Latent Tools for ComfyUI

A collection of nodes for manipulating latent tensors in ComfyUI. These tools provide various operations for working with latent representations in stable diffusion workflows.

Made in collaboration with fzayguler

Installation

1. Clone this repository into your ComfyUI/custom_nodes/ directory
2. Restart ComfyUI

Or use the ComfyUI extension manager.

Available Nodes

Preview and Debug

LTPreviewLatent

Visualizes latent tensors for debugging and inspection.

KSampler with additional noise input

LTKSampler

A KSampler variant that accepts an additional input for starting latent space noise.

| <img src="assets/KSampler.png" alt="LTKSampler" width="70%"> | |------------| | Inputs | | - model: The model used for denoising | | - extra_seed: See for any other noise used by the sampler | | - steps: Number of steps in the denoising process | | - cfg: Classifier-Free Guidance scale | | - sampler_name: Algorithm used for sampling | | - scheduler: Controls how noise is gradually removed | | - positive: Positive conditioning | | - negative: Negative conditioning | | - latent_image: The latent image to denoise | | - latent_noise: Starting noise for the sampler | | - denoise: Amount of denoising to apply | | Outputs | | - latent: The denoised latent tensor |

LTGaussianLatent

Generates latent tensors filled with random values from a normal (Gaussian) distribution.

| <img src="assets/GaussianLatent.png" alt="LTGaussianLatent" width="30%"> | |------------| | Inputs | | - channels: Number of channels (default: 4) | | - width: Width of the latent space (will be divided by 8) | | - height: Height of the latent space (will be divided by 8) | | - batch_size: Number of samples to generate | | - mean: Mean of the normal distribution | | - std: Standard deviation of the normal distribution | | - seed: Random seed | | Outputs | | - latent: Generated latent tensor | | |

Example:

| "quick brown fox", σ=0.9 μ=0 | "quick brown fox", σ=1.05, μ=0 | "quick brown fox", σ=1, μ=0 | "quick brown fox", σ=1, μ=-0.1 | "quick brown fox", σ=1, μ=0.1 | |---|---|---|---|---| | | | | |

LTUniformLatent

Generates latent tensors with values uniformly distributed between min and max.

| <img src="assets/UniformLatent.png" alt="LTUniformLatent" width="30%"> | |------------| | Inputs | | - channels: Number of channels (default: 4) | | - width: Width of the latent space (will be divided by 8) | | - height: Height of the latent space (will be divided by 8) | | - batch_size: Number of samples to generate | | - min: Minimum value | | - max: Maximum value | | - seed: Random seed | | Outputs | | - latent: Generated latent tensor | | |

Note: Stable Diffusion models are usually trained with Gaussian noise, so the generations from Uniform noise will look unusual.

Example:

| "quick brown fox", -1.67 to 1.67 | "quick brown fox", -1.81 to 1.81 | |---|---| | |

Latent Operations

LTBlendLatent

Blends two latent tensors using various blending modes.

| <img src="assets/BlendLatent.png" alt="LTBlendLatent" width="40%"> | |------------| | Inputs | | - latent1: First latent tensor | | - latent2: Second latent tensor | | - mode: Blending mode | | - ratio: Blend ratio (0.0 to 1.0) Only used for mode=sample or mode=interpolate | | - seed: Random seed Only used for mode=sample | | Outputs | | - latent: Blended latent tensor |

| Blending Modes | Description | |-------------------|-----------------| | interpolate | Linear interpolation between latents | | add | Additive blending | | multiply | Multiplicative blending | | abs_max | Maximum of absolute values | | abs_min | Minimum of absolute values | | max | Element-wise maximum | | min | Element-wise minimum | | sample | Randomly sample from either latent based on ratio |

Example:
Inputs: Random Gaussian σ=0.1 μ=0 (top) and Random Uniform [-1s, 1] (bottom)
Blend modes: interpolate (top) and sample (bottom)

LTLatentOp

Applies mathematical operations to a latent tensor.

| <img src="assets/LatentOp-ops.png" alt="LTLatentOp" width="80%"> | |------------| | Inputs | | - latent: Input latent tensor | | - op: Operation to apply | | - arg: Argument to apply (for operations that require an argument) | | Outputs | | - latent: Resulting latent tensor |

| Operation | Description | |---------------|-----------------| | add | Add a value | | mul | Multiply by a value | | pow | Raise to a power | | exp | Exponential | | abs | Absolute value | | clamp_bottom | Clamp minimum value | | clamp_top | Clamp maximum value | | norm | Normalize (zero mean, unit variance) | | mean | Set mean to specified value | | std | Set standard deviation to specified value | | sigmoid | Apply sigmoid function | | nop | No operation |

Example: Inputs: Random Gaussian σ=1 μ=0 Op: abs

LTLatentsConcatenate

Concatenates two latent tensors along a specified dimension.

| <img src="assets/LatentsConcatenate.png" alt="LTLatentsConcatenate" width="50%"> | |------------| | Inputs | | - latent1: First latent tensor | | - latent2: Second latent tensor | | - dim: Dimension to concatenate along (supports negative indexing) | | Outputs | | - latent: Concatenated latent tensor |

Example1: 2 images, concatenated along x axis:

Example2: Stable Video Diffusion xt (24 frames total), concatenating

• 10 frames Gaussian noise (σ=1 μ=0)
• 4 frames Gaussian noise (σ=1.2 μ=0)
• 10 frames Gaussian noise (σ=1 μ=0)

| 10f (σ=1 μ=0) + 4f (σ=1.2 μ=0) + 10f (σ=1 μ=0) | 24f (σ=1 μ=0) | |---|---| | | |

LTLatentToShape

Extracts the shape of a latent tensor.

| <img src="assets/LatentToShape.png" alt="LTLatentToShape" width="40%"> | |------------| | Inputs | | - input: Input latent tensor | | Outputs: | | - Return 7 dimensions of the input latent shape. Non-existing ones are returned as 0 |

LTReshapeLatent

Reshapes a latent tensor to new dimensions.

| <img src="assets/ReshapeLatent.png" alt="LTReshapeLatent" width="50%"> | |------------| | Inputs | | - input: Input latent tensor | | - strict: If True, requires exact size match | | - dim0-dim6: Target dimensions (0 values are ignored) | | Outputs | | - latent: Reshaped latent tensor |

Example: Reshape one latent to match another one:

<img src="assets/ShapeExample.png" alt="LTReshapeLatent" width="70%"> <!--  -->

Batch helpers

Parameter Randomization

LTNumberRangeUniform and LTRandomRangeGaussian are used to randomize inputs to other nodes when scheduling multiple images.

LTNumberRangeUniform

Generates random float values from a uniform distribution.

| <img src="assets/NumberRangeUniform.png" alt="LTNumberRangeUniform" width="50%"> | |------------| | Inputs | | - min: Minimum value | | - max: Maximum value | | - seed: Random seed - randomize it to get different values for each image | | Outputs | | - float: Generated number as float | | - int: Generated number as int |

LTNumberRangeGaussian

Generates random values from a Gaussian distribution.

| <img src="assets/NumberRangeGaussian.png" alt="LTRandomRangeGaussian" width="50%"> | |------------| | Inputs | | - mean: Mean of the normal distribution | | - std: Standard deviation of the normal distribution | | - seed: Random seed - randomize it to get different values for each image | | Outputs | | - float: Generated number as float | | - int: Generated number as int |

Example:

Let's randomize a bumch of parameters for batch generation:

• For the Gaussian Noise that is used as input:
  • the mean value will be itself a random uniform value between -0.2 and 0.2.
  • The standard deviation will be itself a random value with mean 1 and std 0.5, clipped to 0.1 at the low end (using a ComfyMath node).
• The number of diffusion steps is random uniform value betwee 5 and 40.
• The cfg is random gussian with mean 8 and std 1, clipped to 0.1 at the low end (using a ComfyMath node).

Using this setup, generate 100 images with different parameters.

| | | |---|---| | | <img src="assets/random_params.gif" alt="random_params" width="100%"> |

LTFloat_Steps_0001

LTFloat_Steps_0001

LTFloat_Steps_0002

LTFloat_Steps_0005

LTFloat_Steps_001

LTFloat_Steps_002

LTFloat_Steps_005

LTFloat_Steps_01

LTFloat_Steps_02

LTFloat_Steps_05

LTFloat_Steps_1

These nodes are used to increment/decrement a float value by a fixed amount when scheduling multiple images.

| | |------------| | Inputs | | - value: float | | Outputs | | - float: float | | - string: string |

Example:

Generate images with fixed seed Gaussian noise, starting with σ etween 0.8 and 1.2 with 0.001 step increments:

Generate images with fixed seed Gaussian noise, starting with μ between -0.22 and 0.3 with 0.001 step increments:

Result: | Sweeping Standard Deviation (σ) | Sweeping Mean (μ) | |---|---| | | |