Comfyui-QwenEditUtils
A collection of utility nodes for Qwen-based image editing in ComfyUI.
Note: This repository has been succeeded by ComfyUI-EditUtils, which provides enhanced capabilities with support for multiple AI models including Qwen and Flux2Klein. We recommend using the new repository for future projects.
Example
You can find complete ComfyUI workflow examples in the following files:
[qwen edit custom.json](qwen edit custom.json) – demonstrates advanced usage with full configuration options
[qwen edit custom mask.json](qwen edit custom mask.json) – demonstrates mask usage with full configuration options
Update Log
v2.0.5
Added mask option to configs nodes and custom node for improved image editing.
v2.0.3
Added QwenEditAdaptiveLongestEdge for easier to get the best longest edge of the image.
v2.0.0
Added TextEncodeQwenImageEditPlusCustom_lrzjason for highly customized image editing.
Added QwenEditConfigPreparer, QwenEditConfigJsonParser for creating image configurations.
Added QwenEditOutputExtractor for extracting outputs from the custom node.
Added QwenEditListExtractor for extracting items from lists.
Added CropWithPadInfo for cropping images with pad information.
Node
TextEncodeQwenImageEditPlus lrzjason
This node provides text encoding functionality with reference image support for Qwen-based image editing workflows. It allows you to encode prompts while incorporating up to 5 reference images for more controlled image generation.
Inputs
clip: The CLIP model to use for encoding
prompt: The text prompt to encode
vae (optional): The VAE model for image encoding
image1 (optional): First reference image for image editing
image2 (optional): Second reference image for image editing
image3 (optional): Third reference image for image editing
image4 (optional): Fourth reference image for image editing
image5 (optional): Fifth reference image for image editing
enable_resize (optional): Enable automatic resizing of the reference image for VAE encoding
enable_vl_resize (optional): Enable automatic resizing of the reference image for VL encoding
llama_template (optional): Custom Llama template for image description and editing instructions
Outputs
CONDITIONING: The encoded conditioning tensor
image1: The processed first reference image
image2: The processed second reference image
image3: The processed third reference image
image4: The processed fourth reference image
image5: The processed fifth reference image
LATENT: The encoded latent representation of the first reference image
Behavior
Encodes text prompts using CLIP with optional reference image guidance
Supports up to 5 reference images for complex editing tasks
Automatically resizes reference images to optimal dimensions for both VAE and VL encoding
Integrates with VAE models to encode reference images into latent space
Supports custom Llama templates for more precise image editing instructions
Processes images separately for VAE encoding (1024×1024) and VL encoding (384×384)
Returns individual processed images for more flexible workflow connections
TextEncodeQwenImageEditPlusAdvance lrzjason
This advanced node provides enhanced text encoding functionality with reference image support for Qwen-based image editing workflows. It offers more precise control over image resizing and supports flexible image input configurations with separate processing for VAE and VL encoding.
Inputs
clip: The CLIP model to use for encoding
prompt: The text prompt to encode
vae (optional): The VAE model for image encoding
vl_resize_image1 (optional): First reference image for VL encoding with resizing
vl_resize_image2 (optional): Second reference image for VL encoding with resizing
vl_resize_image3 (optional): Third reference image for VL encoding with resizing
not_resize_image1 (optional): First reference image without VL resizing
not_resize_image2 (optional): Second reference image without VL resizing
not_resize_image3 (optional): Third reference image without VL resizing
target_size (optional): Target size for VAE encoding (options: 1024, 1344, 1536, 2048, 768, 512)
target_vl_size (optional): Target size for VL encoding (default: 384)
upscale_method (optional): Method for upscaling images (options: “lanczos”, “bicubic”, “area”)
crop (optional): Cropping method (options: “center”, “disabled”)
instruction (optional): Custom instruction for image editing
Outputs
CONDITIONING: The encoded conditioning tensor
LATENT: The encoded latent representation of the first reference image
target_image1: The processed first target reference image
target_image2: The processed second target reference image
target_image3: The processed third target reference image
vl_resized_image1: The first VL-resized reference image
vl_resized_image2: The second VL-resized reference image
vl_resized_image3: The third VL-resized reference image
Behavior
Provides advanced text encoding with separate control over VAE and VL image processing
Supports 3 reference images with different resize behaviors
Offers multiple target size options for more flexible image processing
Maintains separate image outputs for VAE-encoded and VL-resized images
Provides enhanced upscale and crop controls for optimal image processing
Integrates with custom instructions for tailored image editing
TextEncodeQwenImageEditPlusPro lrzjason
This professional node provides the most flexible text encoding functionality with reference image support for Qwen-based image editing workflows. It offers fine-grained control over which images are VL-resized and includes a main image designation for focused conditioning.
Inputs
clip: The CLIP model to use for encoding
prompt: The text prompt to encode
vae (optional): The VAE model for image encoding
image1 (optional): First reference image for image editing
image2 (optional): Second reference image for image editing
image3 (optional): Third reference image for image editing
image4 (optional): Fourth reference image for image editing
image5 (optional): Fifth reference image for image editing
vl_resize_indexs (optional): Comma-separated indexes of images to apply VL resizing (default: “0,1,2”)
main_image_index (optional): Index of the main reference image for focused conditioning (default: 0)
target_size (optional): Target size for VAE encoding (options: 1024, 1344, 1536, 2048, 768, 512)
target_vl_size (optional): Target size for VL encoding (options: 384, 392)
upscale_method (optional): Method for upscaling images (options: “lanczos”, “bicubic”, “area”)
crop_method (optional): Cropping method (options: “pad”, “center”, “disabled”)
instruction (optional): Custom instruction for image editing
Outputs
CONDITIONING: The encoded conditioning tensor with all reference latents
LATENT: The encoded latent representation of the main reference image
image1: The processed first reference image
image2: The processed second reference image
image3: The processed third reference image
image4: The processed fourth reference image
image5: The processed fifth reference image
CONDITIONING: The encoded conditioning tensor with only the main reference latent
ANY: The pad information dictionary containing scaling and padding details
Behavior
Provides professional-level text encoding with maximum flexibility for image processing
Supports up to 5 reference images with configurable VL-resizing per image
Allows designation of a main reference image for focused conditioning
Offers separate conditioning outputs for full reference and main reference conditioning
Provides multiple target size options for more flexible image processing
Includes pad information for potential image cropping and scaling operations
Provides enhanced upscale and crop controls, including padding for image preservation
TextEncodeQwenImageEditPlusCustom_lrzjason
This node provides maximum flexibility for image editing workflows by allowing custom configurations for each reference image. Each image can have its own specific settings for both reference (VAE) and vision-language (VL) processing, enabling highly customized image editing scenarios.
Inputs
clip: The CLIP model to use for encoding
vae: The VAE model for image encoding
prompt: The text prompt to encode
configs: A list of configuration dictionaries for each image, containing:
image: The reference image for image editing
to_ref: Whether to include image in reference processing (VAE encoding)
ref_main_image: Whether this image is the main reference image for focused conditioning
ref_longest_edge: Target longest edge size for reference processing (default: 1024)
ref_crop: Cropping method for reference processing (options: “pad”, “center”, “disabled”)
ref_upscale: Upscaling method for reference processing (options: “lanczos”, “bicubic”, “area”)
to_vl: Whether to include image in vision-language processing
vl_resize: Whether to resize image for VL processing (default: True)
vl_target_size: Target size for VL processing (default: 384)
vl_crop: Cropping method for VL processing (options: “center”, “disabled”)
vl_upscale: Upscaling method for VL processing (options: “bicubic”, “area”, “lanczos”)
mask: Optional mask for the image to define region of interest for editing
return_full_refs_cond (optional): Whether to return conditioning with all reference images or only with the main reference (default: True)
instruction (optional): Custom instruction for image editing
Outputs
CONDITIONING: The encoded conditioning tensor (with all or only main reference based on return_full_refs_cond)
LATENT: The encoded latent representation of the main reference image
custom_output: A dictionary containing:
pad_info: Padding information dictionary containing scaling and padding details
full_refs_cond: Conditioning with all reference latents
main_ref_cond: Conditioning with only the main reference latent
main_image: The processed main reference image
vae_images: List of all processed VAE images
ref_latents: List of all reference latents
vl_images: List of all processed VL images
full_prompt: The complete prompt with image descriptions
llama_template: The applied system prompt template
mask: The processed mask used for region of interest
Behavior
Provides maximum flexibility by allowing per-image configurations for both reference and VL processing
Supports multiple reference images with different processing requirements simultaneously
Allows fine-grained control over scaling, cropping, and resizing for each image
Returns comprehensive output dictionary with all intermediate results
Integrates with custom instructions for tailored image editing
Provides both full reference and main reference conditioning outputs
QwenEditConfigPreparer
This helper node creates configuration objects for use with the TextEncodeQwenImageEditPlusCustom_lrzjason node. It allows you to define custom processing parameters for individual images.
Inputs
image: The reference image to configure
configs (optional): An existing list of configuration objects to append to
to_ref (optional): Whether to include image in reference processing (default: True)
ref_main_image (optional): Whether this image is the main reference image (default: True)
ref_longest_edge (optional): Target longest edge size for reference processing (default: 1024)
ref_crop (optional): Cropping method for reference processing (options: “pad”, “center”, “disabled”, default: “center”)
ref_upscale (optional): Upscaling method for reference processing (options: “lanczos”, “bicubic”, “area”, default: “lanczos”)
to_vl (optional): Whether to include image in vision-language processing (default: True)
vl_resize (optional): Whether to resize image for VL processing (default: True)
vl_target_size (optional): Target size for VL processing (default: 384)
vl_crop (optional): Cropping method for VL processing (options: “center”, “disabled”, default: “center”)
vl_upscale (optional): Upscaling method for VL processing (options: “bicubic”, “area”, “lanczos”, default: “bicubic”)
mask (optional): Optional mask for the image to define region of interest for editing
Outputs
configs: The updated list of configuration objects
config: The configuration object for the current image
Behavior
Creates configuration objects that define how each image should be processed
Allows appending to existing configuration lists
Provides default values for all configuration parameters
Output config list can be connected directly to TextEncodeQwenImageEditPlusCustom_lrzjason
QwenEditConfigJsonParser
This helper node creates configuration objects from JSON strings for use with the TextEncodeQwenImageEditPlusCustom_lrzjason node. It provides an alternative method to define configuration parameters.
Inputs
image: The reference image to configure
configs (optional): An existing list of configuration objects to append to
config_json (optional): JSON string containing configuration parameters
mask (optional): Optional mask for the image to define region of interest for editing
Outputs
configs: The updated list of configuration objects
config: The configuration object for the current image
Behavior
Creates configuration objects from JSON strings
Allows appending to existing configuration lists
Provides a default JSON configuration template
Output config list can be connected directly to TextEncodeQwenImageEditPlusCustom_lrzjason
QwenEditOutputExtractor
This helper node extracts specific outputs from the custom_output dictionary produced by the TextEncodeQwenImageEditPlusCustom_lrzjason node.
Inputs
custom_output: The custom output dictionary from the custom node
Outputs
pad_info: Padding information dictionary
full_refs_cond: Conditioning with all reference latents
main_ref_cond: Conditioning with only the main reference latent
main_image: The main reference image
vae_images: List of all processed VAE images
ref_latents: List of all reference latents
vl_images: List of all processed VL images
full_prompt: The complete prompt with image descriptions
llama_template: The applied system prompt template
no_refs_cond: Conditioning without any reference latents
mask: The processed mask used for region of interest
Behavior
Extracts individual components from the complex output dictionary
Provides access to all intermediate results from the custom node
Enables modular processing of different output components
QwenEditListExtractor
This helper node extracts a specific item from a list based on its index position.
Inputs
items: The input list
index: The index of the item to extract (default: 0)
Outputs
item: The extracted item from the list
Behavior
Extracts a single item from a list based on index
Useful for extracting specific images from the vae_images list or other collections
Supports any list items regardless of type
QwenEditAdaptiveLongestEdge
This utility node calculates an appropriate longest edge size for an image, ensuring it doesn’t exceed a specified maximum size. This is particularly useful when you want to automatically determine the best longest edge for image processing based on image dimensions, especially when working with large images that need to be downscaled for processing.
Inputs
image: The input image to analyze
max_size: The maximum allowed size for the longest edge (default: 2048, range: 512-4096)
Outputs
longest_edge: The calculated longest edge size for the image that respects the max_size constraint
Behavior
Calculates the longest edge of the input image
If the longest edge exceeds max_size, it calculates a reduced size that fits within the constraint
Returns the appropriate longest edge size that can be used in other nodes like TextEncodeQwenImageEditPlusCustom_lrzjason for dynamic image processing
Useful for workflows where you want to prevent oversized images from causing memory or processing issues
CropWithPadInfo
This utility node crops an image using pad information generated by other nodes, allowing precise cropping back to the original content area after padding operations.
Inputs
image: The image to crop
pad_info: The pad information dictionary containing x, y, width, height and scale values
Outputs
cropped_image: The cropped image with original content dimensions
scale_by: The scale factor used in the original processing
Behavior
Uses pad information to crop images to their original content area
Removes padding that was added during processing
Returns the scale factor for potential additional operations
Key Features
Multi-Image Support: Incorporate up to 3 reference images into your text-to-image generation workflow
Dual Resize Options: Separate resizing controls for VAE encoding (1024px) and VL encoding (384px)
Individual Image Outputs: Each processed reference image is provided as a separate output for flexible connections
Latent Space Integration: Encode reference images into latent space for efficient processing
Qwen Model Compatibility: Specifically designed for Qwen-based image editing models
Customizable Templates: Use custom Llama templates for tailored image editing instructions
Installation
Clone or download this repository into your ComfyUI’s custom_nodes directory.
Restart ComfyUI.
The node will be available in the “advanced/conditioning” category.
Note: This repository has been succeeded by ComfyUI-EditUtils, which provides enhanced capabilities with support for multiple AI models including Qwen and Flux2Klein. We recommend checking out the new repository for future projects.
Contact
Twitter: @Lrzjason
Email: lrzjason@gmail.com
QQ Group: 866612947
Wechatid: fkdeai
Civitai: xiaozhijason
Sponsors me for more open source projects:
