<p align="center">
<img src="web/assets/brand/cmk-logo.png" width="140" alt="CMK Flow logo">
cmk_nodes Introduction
cmk_nodes, presented in ComfyUI as CMK Flow, is a collection of custom nodes and ready-made workflows for creating and refining AI images. It helps you move from model setup and image generation to targeted edits, face processing, comparison, and final output without having to assemble every step from scratch.
The CMK Flow Browser is the main place to explore the package inside ComfyUI. It gives you quick access to nodes, packaged subgraphs, image previews, and reference workflows. Its catalog includes 20 task workflows arranged from simpler to more involved, alongside module examples, comparisons, real-world workflows, a full-flow system reference, and legacy examples.
CMK Flow supports SDXL, Z-Image Turbo, and a HYBRID path that starts with SDXL and uses Z-Image Turbo for the final finish. It keeps these generation paths separate, then brings their images into a shared post-processing stage. That means you can explore different ways to generate an image while using a more consistent set of tools to refine it.
The author designed two kinds of nodes for different needs:
CMK … -Pipe-nodes form a guided workflow with defined connections. They help reduce confusing or incorrect connections between stages.- Other CMK nodes remain available through regular ComfyUI node search for more open-ended experimentation.
The package is free software under GPL-3.0-or-later. Models and other dependencies may have their own licenses.
How cmk_nodes Works
Think of a CMK workflow as an art process with distinct stages: gather your materials, create the first image, refine selected areas, and export the result.
- Prepare. Load the model resources and, if desired, LoRAs. The guided flow begins at
01 START HERE. - Generate. Use a sampler to create the image. Depending on the workflow, you can add optional guidance such as SDXL regional conditioning, ControlNet, or SDXL InstantID and refiner stages.
- Hand off the image. A PostProcess Boundary ends the model-specific generation stage. It passes the active image into a separate refinement context. In a workflow with parallel SDXL, Z-Image Turbo, or HYBRID paths, the
Combinedboundary takes the active path. - Refine and output. Apply modules such as FaceRebuild, FaceSwap, FaceProcess, Detailer, or MaskDetailer, then view the processing stages or upscale and save the image.
You will see connection labels such as MODEL, PROCESS, IMAGE, LOG, and VISUAL. These distinguish model resources, workflow state, image content, diagnostic information, and visualization. Some SDXL generation steps pass a SAMPLED result instead of a finished IMAGE. You do not need to manage the underlying representation to use a reference workflow; the labels mainly help show which connections belong together.
Module boundaries mark completed stages before an image goes to a comparer, another module, or an output. CMK also caches certain completed work. For example, when you change one face-processing branch in a parallel workflow, an unchanged neighboring branch may be reused during the same ComfyUI session instead of running again. A first run may report MISS → STORED; a later reuse may report HIT. These image caches are temporary and are cleared by a ComfyUI restart.
cmk_nodes Features
Guided workflows and reference examples
The Flow Browser offers previews and curated workflows you can examine as starting points. The catalog separates Task Workflows, Module Workflows, Comparisons, Real-World Workflows, the System Workflow, and Legacy examples. Use a task workflow to learn one operation, a module workflow to see a refinement stage on its own, or the system workflow to see how the full flow fits together.
Post-processing modules can also be used outside a complete CMK generation workflow, either as individual steps or as packaged subgraphs.
Image preparation and cropping
CMK Flow · Image Input prepares an image for a target size while preserving its proportions. New nodes default to a visible, aspect-ratio-safe crop instead of stretching the picture to fit.
You can choose center, top, bottom, left, or right to control which part stays in view. For example, choose top when a portrait’s face is close to the upper edge, or left when the subject should remain on the left side of a wide composition. Image-preparation tools also cover mask alignment, Replace, Remove, Extend, and controlled overlap for outpainting.
CMK Swap Image Loader -Pipe- presents the target and source images side by side. Resize and optional advanced crop apply to the target; the source keeps its original pixels.
Targeted refinement
Detailer and MaskDetailer workflows let you concentrate processing on selected areas rather than treating the entire image as one region. CMK supports parallel Smart Detailer and FaceProcess instances, so separate branches can work on different areas or faces. Dynamic SEGS, LOG BLOCK, and DIAGNOSTIC inputs allow these workflows to accommodate additional regions and processing information.
CMK FaceProcess -Pipe- shows the controls relevant to its selected Process Mode. When you switch between Restore and Detailer settings, the inactive settings remain stored and return when you switch back.
A central VISUAL chain can pass registered processing stages to the Visualizer, helping you inspect what happened along the way.
FaceSwap and its safety check
CMK provides local image and video FaceSwap paths. Every active public FaceSwap path uses a mandatory ContentGuard that checks the source and target images; video processing checks each target frame. It stops processing for explicit content, a target estimated to be under 18, a source estimated to be under the conservative threshold of 25, or missing or invalid protection models. There is no public bypass. A disabled FaceSwap node passes its input through without loading the guard.
The guard runs locally using NudeNet and InsightFace. Age estimates can be imperfect, especially for stylized or generated faces. The guard is a technical risk control, not a way to verify consent or a guarantee against incorrect classifications. See CONTENT_GUARD.md for its policy and diagnostic codes.
Video workflows
CMK Split Video into Segments selects a video from input/video/ and writes its segments under output/video/segments/<video_name>/. CMK Merge and Save Video recombines processed segments, shows the result in a player, and saves it under output/video/merged/<video_name>/. It can also publish the result without encoding it again. CMK FaceSwap Video Loader adds video and source-image selection to the segmented workflow.
CMK FaceSwap Video is a separate legacy reference from CMK 1.0. The author has kept it working but has not rebuilt it around the newer CMK 2.5 flow architecture.
Interface and optional LoRA workflows
CMK Flow requires Vue Nodes / Nodes 2.0 in your active ComfyUI profile. This interface supports its dynamic controls, including Advanced switching, dropdowns, shapes, and automatic node sizing. If it is off, CMK displays a startup notice.
The optional LoRA Stack · SDXL, LoRA Stack · ZIT, and LoRA Stack · Combined subgraphs, and workflows that use them, require the ComfyUI LoRA Manager. Other CMK modules remain usable without it. The author credits that project as an inspiration for the Flow Browser; CMK Flow is independent and has no official affiliation with it.
cmk_nodes Models
CMK Flow supports several model paths, but a path is not the same thing as one downloadable model file:
| Path | What it does | When you might use it |
|---|---|---|
| SDXL | Generates an image through the SDXL path, with optional regional conditioning, ControlNet, InstantID, and refiner stages where the workflow supports them. | When you want to work with those SDXL-specific generation and guidance steps. |
| Z-Image Turbo (ZIT) | Uses its own loader and sampler, with an optional ControlNet module. | When you want to build a workflow around the Z-Image Turbo path. |
| HYBRID | Builds the image with SDXL, then hands it to Z-Image Turbo for the final finish. | When you want to explore an SDXL-based image with a Z-Image Turbo finishing stage. |
All three paths can feed the family-neutral post-processing stage through the appropriate boundary. Changing the generation path changes how the initial image is made; it does not require a separate set of post-processing modules.
Features within those paths can require additional model resources. Depending on the workflow you choose, these may include checkpoints, VAEs, ControlNet models, InstantID and InsightFace resources, Ultralytics detector weights, SAM models, face-restore or FaceSwap weights, and upscaling models. These weights are not included with the Python package. A missing resource affects the feature that needs it rather than preventing unrelated CMK nodes from loading.
ZIT-Inpaint is marked EXPERIMENTAL and is temporarily frozen because of its very high memory and runtime demands. It should not be treated as a standard, lightweight inpainting option.
What's New with cmk_nodes
The supplied project overview identifies CMK 2.7.0 as both the development baseline and the stated public release, but it does not provide a version-by-version list of changes through 2.7.0. The changes it does describe are:
- CMK 2.5 flow architecture: Model-specific generation paths are separated from a shared post-processing stage by PostProcess Boundaries. This makes the distinction between creating an image and refining it easier to follow.
- Packaged workflow catalog: The Flow Browser organizes examples by task and complexity, including 20 progressively more involved task workflows. This gives you smaller examples to explore before opening the full system workflow.
- Dedicated Z-Image Turbo path: ZIT has its own loader, sampler, and optional ControlNet modules, alongside the SDXL and HYBRID paths.
- Parallel processing and caching: Smart Detailer and FaceProcess branches can run in parallel, while unchanged work can be reused within a ComfyUI session.
- CMK 1.0 legacy reference retained: The original FaceSwap video workflow remains available as a legacy example, distinct from the newer flow architecture.
For the historical change record, see CHANGELOG.md. It is a history of changes, not the definition of current node interfaces.
Troubleshooting cmk_nodes
Dynamic controls or node shapes do not look right
- Check your active ComfyUI user profile for Vue Nodes / Nodes 2.0 and enable it.
- Return to the CMK workflow and check the Advanced controls, dropdowns, and node layout again. The interface switches when the setting is activated; a reload is not required.
- If CMK displayed a startup notice about this setting, confirm you changed it in the profile you are using.
Sampler or refiner previews are missing
- Open Comfy → Execution → Live preview method.
- Set it to auto.
- Run the workflow again to check whether previews appear during processing.
A workflow reports a missing model
- Read the error to identify which feature is affected. A missing model in one path does not mean all CMK nodes are unusable.
- Check that the model required by your chosen workflow is available in a model location ComfyUI recognizes. The package’s Python dependencies do not include model weights.
- Consult the workflow overview and the project’s resource audit information for the resources used by that workflow. The audit checks specific expected filenames and reports resources individually as
FOUNDorMISSING. - If you use a shared model directory and have just registered it with ComfyUI, restart ComfyUI so it can load the path.
An optional LoRA Stack subgraph is unavailable
Check whether the ComfyUI LoRA Manager is available. It is required by the optional SDXL, ZIT, and Combined LoRA Stack subgraphs and examples that use them. You can continue using CMK modules that do not depend on those subgraphs.
The browser shows duplicate subgraph blueprints
Check for extra copies of CMK’s packaged subgraph JSON files under user/default/subgraphs/. The subgraphs already belong in the node pack’s subgraphs/ folder; extra copies can create duplicate entries. Back up any same-named workflows you created outside the node pack before removing duplicates.
A FaceSwap stops with a ContentGuard error
- Read the diagnostic message or code to determine whether the stop concerns detected content, an age estimate, or a missing or invalid protection model.
- For a protection-model error, check that the required local models are available.
- For a content or age-estimate stop, use different, appropriate source and target images. There is no public bypass.
- See CONTENT_GUARD.md for the policy and the meaning of its diagnostic codes.
Why does a workflow run slowly the first time, or after a restart?
A first run can build temporary cached results and report MISS → STORED. Unchanged branches may then report HIT during the same ComfyUI session. Restarting ComfyUI clears the image-boundary caches, so some processing must run again. Cached files under ComfyUI/temp/cmk/ are accelerators, not a portable way to save a project.
Which versions have been checked?
The stated CMK 2.7.0 reference environment is ComfyUI 0.42.5, Python 3.13.12, PyTorch 2.12.1, and Apple MPS. These values describe a validated environment, not minimum requirements or a guarantee that every workflow and hardware combination has been tested. The project also reports a separate, fully validated CMK 2.5.7 frontend target of ComfyUI 0.38.2 with comfyui-frontend-package 1.53.6.
Learn More about cmk_nodes
- Source repository: Explore the project and its packaged materials.
- English project overview: Read the upstream English overview.
- Workflow guide: Find the subgraph and workflow inventory when choosing an example to explore.
- ContentGuard policy: Understand FaceSwap checks, stop conditions, and limitations.
- Change history: Review recorded changes across versions.
- Architecture reference: Consult the authoritative interface description when you need more technical detail.
- ComfyUI LoRA Manager: Learn about the separate tool required by CMK’s optional LoRA Stack subgraphs.
