ComfyUI Node Reference — Integrations

This reference documents the most commonly used ComfyUI nodes for image generation workflows in the Oshun platform.

Table of Contents#

Core Nodes — checkpoint loading and the KSampler samplers
Loading Nodes — LoRA, ControlNet, VAE, CLIP, and upscale-model loaders
Conditioning Nodes — text encoding and ControlNet application
Latent Nodes
VAE Nodes
Image Nodes
Upscaling Nodes
Utility Nodes
Custom Manager Nodes

Core Nodes#

CheckpointLoaderSimple#

Loads a Stable Diffusion checkpoint file.

json

{
  "class_type": "CheckpointLoaderSimple",
  "inputs": {
    "ckpt_name": "sd_xl_base_1.0.safetensors"
  }
}

Input	Type	Description
`ckpt_name`	string	Checkpoint filename from models/checkpoints/

Output	Index	Type	Description
MODEL	0	MODEL	Diffusion model
CLIP	1	CLIP	Text encoder
VAE	2	VAE	Variational autoencoder

Supported Checkpoints:

Model	VRAM	Resolution	Notes
sd_xl_base_1.0	8GB	1024x1024	SDXL base model
sd_xl_refiner_1.0	8GB	1024x1024	SDXL refiner
sd_v1-5	4GB	512x512	SD 1.5
sd_v2-1	5GB	768x768	SD 2.1
flux_dev	24GB	Variable	Flux.1 Dev

KSampler#

The primary sampling node for generating images.

json

{
  "class_type": "KSampler",
  "inputs": {
    "seed": 12345,
    "steps": 30,
    "cfg": 7.5,
    "sampler_name": "euler_ancestral",
    "scheduler": "normal",
    "denoise": 1.0,
    "model": ["1", 0],
    "positive": ["2", 0],
    "negative": ["3", 0],
    "latent_image": ["4", 0]
  }
}

Input	Type	Description
`seed`	int	Random seed for reproducibility
`steps`	int	Number of sampling steps (1-150)
`cfg`	float	Classifier-free guidance scale (1-30)
`sampler_name`	string	Sampling algorithm
`scheduler`	string	Noise schedule
`denoise`	float	Denoising strength (0.0-1.0)
`model`	MODEL	Diffusion model connection
`positive`	CONDITIONING	Positive prompt conditioning
`negative`	CONDITIONING	Negative prompt conditioning
`latent_image`	LATENT	Input latent (empty or encoded)

Output	Index	Type	Description
LATENT	0	LATENT	Sampled latent

Sampler Options:

Sampler	Speed	Quality	Notes
`euler`	Fast	Good	Basic Euler method
`euler_ancestral`	Fast	Good	Adds noise, more variation
`heun`	Medium	Better	2nd order method
`dpm_2`	Medium	Better	DPM-Solver 2nd order
`dpm_2_ancestral`	Medium	Better	DPM-2 with noise
`dpmpp_2m`	Fast	Better	DPM++ 2M
`dpmpp_2m_sde`	Medium	Best	DPM++ 2M with SDE
`dpmpp_3m_sde`	Slow	Best	DPM++ 3M with SDE
`ddim`	Fast	Good	Deterministic
`uni_pc`	Fast	Better	UniPC sampler

Scheduler Options:

Scheduler	Description
`normal`	Linear noise schedule
`karras`	Karras noise schedule (recommended)
`exponential`	Exponential schedule
`sgm_uniform`	SGM uniform schedule
`simple`	Simple schedule
`ddim_uniform`	DDIM uniform

KSamplerAdvanced#

Advanced sampler with additional control.

json

{
  "class_type": "KSamplerAdvanced",
  "inputs": {
    "add_noise": "enable",
    "noise_seed": 12345,
    "steps": 30,
    "cfg": 7.5,
    "sampler_name": "dpmpp_2m_sde",
    "scheduler": "karras",
    "start_at_step": 0,
    "end_at_step": 30,
    "return_with_leftover_noise": "disable",
    "model": ["1", 0],
    "positive": ["2", 0],
    "negative": ["3", 0],
    "latent_image": ["4", 0]
  }
}

Additional Inputs	Type	Description
`add_noise`	enable/disable	Whether to add initial noise
`noise_seed`	int	Separate seed for noise
`start_at_step`	int	Starting step (for img2img)
`end_at_step`	int	Ending step
`return_with_leftover_noise`	enable/disable	Keep noise for chaining

Loading Nodes#

LoraLoader#

Loads and applies LoRA weights to model and CLIP.

json

{
  "class_type": "LoraLoader",
  "inputs": {
    "model": ["1", 0],
    "clip": ["1", 1],
    "lora_name": "my_lora.safetensors",
    "strength_model": 0.8,
    "strength_clip": 0.8
  }
}

Input	Type	Description
`model`	MODEL	Input model
`clip`	CLIP	Input CLIP
`lora_name`	string	LoRA filename
`strength_model`	float	Model weight strength (-2.0 to 2.0)
`strength_clip`	float	CLIP weight strength (-2.0 to 2.0)

Output	Index	Type	Description
MODEL	0	MODEL	LoRA-enhanced model
CLIP	1	CLIP	LoRA-enhanced CLIP

Stacking LoRAs:

json

{
  "1": {
    "class_type": "CheckpointLoaderSimple",
    "inputs": { "ckpt_name": "sd_xl_base_1.0.safetensors" }
  },
  "2": {
    "class_type": "LoraLoader",
    "inputs": {
      "model": ["1", 0],
      "clip": ["1", 1],
      "lora_name": "style_lora.safetensors",
      "strength_model": 0.7,
      "strength_clip": 0.7
    }
  },
  "3": {
    "class_type": "LoraLoader",
    "inputs": {
      "model": ["2", 0],
      "clip": ["2", 1],
      "lora_name": "detail_lora.safetensors",
      "strength_model": 0.5,
      "strength_clip": 0.5
    }
  }
}

ControlNetLoader#

Loads a ControlNet model.

json

{
  "class_type": "ControlNetLoader",
  "inputs": {
    "control_net_name": "controlnet-canny-sdxl-1.0.safetensors"
  }
}

Input	Type	Description
`control_net_name`	string	ControlNet filename

Output	Index	Type	Description
CONTROL_NET	0	CONTROL_NET	ControlNet model

Common ControlNets:

ControlNet	Type	Description
controlnet-canny	Edge	Canny edge detection
controlnet-depth	Depth	Depth estimation
controlnet-openpose	Pose	Human pose
controlnet-scribble	Sketch	Hand-drawn sketches
controlnet-softedge	Edge	Soft edge detection
controlnet-lineart	Line	Line art
controlnet-tile	Tile	Tile/upscale
controlnet-inpaint	Inpaint	Inpainting

VAELoader#

Loads a standalone VAE.

json

{
  "class_type": "VAELoader",
  "inputs": {
    "vae_name": "sdxl_vae.safetensors"
  }
}

Input	Type	Description
`vae_name`	string	VAE filename

Output	Index	Type	Description
VAE	0	VAE	Variational autoencoder

CLIPLoader#

Loads a standalone CLIP model.

json

{
  "class_type": "CLIPLoader",
  "inputs": {
    "clip_name": "clip_l.safetensors",
    "type": "stable_diffusion"
  }
}

UpscaleModelLoader#

Loads an upscaling model.

json

{
  "class_type": "UpscaleModelLoader",
  "inputs": {
    "model_name": "RealESRGAN_x4plus.pth"
  }
}

Output	Index	Type	Description
UPSCALE_MODEL	0	UPSCALE_MODEL	Upscaler model

Common Upscalers:

Upscaler	Scale	Type	Notes
RealESRGAN_x4plus	4x	General	Good all-around
RealESRGAN_x4plus_anime	4x	Anime	Anime/illustration
4x_NMKD-Siax	4x	Photo	Photorealistic
4x-UltraSharp	4x	Photo	Sharp details
8x_NMKD-Superscale	8x	General	Extreme upscale

Conditioning Nodes#

CLIPTextEncode#

Encodes text prompts to CLIP embeddings.

json

{
  "class_type": "CLIPTextEncode",
  "inputs": {
    "text": "A beautiful sunset over mountains, highly detailed, 8k",
    "clip": ["1", 1]
  }
}

Input	Type	Description
`text`	string	Prompt text
`clip`	CLIP	CLIP model

Output	Index	Type	Description
CONDITIONING	0	CONDITIONING	Text embeddings

Prompt Tips:

Use commas to separate concepts
Add quality boosters: "highly detailed, 8k, masterpiece"
Specify style: "digital art, oil painting, photograph"
Add negative prompts in a separate node

CLIPTextEncodeSDXL#

SDXL-specific text encoder with dual prompts.

json

{
  "class_type": "CLIPTextEncodeSDXL",
  "inputs": {
    "text_g": "A beautiful sunset over mountains",
    "text_l": "detailed landscape, golden hour lighting",
    "clip": ["1", 1],
    "width": 1024,
    "height": 1024,
    "crop_w": 0,
    "crop_h": 0,
    "target_width": 1024,
    "target_height": 1024
  }
}

Input	Type	Description
`text_g`	string	Global/main prompt (CLIP-G)
`text_l`	string	Local/detail prompt (CLIP-L)
`width/height`	int	Original image dimensions
`crop_w/crop_h`	int	Crop offset
`target_width/target_height`	int	Target dimensions

ConditioningCombine#

Combines multiple conditioning inputs.

json

{
  "class_type": "ConditioningCombine",
  "inputs": {
    "cond_1": ["2", 0],
    "cond_2": ["3", 0]
  }
}

ConditioningSetArea#

Sets conditioning to a specific area (for regional prompts).

json

{
  "class_type": "ConditioningSetArea",
  "inputs": {
    "conditioning": ["2", 0],
    "width": 512,
    "height": 512,
    "x": 0,
    "y": 0,
    "strength": 1.0
  }
}

ControlNetApply / ControlNetApplyAdvanced#

Applies ControlNet conditioning.

json

{
  "class_type": "ControlNetApply",
  "inputs": {
    "conditioning": ["2", 0],
    "control_net": ["5", 0],
    "image": ["6", 0],
    "strength": 0.8
  }
}

json

{
  "class_type": "ControlNetApplyAdvanced",
  "inputs": {
    "positive": ["2", 0],
    "negative": ["3", 0],
    "control_net": ["5", 0],
    "image": ["6", 0],
    "strength": 0.8,
    "start_percent": 0.0,
    "end_percent": 1.0
  }
}

Input	Type	Description
`strength`	float	ControlNet influence (0.0-2.0)
`start_percent`	float	When to start applying (0.0-1.0)
`end_percent`	float	When to stop applying (0.0-1.0)

Latent Nodes#

EmptyLatentImage#

Creates an empty latent for txt2img.

json

{
  "class_type": "EmptyLatentImage",
  "inputs": {
    "width": 1024,
    "height": 1024,
    "batch_size": 1
  }
}

Input	Type	Description
`width`	int	Image width (must be divisible by 8)
`height`	int	Image height (must be divisible by 8)
`batch_size`	int	Number of images to generate

Resolution Guidelines:

Model	Optimal	Minimum	Maximum
SD 1.5	512x512	256x256	768x768
SD 2.1	768x768	512x512	1024x1024
SDXL	1024x1024	768x768	2048x2048
Flux	1024x1024	512x512	2048x2048

LatentUpscale#

Upscales latent images.

json

{
  "class_type": "LatentUpscale",
  "inputs": {
    "samples": ["5", 0],
    "upscale_method": "nearest-exact",
    "width": 2048,
    "height": 2048,
    "crop": "disabled"
  }
}

Input	Type	Description
`upscale_method`	string	nearest-exact, bilinear, area, bicubic, bislerp
`width/height`	int	Target dimensions
`crop`	string	disabled, center

LatentComposite#

Composites two latent images.

json

{
  "class_type": "LatentComposite",
  "inputs": {
    "samples_to": ["5", 0],
    "samples_from": ["6", 0],
    "x": 0,
    "y": 0,
    "feather": 64
  }
}

SetLatentNoiseMask#

Sets a mask for partial denoising (inpainting).

json

{
  "class_type": "SetLatentNoiseMask",
  "inputs": {
    "samples": ["5", 0],
    "mask": ["10", 0]
  }
}

VAE Nodes#

VAEDecode#

Decodes latents to RGB images.

json

{
  "class_type": "VAEDecode",
  "inputs": {
    "samples": ["5", 0],
    "vae": ["1", 2]
  }
}

Input	Type	Description
`samples`	LATENT	Latent to decode
`vae`	VAE	VAE model

Output	Index	Type	Description
IMAGE	0	IMAGE	Decoded RGB image

VAEEncode#

Encodes RGB images to latents.

json

{
  "class_type": "VAEEncode",
  "inputs": {
    "pixels": ["8", 0],
    "vae": ["1", 2]
  }
}

Input	Type	Description
`pixels`	IMAGE	RGB image
`vae`	VAE	VAE model

Output	Index	Type	Description
LATENT	0	LATENT	Encoded latent

VAEEncodeForInpaint#

Encodes image with mask for inpainting.

json

{
  "class_type": "VAEEncodeForInpaint",
  "inputs": {
    "pixels": ["8", 0],
    "vae": ["1", 2],
    "mask": ["9", 0],
    "grow_mask_by": 6
  }
}

Input	Type	Description
`mask`	MASK	Inpainting mask
`grow_mask_by`	int	Expand mask by pixels

Image Nodes#

LoadImage#

Loads an image from disk or URL.

json

{
  "class_type": "LoadImage",
  "inputs": {
    "image": "input_image.png"
  }
}

Output	Index	Type	Description
IMAGE	0	IMAGE	Loaded image
MASK	1	MASK	Alpha channel as mask

SaveImage#

Saves images to disk.

json

{
  "class_type": "SaveImage",
  "inputs": {
    "filename_prefix": "ComfyUI",
    "images": ["7", 0]
  }
}

Input	Type	Description
`filename_prefix`	string	Output filename prefix
`images`	IMAGE	Images to save

PreviewImage#

Previews images without saving.

json

{
  "class_type": "PreviewImage",
  "inputs": {
    "images": ["7", 0]
  }
}

ImageScale#

Scales images.

json

{
  "class_type": "ImageScale",
  "inputs": {
    "image": ["8", 0],
    "upscale_method": "bicubic",
    "width": 2048,
    "height": 2048,
    "crop": "disabled"
  }
}

ImageInvert#

Inverts image colors.

json

{
  "class_type": "ImageInvert",
  "inputs": {
    "image": ["8", 0]
  }
}

ImageBlend#

Blends two images.

json

{
  "class_type": "ImageBlend",
  "inputs": {
    "image1": ["8", 0],
    "image2": ["9", 0],
    "blend_factor": 0.5,
    "blend_mode": "normal"
  }
}

Input	Type	Description
`blend_factor`	float	Blend amount (0.0-1.0)
`blend_mode`	string	normal, multiply, screen, overlay, soft_light

Upscaling Nodes#

ImageUpscaleWithModel#

Upscales using an AI model.

json

{
  "class_type": "ImageUpscaleWithModel",
  "inputs": {
    "upscale_model": ["10", 0],
    "image": ["8", 0]
  }
}

Input	Type	Description
`upscale_model`	UPSCALE_MODEL	Loaded upscaler
`image`	IMAGE	Image to upscale

UltimateSDUpscale#

Advanced upscaling with tiled sampling.

json

{
  "class_type": "UltimateSDUpscale",
  "inputs": {
    "image": ["8", 0],
    "model": ["1", 0],
    "positive": ["2", 0],
    "negative": ["3", 0],
    "vae": ["1", 2],
    "upscale_by": 2,
    "seed": 12345,
    "steps": 20,
    "cfg": 7.5,
    "sampler_name": "euler_ancestral",
    "scheduler": "normal",
    "denoise": 0.3,
    "tile_width": 512,
    "tile_height": 512,
    "mask_blur": 8,
    "tile_padding": 32
  }
}

Utility Nodes#

PrimitiveNode#

Provides a constant value.

json

{
  "class_type": "PrimitiveNode",
  "inputs": {},
  "widgets_values": [12345]
}

Reroute#

Organizes workflow connections.

json

{
  "class_type": "Reroute",
  "inputs": {
    "": ["5", 0]
  }
}

Note#

Adds documentation to workflows (no execution).

json

{
  "class_type": "Note",
  "inputs": {},
  "widgets_values": ["This is a workflow note"]
}

Custom Manager Nodes#

These nodes are from popular custom node packs.

Efficient Loader (Efficiency Nodes)#

Combined checkpoint + VAE + LoRA loading.

json

{
  "class_type": "Efficient Loader",
  "inputs": {
    "ckpt_name": "sd_xl_base_1.0.safetensors",
    "vae_name": "Baked VAE",
    "clip_skip": -1,
    "lora_name": "None",
    "lora_model_strength": 1,
    "lora_clip_strength": 1,
    "positive": "A beautiful landscape",
    "negative": "blurry, low quality",
    "batch_size": 1
  }
}

KSampler (Efficient)#

Optimized sampler with built-in preview.

json

{
  "class_type": "KSampler (Efficient)",
  "inputs": {
    "seed": 12345,
    "steps": 30,
    "cfg": 7.5,
    "sampler_name": "euler_ancestral",
    "scheduler": "normal",
    "denoise": 1,
    "preview_method": "auto",
    "model": ["1", 0],
    "positive": ["1", 1],
    "negative": ["1", 2],
    "latent_image": ["1", 3],
    "optional_vae": ["1", 4]
  }
}

IPAdapter#

IP-Adapter for image-based conditioning.

json

{
  "class_type": "IPAdapterApply",
  "inputs": {
    "model": ["1", 0],
    "ipadapter": ["10", 0],
    "image": ["11", 0],
    "weight": 0.8,
    "noise": 0.0,
    "weight_type": "standard"
  }
}

FaceDetailer (Impact Pack)#

Automatic face detection and enhancement.

json

{
  "class_type": "FaceDetailer",
  "inputs": {
    "image": ["8", 0],
    "model": ["1", 0],
    "clip": ["1", 1],
    "vae": ["1", 2],
    "positive": ["2", 0],
    "negative": ["3", 0],
    "bbox_detector": ["20", 0],
    "sam_model": ["21", 0],
    "guide_size": 384,
    "guide_size_for": "bbox",
    "max_size": 1024,
    "seed": 12345,
    "steps": 20,
    "cfg": 7.5,
    "sampler_name": "euler_ancestral",
    "scheduler": "normal",
    "denoise": 0.4
  }
}

Table of Contents#

Core Nodes#

CheckpointLoaderSimple#

KSampler#

KSamplerAdvanced#

Loading Nodes#

LoraLoader#

ControlNetLoader#

VAELoader#

CLIPLoader#

UpscaleModelLoader#

Conditioning Nodes#

CLIPTextEncode#

CLIPTextEncodeSDXL#

ConditioningCombine#

ConditioningSetArea#

ControlNetApply / ControlNetApplyAdvanced#

Latent Nodes#

EmptyLatentImage#

LatentUpscale#

LatentComposite#

SetLatentNoiseMask#

VAE Nodes#

VAEDecode#

VAEEncode#

VAEEncodeForInpaint#

Image Nodes#

LoadImage#

SaveImage#

PreviewImage#

ImageScale#

ImageInvert#

ImageBlend#

Upscaling Nodes#

ImageUpscaleWithModel#

UltimateSDUpscale#

Utility Nodes#

PrimitiveNode#

Reroute#

Note#

Custom Manager Nodes#

Efficient Loader (Efficiency Nodes)#

KSampler (Efficient)#

IPAdapter#

FaceDetailer (Impact Pack)#

Related Documentation#