AI models used to generate or modify images

Usually a pre-trained diffusion model

positive cue word

Describe what you want to see in the generated image

For example: "An orange cat sitting on the grass with the sun shining brightly"

negative cue words

Describe what you don't want to see in the generated image

For example: "blurred, low-quality, distorted features"

potential image

An intermediate representation, not a directly visible image, but an encoded form

Possibly from previous processing steps or random initialization

seed value

Used to initialize the random number generator

The same results can be reproduced using the same seed value

Allows additional control or modification of results after generation

Specific functions may vary depending on the version of ComfyUI (usually modifying the seed value)

Number of sampling steps

Determine the number of iterations of the diffusion process

More steps usually produce finer details but increase calculation time

No classifier guidance scale

Controls how much the generation process respects prompt words

Higher values ​​will make the results closer to the prompt word, but may reduce creativity

sampler name

Specifies the algorithm used to generate images from noise

Common ones include Euler, Euler a, DDIM, etc.

Scheduler

Decide how to adjust the noise level during sampling

Different schedulers may affect build speed and quality

Denoise intensity

Control how much the input latent image is modified

When the value is 1, it is completely regenerated. The smaller the value, the more features of the original image are retained.

refers to the latent space representation of the output

Is an encoded form of an image that can be used for further processing or conversion into a visual image

LatentUpscale

LatentComposite

LatentRotate

LatentFlip

LatentCrop

VAEDecode

Savelmage

LatentMixRepeat

ControlNetApply

LatentBlend

EmptyLatentlmage

VAEEncode

KSampler(previous)

LatentComposite

LatentUpscale

ControlNetApply

LatentRotate/LatentFlip

LatentBlend

Dimension consistency: ensure that the input latent image dimensions are compatible with other settings of KSampler (such as the model)

Denoising strength: When using a non-empty latent image, adjusting the denoising strength can control the degree to which the original image features are retained.

Batch processing: Some nodes can output batches of latent images, suitable for generating multiple variants

Simple yet effective benchmark sampler, fast

May be less detailed than more complex samplers

Euler Ancestral

An improved version of Euler

Typically produces richer detail while maintaining speed

Denoising Diffusion Implicit Models

Deterministic sampler

Produce stable and consistent results

Typically faster than other samplers while maintaining image quality

Includes variants such as DPM 2M and DPM SDE

Typically produces high-quality results, especially when working with intricate details

Latent Consistency Model

Newer sampler designed for fast inference

Produce high-quality images in fewer sampling steps

Particularly suitable for real-time or near-real-time applications

High efficiency sampler

Produce high-quality results in fewer steps

Strike a good balance between speed and quality

One of the basic samplers for diffusion models

May not be as fast as some newer samplers

Typically produces stable and high-quality results

Description: Standard scheduler, suitable for most situations.

Features: Consistent noise reduction rate throughout the entire sampling process.

Applicable scenarios: General purpose, suitable for most model and generation tasks.

Description: An optimized scheduler based on the work of Karras et al.

Features: Dynamically adjusts the noise reduction rate at different stages of the sampling process, often producing higher quality results.

Applicable scenarios: When pursuing higher quality output, especially in high-resolution or detailed image generation.

Description: Use an exponential function to schedule noise reduction.

Features: More noise reduction in early steps, less noise in later steps.

Suitable for: When you want to quickly establish image structure in the early stages of the generation process.

Description: A simplified scheduling method.

Features: Linear noise reduction, light computational burden.

Applicable scenarios: When rapid generation is required or when computing resources are limited.

Description: Uniform sampling based on the Denoising Diffusion Implicit Model (DDIM).

Features: Provides a more consistent sampling process, potentially leading to more stable results.

Applicable scenarios: When more predictable and consistent generation results are needed.