Giters

rheros / ComfyUI-Impact-Pack

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Youtube Badge

ComfyUI-Impact-Pack

Custom nodes pack for ComfyUI This custom node helps to conveniently enhance images through Detector, Detailer, Upscaler, Pipe, and more.

NOTICE

  • Selection weight syntax is changed(: -> ::) since V3.16. (tutorial)
  • Starting from V3.6, requires latest version(Aug 8, 9ccc965) of ComfyUI.
  • In versions below V3.3.1, there was an issue with the image quality generated after using the UltralyticsDetectorProvider. Please make sure to upgrade to a newer version.
  • Starting from V3.0, nodes related to mmdet are optional nodes that are activated only based on the configuration settings.
    • Through ComfyUI-Impact-Subpack, you can utilize UltralyticsDetectorProvider to access various detection models.
  • Between versions 2.22 and 2.21, there is partial compatibility loss regarding the Detailer workflow. If you continue to use the existing workflow, errors may occur during execution. An additional output called "enhanced_alpha_list" has been added to Detailer-related nodes.
  • The permission error related to cv2 that occurred during the installation of Impact Pack has been patched in version 2.21.4. However, please note that the latest versions of ComfyUI and ComfyUI-Manager are required.
  • The "PreviewBridge" feature may not function correctly on ComfyUI versions released before July 1, 2023.
  • Attempting to load the "ComfyUI-Impact-Pack" on ComfyUI versions released before June 27, 2023, will result in a failure.
  • With the addition of wildcard support in FaceDetailer, the structure of DETAILER_PIPE-related nodes and Detailer nodes has changed. There may be malfunctions when using the existing workflow.

Custom Nodes

  • SAMLoader - Loads the SAM model.

  • UltralyticsDetectorProvider - Loads the Ultralystics model to provide SEGM_DETECTOR, BBOX_DETECTOR.

    • Unlike MMDetDetectorProvider, for segm models, BBOX_DETECTOR is also provided.
    • The various models available in UltralyticsDetectorProvider can be downloaded through ComfyUI-Manager.

  • ONNXDetectorProvider - Loads the ONNX model to provide SEGM_DETECTOR.

  • CLIPSegDetectorProvider - Wrapper for CLIPSeg to provide BBOX_DETECTOR.

    • You need to install the ComfyUI-CLIPSeg node extension.

  • SEGM Detector (combined) - Detects segmentation and returns a mask from the input image.

  • BBOX Detector (combined) - Detects bounding boxes and returns a mask from the input image.

  • SAMDetector (combined) - Utilizes the SAM technology to extract the segment at the location indicated by the input SEGS on the input image and outputs it as a unified mask.

  • SAMDetector (Segmented) - It is similar to SAMDetector (combined), but it separates and outputs the detected segments. Multiple segments can be found for the same detected area, and currently, a policy is in place to group them arbitrarily in sets of three. This aspect is expected to be improved in the future.

    • As a result, it outputs the combined_mask, which is a unified mask, and batch_masks, which are multiple masks grouped together in batch form.
    • While batch_masks may not be completely separated, it provides functionality to perform some level of segmentation.

  • Simple Detector (SEGS) - Operating primarily with BBOX_DETECTOR, and with the additional provision of SAM_MODEL or SEGM_DETECTOR, this node internally generates improved SEGS through mask operations on both bbox and silhouette. It serves as a convenient tool to simplify a somewhat intricate workflow.

  • ControlNetApply (SEGS) - To apply ControlNet in SEGS, you need to use the Preprocessor Provider node from the Inspire Pack to utilize this node.

  • Bitwise(SEGS & SEGS) - Performs a 'bitwise and' operation between two SEGS.

  • Bitwise(SEGS - SEGS) - Subtracts one SEGS from another.

  • Bitwise(SEGS & MASK) - Performs a bitwise AND operation between SEGS and MASK.

  • Bitwise(SEGS & MASKS ForEach) - Performs a bitwise AND operation between SEGS and MASKS.

    • Please note that this operation is performed with batches of MASKS, not just a single MASK.

  • Bitwise(MASK & MASK) - Performs a 'bitwise and' operation between two masks.

  • Bitwise(MASK - MASK) - Subtracts one mask from another.

  • Bitwise(MASK + MASK) - Combine two masks.

  • SEGM Detector (SEGS) - Detects segmentation and returns SEGS from the input image.

  • BBOX Detector (SEGS) - Detects bounding boxes and returns SEGS from the input image.

  • ONNX Detector (SEGS) - Utilizes the ONNX model to identify the bbox and retrieve the SEGS from the input image.

  • Detailer (SEGS) - Refines the image based on SEGS.

  • DetailerDebug (SEGS) - Refines the image based on SEGS. Additionally, it provides the ability to monitor the cropped image and the refined image of the cropped image.

    • To prevent regeneration caused by the seed that does not change every time when using 'external_seed', please disable the 'seed random generate' option in the 'Detailer...' node.

  • MASK to SEGS - Generates SEGS based on the mask.

  • ToBinaryMask - Separates the mask generated with alpha values between 0 and 255 into 0 and 255. The non-zero parts are always set to 255.

  • Masks to Mask List - MASKS

    • This node converts the MASKS in batch form to a list of individual masks.

  • EmptySEGS - Provides an empty SEGS.

  • MaskPainter - Provides a feature to draw masks.

  • FaceDetailer - Easily detects faces and improves them.

  • FaceDetailer (pipe) - Easily detects faces and improves them (for multipass).

  • FromDetailer (SDXL/pipe), BasicPipe -> DetailerPipe (SDXL), Edit DetailerPipe (SDXL) - These are pipe functions used in Detailer for utilizing the refiner model of SDXL.

  • SEGSDetailer - Performs detailed work on SEGS without pasting it back onto the original image.

  • SEGSPaste - Pastes the results of SEGS onto the original image.

    • If ref_image_opt is present, the images contained within SEGS are ignored. Instead, the image within ref_image_opt corresponding to the crop area of SEGS is taken and pasted. The size of the image in ref_image_opt should be the same as the original image size.

  • SEGSPreview - Provides a preview of SEGS.

    • This option is used to preview the improved image through SEGSDetailer before merging it into the original. Prior to going through SEGSDetailer, SEGS only contains mask information without image information. If fallback_image_opt is connected to the original image, SEGS without image information will generate a preview using the original image. However, if SEGS already contains image information, fallback_image_opt will be ignored.

  • SEGSToImageList - Convert SEGS To Image List

  • SEGSToMaskList - Convert SEGS To Mask List

  • SEGS Filter (label) - This node filters SEGS based on the label of the detected areas.

  • SEGS Filter (ordered) - This node sorts SEGS based on size and position and retrieves SEGs within a certain range.

  • SEGS Filter (range) - This node retrieves only SEGs from SEGS that have a size and position within a certain range.

  • SEGSConcat - Concatenate segs1 and segs2. If source shape of segs1 and segs2 are different then segs2 will be ignored.

  • Pipe nodes

    • ToDetailerPipe, FromDetailerPipe - These nodes are used to bundle multiple inputs used in the detailer, such as models and vae, ..., into a single DETAILER_PIPE or extract the elements that are bundled in the DETAILER_PIPE.
    • ToBasicPipe, FromBasicPipe - These nodes are used to bundle model, clip, vae, positive conditioning, and negative conditioning into a single BASIC_PIPE, or extract each element from the BASIC_PIPE.
    • EditBasicPipe, EditDetailerPipe - These nodes are used to replace some elements in BASIC_PIPE or DETAILER_PIPE.
    • FromDetailerPipe_v2, FromBasicPipe_v2 - It has the same functionality as FromDetailerPipe and FromBasicPipe, but it has an additional output that directly exports the input pipe. It is useful when editing EditBasicPipe and EditDetailerPipe.

  • Latent Scale (on Pixel Space) - This node converts latent to pixel space, upscales it, and then converts it back to latent.

    • If upscale_model_opt is provided, it uses the model to upscale the pixel and then downscales it using the interpolation method provided in scale_method to the target resolution.

  • PixelKSampleUpscalerProvider - An upscaler is provided that converts latent to pixels using VAEDecode, performs upscaling, converts back to latent using VAEEncode, and then performs k-sampling. This upscaler can be attached to nodes such as 'Iterative Upscale' for use.

    • Similar to 'Latent Scale (on Pixel Space)', if upscale_model_opt is provided, it performs pixel upscaling using the model.

  • PixelTiledKSampleUpscalerProvider - It is similar to PixelKSampleUpscalerProvider, but it uses ComfyUI_TiledKSampler and Tiled VAE Decoder/Encoder to avoid GPU VRAM issues at high resolutions.

  • DenoiseScheduleHookProvider - IterativeUpscale provides a hook that gradually changes the denoise to target_denoise as the step progresses.

  • CfgScheduleHookProvider - IterativeUpscale provides a hook that gradually changes the cfg to target_cfg as the step progresses.

  • PixelKSampleHookCombine - This is used to connect two PK_HOOKs. hook1 is executed first and then hook2 is executed.

    • If you want to simultaneously change cfg and denoise, you can combine the PK_HOOKs of CfgScheduleHookProvider and PixelKSampleHookCombine.

  • NoiseInjectionHookProvider - During each iteration of IterativeUpscale, noise is injected into the latent space while varying the strength according to a schedule.

    • You need to install the BlenderNeko/ComfyUI_Noise node extension.
    • The seed serves as the initial value required for generating noise, and it increments by 1 with each iteration as the process unfolds.
    • The source determines the types of CPU noise and GPU noise to be configured.
    • Currently, there is only a simple schedule available, where the strength of the noise varies from start_strength to end_strength during the progression of each iteration.

  • NoiseInjectionDetailerHookProvider - The detailer_hook is a hook in the Detailer that injects noise during the processing of each SEGS.

  • Iterative Upscale (Latent) - The upscaler takes the input upscaler and splits the scale_factor into steps, then iteratively performs upscaling. This takes latent as input and outputs latent as the result.

  • Iterative Upscale (Image) - The upscaler takes the input upscaler and splits the scale_factor into steps, then iteratively performs upscaling. This takes image as input and outputs image as the result.

    • Internally, this node uses 'Iterative Upscale (Latent)'.

  • TwoSamplersForMask - This node can apply two samplers depending on the mask area. The base_sampler is applied to the area where the mask is 0, while the mask_sampler is applied to the area where the mask is 1.

    • Note: The latent encoded through VAEEncodeForInpaint cannot be used.

  • KSamplerProvider - This is a wrapper that enables KSampler to be used in TwoSamplersForMask TwoSamplersForMaskUpscalerProvider.

  • TiledKSamplerProvider - ComfyUI_TiledKSampler is a wrapper that provides KSAMPLER.

  • TwoAdvancedSamplersForMask - TwoSamplersForMask is similar to TwoAdvancedSamplersForMask, but they differ in their operation. TwoSamplersForMask performs sampling in the mask area only after all the samples in the base area are finished. On the other hand, TwoAdvancedSamplersForMask performs sampling in both the base area and the mask area sequentially at each step.

  • KSamplerAdvancedProvider - This is a wrapper that enables KSampler to be used in TwoAdvancedSamplersForMask.

  • TwoSamplersForMaskUpscalerProvider - This is an Upscaler that extends TwoSamplersForMask to be used in Iterative Upscale.

    • TwoSamplersForMaskUpscalerProviderPipe - pipe version of TwoSamplersForMaskUpscalerProvider.

  • PreviewBridge - This custom node can be used with a bridge when using the MaskEditor feature of Clipspace.

  • ImageSender, ImageReceiver - The images generated in ImageSender are automatically sent to the ImageReceiver with the same link_id.

  • LatentSender, LatentReceiver - The latent generated in LatentSender are automatically sent to the LatentReceiver with the same link_id.

    • Furthermore, LatentSender is implemented with PreviewLatent, which stores the latent in payload form within the image thumbnail.
    • Due to the current structure of ComfyUI, it is unable to distinguish between SDXL latent and SD1.5/SD2.1 latent. Therefore, it generates thumbnails by decoding them using the SD1.5 method.

  • Switch (image,mask), Switch (latent), Switch (SEGS) - Among multiple inputs, it selects the input designated by the selector and outputs it. The first input must be provided, while the others are optional. However, if the input specified by the selector is not connected, an error may occur.

  • ImpactWildcardProcessor - The text is generated by processing the wildcard in the Text. If the mode is set to "populate", a dynamic prompt is generated with each execution and the input is filled in the second textbox. If the mode is set to "fixed", the content of the second textbox remains unchanged.

    • When an image is generated with the "fixed" mode, the prompt used for that particular generation is stored in the metadata.

  • ImpactWildcardEncode - Similar to ImpactWildcardProcessor, this provides the loading functionality of LoRAs (e.g. <lora:some_awesome_lora:0.7:1.2>). Populated prompts are encoded using the clip after all the lora loading is done.

  • RegionalSampler, CombineRegionalPrompts, RegionalPrompt - experimental feature

  • multiple region version of TwoAdvancedSamplersForMask

  • KSampler (pipe), KSampler (advanced/pipe)

  • ImpactCompare, ImpactConditionalBranch, ImpactInt, ImpactValueSender, ImpactValueReceiver, ImpactImageInfo, ImpactMinMax, ImpactNeg, ImpactConditionalStopIteration

  • Experimental set of nodes for implementing loop functionality (tutorial to be prepared later / example workflow).
  • Image batch To Image List - Convert Image batch to Image List
  • You can use images generated in a multi batch to handle them
  • Make Image List - Convert multiple images into a single image list
  • The input of images can be scaled up as needed
  • String Selector - It selects and returns a portion of the string. When multiline mode is disabled, it simply returns the string of the line pointed to by the selector. When multiline mode is enabled, it divides the string based on lines that start with # and returns them. If the select value is larger than the number of items, it will start counting from the first line again and return accordingly.

MMDet nodes

  • MMDetDetectorProvider - Loads the MMDet model to provide BBOX_DETECTOR and SEGM_DETECTOR.
  • To use the existing MMDetDetectorProvider, you need to enable the MMDet usage configuration.

Feature

  • Interactive SAM Detector (Clipspace) - When you right-click on a node that has 'MASK' and 'IMAGE' outputs, a context menu will open. From this menu, you can either open a dialog to create a SAM Mask using 'Open in SAM Detector', or copy the content (likely mask data) using 'Copy (Clipspace)' and generate a mask using 'Impact SAM Detector' from the clipspace menu, and then paste it using 'Paste (Clipspace)'.
  • Providing a feature to detect errors that occur when mixing models and clips from checkpoints such as SDXL Base, SDXL Refiner, SD1.x, SD2.x during sample execution, and reporting appropriate errors.

Deprecated

  • The following nodes have been kept only for compatibility with existing workflows, and are no longer supported. Please replace them with new nodes.
    • MMDetLoader -> MMDetDetectorProvider
    • SegsMaskCombine -> SEGS to MASK (combined)
    • BboxDetectorForEach -> BBOX Detector (SEGS)
    • SegmDetectorForEach -> SEGM Detector (SEGS)
    • BboxDetectorCombined -> BBOX Detector (combined)
    • SegmDetectorCombined -> SEGM Detector (combined)
    • MaskPainter -> PreviewBridge
  • To use the existing deprecated legacy nodes, you need to enable the MMDet usage configuration.

How to activate 'MMDet usage'

  • Upon the initial execution, an impact-pack.ini file will be generated in the custom_nodes/ComfyUI-Impact-Pack directory.
[default]
dependency_version = 2
mmdet_skip = True
  • Change mmdet_skip = True to mmdet_skip = False
[default]
dependency_version = 2
mmdet_skip = False
  • Restart ComfyUI

Installation

  1. cd custom_nodes

  2. git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack.git

  3. cd ComfyUI-Impact-Pack

  4. (optional) git submodule update --init --recursive

    • Impact Pack will automatically download subpack during its initial launch.

  5. (optional) python install.py

    • Impact Pack will automatically install its dependencies during its initial launch.
    • For the portable version, you should execute the command ..\..\..\python_embeded\python.exe install.py to run the installation script.

  6. Restart ComfyUI

  • NOTE: If an error occurs during the installation process, please refer to Troubleshooting Page for assistance.
  • You can use this colab notebook colab notebook to launch it. This notebook automatically downloads the impact pack to the custom_nodes directory, installs the tested dependencies, and runs it.

Package Dependencies (If you need to manual setup.)

  • pip install

    • openmim
    • segment-anything
    • ultralytics
    • scikit-image
    • piexif
    • (optional) pycocotools
    • (optional) onnxruntime

  • mim install (optional)

    • mmcv==2.0.0, mmdet==3.0.0, mmengine==0.7.2

  • linux packages (ubuntu)

    • libgl1-mesa-glx
    • libglib2.0-0

Config example

  • Once you run the Impact Pack for the first time, an impact-pack.ini file will be automatically generated in the Impact Pack directory. You can modify this configuration file to customize the default behavior.
    • dependency_version - don't touch this
    • mmdet_skip - disable MMDet based nodes and legacy nodes if True
    • sam_editor_cpu - use cpu for SAM editor instead of gpu
    • sam_editor_model: Specify the SAM model for the SAM editor.
      • You can download various SAM models using ComfyUI-Manager.
      • Path to SAM model: ComfyUI/models/sams
[default]
dependency_version = 9
mmdet_skip = True
sam_editor_cpu = False
sam_editor_model = sam_vit_b_01ec64.pth

Other Materials (auto-download on initial startup)

Troubleshooting page

How to use (DDetailer feature)

1. Basic auto face detection and refine exapmle.

simple

  • The face that has been damaged due to low resolution is restored with high resolution by generating and synthesizing it, in order to restore the details.
  • The FaceDetailer node is a combination of a Detector node for face detection and a Detailer node for image enhancement. See the Advanced Tutorial for a more detailed explanation.
  • Pass the MMDetLoader 's bbox model and the detection model loaded by SAMLoader to FaceDetailer . Since it performs the function of KSampler for image enhancement, it overlaps with KSampler's options.
  • The MASK output of FaceDetailer provides a visualization of where the detected and enhanced areas are.

simple-orig simple-refined

  • You can see that the face in the image on the left has increased detail as in the image on the right.

2. 2Pass refine (restore a severely damaged face)

2pass-workflow-example

  • Although two FaceDetailers can be attached together for a 2-pass configuration, various common inputs used in KSampler can be passed through DETAILER_PIPE, so FaceDetailerPipe can be used to configure easily.
  • In 1pass, only rough outline recovery is required, so restore with a reasonable resolution and low options. However, if you increase the dilation at this time, not only the face but also the surrounding parts are included in the recovery range, so it is useful when you need to reshape the face other than the facial part.

2pass-example-original 2pass-example-middle 2pass-example-result

  • In the first stage, the severely damaged face is restored to some extent, and in the second stage, the details are restored

3. Face Bbox(bounding box) + Person silhouette segmentation (prevent distortion of the background.)

combination-workflow-example combination-example-original combination-example-refined

  • Facial synthesis that emphasizes details is delicately aligned with the contours of the face, and it can be observed that it does not affect the image outside of the face.

  • The BBoxDetectorForEach node is used to detect faces, and the SAMDetectorCombined node is used to find the segment related to the detected face. By using the Segs & Mask node with the two masks obtained in this way, an accurate mask that intersects based on segs can be generated. If this generated mask is input to the DetailerForEach node, only the target area can be created in high resolution from the image and then composited.

4. Iterative Upscale

upscale-workflow-example

  • The IterativeUpscale node is a node that enlarges an image/latent by a scale_factor. In this process, the upscale is carried out progressively by dividing it into steps.

  • IterativeUpscale takes an Upscaler as an input, similar to a plugin, and uses it during each iteration. PixelKSampleUpscalerProvider is an Upscaler that converts the latent representation to pixel space and applies ksampling.

    • The upscale_model_opt is an optional parameter that determines whether to use the upscale function of the model base if available. Using the upscale function of the model base can significantly reduce the number of iterative steps required. If an x2 upscaler is used, the image/latent is first upscaled by a factor of 2 and then downscaled to the target scale at each step before further processing is done.

  • The following image is an image of 304x512 pixels and the same image scaled up to three times its original size using IterativeUpscale.

combination-example-original combination-example-refined

5. Interactive SAM Detector (Clipspace)

  • When you right-click on the node that outputs 'MASK' and 'IMAGE', a menu called "Open in SAM Detector" appears, as shown in the following picture. Clicking on the menu opens a dialog in SAM's functionality, allowing you to generate a segment mask. samdetector-menu

  • By clicking the left mouse button on a coordinate, a positive prompt in blue color is entered, indicating the area that should be included. Clicking the right mouse button on a coordinate enters a negative prompt in red color, indicating the area that should be excluded. Positive prompts represent the areas that should be included, while negative prompts represent the areas that should be excluded.

  • You can remove the points that were added by using the "undo" button. After selecting the points, pressing the "detect" button generates the mask. Additionally, you can adjust the fidelity slider to determine the extent to which the mask belongs to the confidence region.

samdetector-dialog

  • If you opened the dialog through "Open in SAM Detector" from the node, you can directly apply the changes by clicking the "Save to node" button. However, if you opened the dialog through the "clipspace" menu, you can save it to clipspace by clicking the "Save" button.

samdetector-result

  • When you execute using the reflected mask in the node, you can observe that the image and mask are displayed separately.

Others Tutorials

Credits

ComfyUI/ComfyUI - A powerful and modular stable diffusion GUI.

dustysys/ddetailer - DDetailer for Stable-diffusion-webUI extension.

Bing-su/dddetailer - The anime-face-detector used in ddetailer has been updated to be compatible with mmdet 3.0.0, and we have also applied a patch to the pycocotools dependency for Windows environment in ddetailer.

facebook/segment-anything - Segmentation Anything!

hysts/anime-face-detector - Creator of anime-face_yolov3, which has impressive performance on a variety of art styles.

open-mmlab/mmdetection - Object detection toolset. dd-person_mask2former was trained via transfer learning using their R-50 Mask2Former instance segmentation model as a base.

biegert/ComfyUI-CLIPSeg - This is a custom node that enables the use of CLIPSeg technology, which can find segments through prompts, in ComfyUI.

BlenderNeok/ComfyUI-TiledKSampler - The tile sampler allows high-resolution sampling even in places with low GPU VRAM.

WASasquatch/was-node-suite-comfyui - A powerful custom node extensions of ComfyUI.

About

License:GNU General Public License v3.0

Languages

Language:Python 85.7%Language:JavaScript 12.6%Language:Jupyter Notebook 1.7%