ComfyUI-LukutarNodes Introduction
ComfyUI-LukutarNodes is a collection of custom nodes for ComfyUI. The author brings together tools for image processing, region-based prompts, workflow organization, sampling, and image metadata.
You can use the nodes independently. For example, you might add Color Match (Frequency Split) to an existing upscale workflow without changing how you load your model or save your images. Other nodes are designed to work together: Prompt Annotate creates labeled prompt regions that the annotation nodes can edit or send to an Impact Pack Detailer.
The collection is organized into these ComfyUI menu categories:
| Category | What it helps you do |
|---|---|
Lukutar/Image | Correct colors, inspect detections, adjust detection crops, and save images. |
Lukutar/Prompt | Write one prompt with labeled regions and turn those labels into useful outputs. |
Lukutar/Variables | Move one value or a group of values across a large workflow without long visible wires. |
Lukutar/Groups | Switch canvas groups on or off from one panel. |
Lukutar/Seeds | Generate a seed and keep a visible history of seeds used. |
Lukutar/Bundle | Carry a checkpoint, prompts, latent, and related generation components together. |
Lukutar/Info and Lukutar/Metadata | Collect generation details and prepare them for image metadata. |
How ComfyUI-LukutarNodes Works
A ComfyUI workflow is a set of connected steps. Each Lukutar node performs a particular job and passes its result to the next step. Some nodes change an image; others organize information so that different parts of your workflow can use it.
Three ideas are especially useful to understand:
- Separate detail from color. Color Match treats an image somewhat like two layers: broad areas of color and tone, and fine detail. It adjusts the broad layer using a reference image, then keeps the processed image’s fine detail. This is useful when an upscale looks sharp but its colors have drifted.
- Write one prompt for several regions. Prompt Annotate lets you mark text for areas such as
faceorbackground. The unmarked text remains common to the whole image. Other annotation nodes can then extract a region’s instructions, edit them, or format them for an Impact Pack Detailer. - Carry related values together. A bundle is like a labeled kit containing the model, VAE, prompts, and other values needed for generation. Instead of connecting each value separately through every stage, you can pass the kit along and unpack it when you need an individual part.
Some conveniences depend on the extension’s ComfyUI interface assets. For instance, they provide the highlighted prompt editor, automatically managed bundle sockets, and invisible links between named variable nodes. Where the underlying node also works without those assets, you may see a simpler interface or need to connect it manually.
ComfyUI-LukutarNodes Features
Color correction and detection previews
Color Match (Frequency Split) recolors a processed image using a reference while retaining the processed image’s detail. A practical example is to connect your original image as reference and a color-drifted upscale as image.
reinhard, the default method, adjusts the broad color distribution while preserving more of the processed image’s tonal character.replaceuses the reference’s broad color layer directly, producing a stronger correction.sigmacontrols the separation between broad color and fine detail; its default is30.0. If you are unsure which value to use,auto_tunetests a range and outputs the selected sigma. Theenvelopemetric focuses on broad color and tone, whilefullalso considers image structure.
For a batch, one reference frame can be used for every processed frame, or equally sized batches can be matched frame by frame. Auto-tuning tests the first pair and uses the chosen setting for the entire batch.
SEGS BBox Overlay draws labeled detection boxes over an image so you can check what an Impact Pack detector found before a Detailer repaints those areas. You can show a class name alone or add its confidence percentage, tint detected masks, change tint strength, and adjust box and label sizes. auto color mode gives each detected class a consistent color; single uses one RGB color for all detections. Its segs output passes the detections onward unchanged.
SEGS Set Crop Size changes the crop regions used around Impact Pack detections. This helps when a detector’s relative crop setting produces very different crop sizes for small and large subjects. In exact mode, you request a width and height such as 512 × 512. In aspect mode, the crop follows the detection’s proportions while scaling toward the requested size. The node keeps the detection box, label, confidence, and mask content; it changes the crop region and aligns the mask to it. A crop can grow beyond the exact target if necessary to avoid cutting off a detection, and crops larger than the image are limited to the image size.
Labeled prompts and regional edits
Prompt Annotate lets you keep a single readable prompt while assigning instructions to specific labels:
masterpiece, |face: blue eyes, smirk|, |hair: red hair|, |background: outdoors, park|
The text outside tags, masterpiece in this example, is common to the prompt. The clean_prompt output removes the tags for base generation; the annotations output retains the labeled structure for the other nodes. Labels are names you choose, but they should match the labels used by the rest of your region-based workflow. Tags cannot be nested. One tag can name several labels, such as |body,hair: red hair|.
With the interface assets loaded, the prompt field highlights labeled text as you type. Without them, you can still type the markup as plain text.
The companion nodes let you use that prompt in different ways:
| Node | What you can do with it |
|---|---|
| Annotations to Wildcard (LAB) | Produce [LAB] text for an Impact Pack Detailer’s label-mode wildcard input. Common text goes under [ALL]; each region receives its labeled text. |
| Annotation Segment | Extract one label’s text. Turn on include_common to add the unmarked, shared prompt text. |
| Annotation Labels | Get a comma-separated list of the labels you used, in first-appearance order. You can optionally include the implicit all label. |
| Annotations to String | Convert annotations back to ordinary text. Choose clean for text without markup or raw for the original marked-up prompt. |
| Annotation Segment Edit | Change one or more labels without rewriting the shared source prompt. |
With Annotation Segment Edit, prepend and append add text before or after a label’s existing text; remove deletes matching comma-separated tags; new creates a labeled span; and delete removes selected labels and their content. For example, you could keep your main face prompt unchanged but append smile only in a smiling-portrait branch. If several labels share the same span, editing its text affects all of them.
Organizing large workflows
Set Variable / Get Variable let you give a value a name and retrieve it elsewhere on the canvas. This is useful when a wire carrying an image, model, conditioning, or another value would otherwise cross several groups. The interface connects matching nodes with a real link that does not appear on the canvas. A manual wire into Get Variable overrides name-based matching.
A name can be reused by Set nodes on alternative branches only when one branch is active. Two simultaneously active Set nodes with the same name are ambiguous and cause an error. Values belong to the current queue run; these nodes are not storage between runs.
Set Bundle / Get Bundle extend that idea to several values under one name. You might group model, vae, and clip as a kit named sdxl. Connecting a value to Set Bundle creates a typed, named slot; a matching Get Bundle presents corresponding outputs. You can rename slots for clarity, add slots ahead of time, or remove them. A bundle supports up to 16 slots. These dynamically managed sockets require the extension’s interface assets for normal canvas wiring.
Group Switch displays a row for each root-canvas group, letting you switch groups from one panel. Its bypass setting lets inputs pass through bypassed nodes; disable mutes the group’s nodes so they leave the queued prompt. You can also change the color of switched-off group frames and use title prefixes to order the panel’s rows. The switch reflects changes made through ComfyUI’s native group controls, but does not list groups inside subgraphs.
Seed History supplies a seed to a sampler and displays a rolling, copyable list of seeds sent to the backend. You choose how many entries to keep with history_size (default 5). It uses ComfyUI’s familiar after-generation seed controls, such as fixed, increment, decrement, and randomize. Repeated consecutive uses of a fixed seed do not fill the log with duplicate entries. Without the interface assets, the node still supplies a seed, but the history is not displayed.
Checkpoint bundles and sampling
Load Checkpoint (Bundle) combines several familiar generation steps. It loads a checkpoint, can apply an optional LoRA List, encodes positive and negative prompts, creates the starting latent, and outputs both a generation bundle and an info record of its settings. You can choose the checkpoint, baked-in or separate VAE, image dimensions, batch size, CLIP skip, and prompt-weight interpretation.
LoRA List collects up to eight LoRAs with individual weights for that loader. An unused choice or a zero-weight LoRA is skipped.
KSampler (Bundle) takes the bundle and info record, samples the latent, and returns updated versions of both. Its controls include seed, steps, CFG, sampler, scheduler, denoise strength, and vae_decode. With decoding enabled, the finished image replaces the noisy progress preview in the node. The node also saves temporary preview images for its image context-menu actions. With decoding disabled, the bundle does not gain a newly decoded image and the latent preview remains.
Checkpoint Bundle Unpack exposes the bundle’s individual model, CLIP, VAE, conditioning, latent, and image sockets for use with standard nodes. Checkpoint Bundle Pack assembles those components into a bundle again. This makes it possible to use the compact bundle path for most of a workflow while still accessing individual components where needed.
Generation information, metadata, and saving
The extension provides several ways to collect and prepare the details of a generation:
- Info Unpack exposes standard generation-info entries as separate text outputs. Info Pack combines an optional existing info record with up to four additional key/value pairs.
- Info to A1111 Parameters turns a bundle’s info record into an A1111-style
parametersmetadata block. It omits pack-only fields and adjusts supported values to A1111 infotext conventions. - Metadata Blocks creates a metadata block from up to 16 editable name/value pairs. Standard Metadata supplies workflow, prompt, and datetime entries. Merge Metadata combines blocks without leaving duplicate block names.
- A1111 Parameters assembles an A1111-style infotext from prompts and settings. A1111 Sampling and A1111 Hires Fix supply their respective settings as blocks. A1111 Model Info and A1111 Lora Hashes add model or LoRA hash information.
Save Image (Metadata) saves image batches as PNG, JPG, WEBP, TIFF, or BMP, with optional metadata where the format supports it. You choose a destination path, base filename, format, and, for JPG or WEBP, a quality setting. Save Image JXL is a separate JPEG XL saver with lossy and lossless options and additional encoder controls.
Finally, Paste (Clipspace, Keep Mask) and Paste Image (Keep Mask) add context-menu paste actions for image upload nodes. They let you replace an image while retaining a mask you have painted on that node.
Troubleshooting ComfyUI-LukutarNodes
A named variable or bundle does not connect
- Check that the Set and Get nodes use the same name.
- Check whether two Set nodes with that name are active at once. If so, rename one or mute an alternative branch.
- For a bundle, check that its declared slots are connected. An unconnected slot produces an error identifying the missing slot.
- If you are working without the extension’s interface assets, connect a single-value Get Variable manually. Bundle sockets depend on those assets for normal UI wiring.
A regional prompt produces an error
- Check that every tag follows the flat
|label: text|pattern; do not put a tag inside another tag. - Remove any literal
|characters that are not part of a tag. - Make sure the label requested by Annotation Segment or Annotation Segment Edit exists. An unknown-label error lists the available labels.
- When using the
newedit mode, choose a label that does not already exist.
A detection crop is larger than the size requested
In SEGS Set Crop Size, an exact crop grows when its target would cut off the detected box. It may also be limited by the image boundary. Check the detection size against your requested width and height, then increase the target dimensions if you need a consistent crop that contains every detection.
The sampler shows a noisy preview instead of a finished image
Check whether KSampler (Bundle) has vae_decode enabled. When it is disabled, the node does not decode a new finished image into the bundle, so its last latent preview remains visible.
Frequently asked questions
Do I have to use the checkpoint bundle to use the other nodes?
No. Nodes such as Color Match, Seed History, and the SEGS tools can serve separate purposes in an existing workflow.
Can I use one reference image for a batch of upscales?
Yes. Color Match applies a single reference frame to every frame of the processed-image batch. If both inputs contain multiple frames, their batch sizes must match.
Will SEGS BBox Overlay change what the Detailer receives?
No. It draws a preview image and passes its input SEGS through unchanged.
Are named variables saved between generations?
No. Their values flow through the workflow during a queue run; they are not persistent storage.
Does changing the seed-history length change the seeds themselves?
No. history_size controls how many used seeds appear in the rolling log, not how a seed is generated.
Learn More about ComfyUI-LukutarNodes
- ComfyUI is the application in which these nodes are used. Its project page is a starting point for understanding workflows, nodes, and standard connections.
- ComfyUI Impact Pack provides the SEGS detectors and Detailer workflows referenced by the overlay, crop-size, and
[LAB]wildcard features.
