512"
Crop height. Range: 64 to 16384, step 8
​
x
int
default:"0"
X offset. Range: 0 to 16384, step 8
​
y
int
default:"0"
Y offset. Range: 0 to 16384, step 8
Returns:
  • LATENT: Cropped latent

​
Latent Composition

​
LatentComposite

Composites one latent onto another. Category: latent
​
samples_to
LATENT
required
The destination latent (background)
​
samples_from
LATENT
required
The source latent to composite
​
x
int
default:"0"
X position. Range: 0 to 16384, step 8
​
y
int
default:"0"
Y position. Range: 0 to 16384, step 8
​
feather
int
default:"0"
Feathering amount for blending. Range: 0 to 16384, step 8
Returns:
  • LATENT: Composited latent
Description: Pastes one latent onto another at a specific position. Feathering creates smooth transitions at edges.

​
LatentBlend

Blends two latents together. Category: _for_testing
​
samples1
LATENT
required
First latent
​
samples2
LATENT
required
Second latent
​
blend_factor
float
default:"0.5"
Blend amount. 0.0 = only samples1, 1.0 = only samples2. Range: 0.0 to 1.0
Returns:
  • LATENT: Blended latent

​
Latent Batch Operations

​
LatentFromBatch

Extracts a subset from a latent batch. Category: latent/batch
​
samples
LATENT
required
The batch to extract from
​
batch_index
int
default:"0"
Starting index. Range: 0 to 63
​
length
int
default:"1"
Number of items to extract. Range: 1 to 64
Returns:
  • LATENT: Extracted latent subset

​
RepeatLatentBatch

Duplicates a latent batch. Category: latent/batch
​
samples
LATENT
required
The latent to repeat
​
amount
int
default:"1"
Number of times to repeat. Range: 1 to 64
Returns:
  • LATENT: Repeated latent batch

​
Advanced Latent Operations

​
LatentAdd

Adds two latents together. Category: latent/advanced
​
samples1
LATENT
required
First latent
​
samples2
LATENT
required
Second latent
Returns:
  • LATENT: Sum of the two latents

​
LatentSubtract

Subtracts one latent from another. Category: latent/advanced
​
samples1
LATENT
required
Latent to subtract from
​
samples2
LATENT
required
Latent to subtract
Returns:
  • LATENT: Difference between latents

​
LatentMultiply

Multiplies a latent by a scalar value. Category: latent/advanced
​
samples
LATENT
required
The latent to multiply
​
multiplier
float
default:"1.0"
Scale factor. Range: -10.0 to 10.0
Returns:
  • LATENT: Scaled latent

​
LatentInterpolate

Interpolates between two latents. Category: latent/advanced
​
samples1
LATENT
required
First latent
​
samples2
LATENT
required
Second latent
​
ratio
float
default:"1.0"
Interpolation ratio. Range: 0.0 to 1.0
Returns:
  • LATENT: Interpolated latent
Description: Creates a smooth blend between two latents using spherical interpolation.

​
Inpainting

​
SetLatentNoiseMask

Adds a noise mask to a latent for inpainting. Category: latent/inpaint
​
samples
LATENT
required
The latent to add mask to
​
mask
MASK
required
The noise mask
Returns:
  • LATENT: Latent with noise mask
Description: The noise mask tells the sampler which areas to denoise. White areas are regenerated, black areas are preserved.

​
Latent Storage

​
SaveLatent

Saves a latent to disk. Category: _for_testing
​
samples
LATENT
required
The latent to save
​
filename_prefix
string
default:"latents/ComfyUI"
Filename prefix for saved latent
Description: Saves latents as .latent files that can be loaded later.

​
LoadLatent

Loads a previously saved latent from disk. Category: _for_testing
​
latent
string
required
The latent file to load
Returns:
  • LATENT: The loaded latent

​
Workflow Tips

​
Latent Upscaling Workflow

  1. Generate base image at 512×512
  2. Use LatentUpscale or LatentUpscaleBy to upscale latent to 1024×1024
  3. Use KSampler with denoise=0.4-0.6 to add high-frequency details
  4. Decode with VAEDecode
This is faster and uses less VRAM than generating at high resolution directly.

​
Latent Space Benefits

  • 8x compression: A 512×512 pixel image is 64×64 in latent space
  • Faster operations: Transforms are 64x faster than in pixel space
  • Lower memory: Operations use significantly less VRAM
  • Native space: Diffusion models work directly in latent space

Build docs developers (and LLMs) love

Get started for freeTalk to us