Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > ComfyUI-TJ_NODE

ComfyUI Extension: ComfyUI-TJ_NODE

Repo Name

ComfyUI-TJ_NODE

Author
designloves2 (Account age: 437 days)
Nodes
View all nodes(2)
Latest Updated
2026-11-02
Github Stars
0.02K

How to Install ComfyUI-TJ_NODE

Install this extension via the ComfyUI Manager by searching for ComfyUI-TJ_NODE
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-TJ_NODE in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

ComfyUI-TJ_NODE Description

ComfyUI-TJ_NODE organizes large ComfyUI workflows with wireless routing, stage handoffs, and branch control, plus tools for prompts, reference assets, image processing, and generation records.

ComfyUI-TJ_NODE Introduction

ComfyUI-TJ_NODE is a toolkit for building and managing large ComfyUI workflows. Its main purpose is to help you keep a growing canvas understandable: fewer long wires, clearer groups of nodes, and more control over which parts of a workflow run.

The author recommends thinking of a workflow as a series of sections:

INPUT → GENERATION → EDIT → UPSCALE → PREVIEW → SAVE

You can keep ordinary connections short within each section and use the extension’s routing tools to pass information between sections. This is useful when, for example, you want to generate an image, try several edits, upscale only the version you like, and save the result without repeatedly running every earlier step.

The extension also includes tools for writing prompts, organizing reference assets, comparing LoRAs, processing images and video, and keeping a searchable record of generated images and their settings. You do not need to use all of these systems at once; individual nodes can serve a small workflow, while the routing tools become more valuable as your canvas grows.

Example ComfyUI-TJ_NODE workflow

How ComfyUI-TJ_NODE Works

A ComfyUI workflow is a set of connected steps. A connection carries something, such as an image, text prompt, or model output, from one node to another. ComfyUI-TJ_NODE adds ways to organize those steps and control how information moves between them.

  • Wireless routing reduces visual clutter. Set/Get-style tools let you pass compatible values across the canvas by name. Think of them as labeled handoff points: you can arrange related nodes together without drawing a long wire through every section.
  • Send and Send Point separate work into stages. A Send node can copy a result to a named Send Point when you click its button. The receiving section can then run on its own. For instance, you can generate an image in one group and later run an upscale group without generating the image again. This is a separate system from the extension’s Set/Get nodes.
  • Multi Switch chooses which route runs. An A/B switch lets you compare two approaches. Before a job is queued, the extension removes the unselected branch from that queued workflow, so its upstream nodes do not run. This matters when each branch would otherwise load a large model.
  • Pipes carry related settings together. PromptDB, for example, can send a group of saved settings through one pipe connection. A Bridge node can unpack those values near the nodes that need them. This keeps a canvas tidier without discarding the individual settings.

A practical approach is to start with clearly named sections, use short wires where they remain easy to follow, and add wireless handoffs where wires would otherwise cross the canvas.

ComfyUI-TJ_NODE Features

Prompt and scene planning

Prompt Studio (TJ) combines two prompt-writing tasks. With an image connected, its Auto mode works as Image to Prompt; without one, it works as Prompt Enhancer. You can also choose either mode explicitly. It supports embedded wireless Get/Set connections.

Prompt Enhancer (TJ) rewrites or refines a text prompt for a selected model_format. The listed formats include natural-language and tag-oriented prompts, KREA2 Prompt Enhance, and Qwen Image 2.1 text-to-image and edit formats. The Qwen edit format is designed to state what should change while preserving other attributes of the input image. Descriptive prose is produced in English, except where a format calls for dialogue or visible text to be preserved exactly.

Image to Prompt (TJ) uses a vision-capable language model to turn an image into a caption, tags, or another structured description. Its vision_task setting changes the kind of analysis requested; examples in the author’s documentation include pose analysis and content-quality checks.

Scene Maker (TJ) organizes prompts around visual beats, or individual moments in a scene. It supports translation among Korean, English, Japanese, and Chinese. With suggest_references enabled, it can sort needed references into images you provided, references you already have, and missing references. You can list existing assets in existing_references so they are not suggested again. For missing items, Scene Maker can provide a reference plan and prompts to help create them. Its @name references can be used with matching aliases in the Reference Asset Library. Scene Maker Result - pipe(TJ) carries the results onward and lets you collapse or resize its display sections.

Prompt Show & Locker (TJ) displays a text value and can hold it steady across workflow runs. Use it when you want to inspect a generated prompt, copy it, or keep a chosen version while testing other parts of the workflow.

References and video

Reference Asset Library lets you register reference images, video, and audio once and reuse them by ID, alias, or asset name. Its browser can add a single image, a set of 2–10 images, video, or audio. A “From Gallery” option integrates with STUDIO_ONE when that gallery is available. The documented attachment limits are nine images, three videos, and three audio items.

MiniMax H3 Reference / Image to Video (TJ) can interpret prompt references such as @12, @alias, or @asset-name and pass them to the underlying node in its expected reference format. This saves you from manually keeping reference numbers in sync with a prompt. The pack also provides separate MiniMax H3 canvas nodes for prompt queues, sequencing, audio lock, one-take sampling, output, latent continuation, and saving or loading latent checkpoints.

Video Resize (TJ) offers five ways to resize frames: long side, short side, ratio with a center crop, megapixels, or an exact width and height. Choose a side-based mode when you want to preserve a general size relationship; choose exact dimensions when a later step requires a particular frame size.

Images, sampling, and LoRAs

Resolution (TJ) offers nine aspect-ratio presets, including square, landscape, and portrait options, with eight representative resolutions for each. You can also enter a custom ratio or resolution, swap dimensions, choose a snap value of 8, 16, 32, or 64, and preview the resulting shape. This helps you choose dimensions before generating an image.

Enhanced KSampler (TJ) adds optional prompt-adherence controls for Krea2, Klein, and Z-Image. Select the architecture that matches your loaded model; the node shows the relevant controls and warns if the selection does not match. Turn enhance_enabled off to use standard KSampler behavior. Its Krea2 controls include text-chunk amplification and text_scale; the Klein and Z-Image modes use different architecture-specific methods, so a setting from one mode should not be assumed to have the same effect in another.

Multi Image Loader (TJ) helps you work with image batches. Its file picker supports sorting and folder bookmarks, and its FILENAMES output provides the original file names. Connect that output to Save With Original Names (TJ) when you want processed batch images saved under their source names. The save node also offers choices for output subfolder, extension, and overwrite or automatic numbering behavior.

Batch to MinimaxH3 (TJ) splits a batch into as many as 12 optional image outputs. Outputs beyond the batch size are empty rather than black images, which is useful for optional MiniMax H3 reference slots. Do not connect those potentially empty outputs directly to a node that requires an image, such as PreviewImage.

Skin Retouch (TJ) applies non-generative adjustments to skin tone, blemishes, shine, and texture while aiming to preserve the person’s identity. Although it was designed for a STUDIO_ONE MiniMax H3 post-processing chain, you can connect it independently in another image workflow.

RTX Deblur, RTX Denoise, and RTX VSR (TJ) provide separate image-processing filters. RTX VSR also offers short-edge, long-edge, and target-dimensions sizing; the target-dimensions mode includes a crop-anchor choice.

LoRA Analyzer has separate nodes for Krea2, Klein 4B, Klein 9B, and Z-Image LoRAs. Each shows block-level effects as you adjust strengths, offers one-click starting presets, and lets you toggle back to the original LoRA for comparison without losing your settings. Use the analyzer that matches the LoRA architecture; the extension warns about mismatches.

Workflow control and records

Multi Switch (TJ) supports up to 12 A/B groups. Global mode makes the groups share one selection; Per-Group mode lets each group choose independently. An unconnected selected input is skipped for that group without raising an error.

Send (TJ) and Send Point (TJ) pass a result between separately runnable workflow groups. Give a Send Point a unique point_name to create its transfer button on Send. Image, video, and text transfers have corresponding previews. Received values can be kept with the workflow and restored when it is reopened; values that cannot be serialized, such as some latent data, are session-only. Turning a Send Point’s output off stops that value from continuing into its downstream graph without deleting it.

PromptDBSave(TJ) records generated images and settings in an Excel workbook, with one row and thumbnail per image. It can collect settings from the workflow automatically, while directly connected inputs take priority. Set it to BYPASS to let images pass through without recording them. When supplied with STUDIO_ONE settings through json_input, it uses those JSON values for the corresponding model, seed, sampling, and prompt fields. A status area reports whether a run saved, skipped, or encountered an error.

PromptDBLoader(TJ) presents saved entries as a thumbnail gallery. You can select an entry for a later run, open its details to edit or delete it, search notes, change sorting, replace a thumbnail, or move an entry to another library. PromptDBBridge(TJ) unpacks the Loader’s pipe into individual fields near the nodes where you use them. Library controls also provide switching, import and export of the library list, and workbook backups.

Queue Loop (TJ) can use a count or index supplied by another node. Index LoRA Loader Counter (TJ) reads the number of active LoRAs in an Index LoRA Loader, excluding [none], so you can use that count for a loop without creating a circular graph connection.

TQD Score Estimate (TJ) is an experimental, test-stage tool for preparing scored image datasets for Krea2 LoRA training. It evaluates images one at a time with a local vision language model, records separate structure and detail scores, and saves images, captions, and scores together. It is aimed at dataset preparation rather than everyday image generation.

ComfyUI-TJ_NODE Models

ComfyUI-TJ_NODE is a collection of nodes, not a single image model. Several nodes use models or model families that you select separately. The context provided does not establish a complete list of required downloads or claim that model weights are included.

Model or backendWhere it is usedWhen you might choose it
GGUF / llama.cppLLM, Prompt Studio, Prompt Enhancer, Image to Prompt, and Scene Maker workflowsFor local prompt writing or image analysis with a compatible text or vision model. The author notes that a basic llama-cpp-python installation can be CPU-only and slow; see LLAMA_GPU_SETUP.md for the documented GPU guidance.
ComfyUI TextGenerateSupported prompt and LLM nodesWhen you already use a compatible ComfyUI text-generation setup.
Open Router or Connect CustomLLM (TJ) and the newer Scene Maker backend selectionWhen you want to use a configured service or custom connection rather than a local GGUF backend.
OllamaOllama LLM Loader (TJ)When you have an Ollama server running locally and want to use its available text or image-capable models. The documented loader limits connections to localhost by default.
Krea2, Klein, or Z-ImageEnhanced KSampler and matching LoRA Analyzer nodesChoose the setting or analyzer that matches the image model or LoRA you are actually using.
Qwen Image 2.1Text Encode Qwen Image 2.1 (TJ) and related prompt formatsWhen working in the corresponding text-to-image or image-editing workflow.
MiniMax H3Reference, image-to-video, and one-take canvas nodesWhen building a MiniMax H3 reference-driven or video workflow.
LTX 2.5LTX25 CLIP GGUF Loader (TJ)When your workflow needs the documented LTX 2.5 Gemma4 text-encoder GGUF support.
Qwen3-VL or Qwen2.5-VL GGUFExamples given for TQD Score Estimate (TJ)When testing the experimental local image-scoring workflow with a compatible vision language model.

A model_format in a prompt node is an output-writing format, not necessarily a model download. For example, selecting a Qwen Image 2.1 prompt format changes how the prompt is written; it does not by itself load Qwen Image 2.1.

What's New with ComfyUI-TJ_NODE

The original README identifies v2.27.0 as its headline version, but also warns that some listed additions were pre-release or still being tested. This is a summary of the changes explicitly described in the provided README, not a complete release history. Check CHANGELOG.md for exact release tags and fuller change details.

VersionDocumented changeWhy it matters to you
v2.27.0LLM (TJ) gained model_format, aesthetic, and vision_task prompt choices, plus progress reporting for GGUF/llama.cpp runs.You get more control over the requested text and clearer feedback during a local generation.
v2.23.0PromptDBSave gained json_input and a visible save-status area. Scene Maker gained reference suggestions and reference-plan outputs; its result display became collapsible and resizable.You can reuse settings supplied by STUDIO_ONE, see whether a record was saved, and plan missing shot references more easily.
v2.22.0Scene Maker adopted LLM (TJ)-style backend selection and connection controls.Its model connection choices are more consistent with the general LLM node.
v2.18.0–v2.20.0Reference Asset Library and MiniMax H3 reference tools were added and expanded, including browser add options and storage and attachment-limit changes.You can reuse named visual and audio references across shots with less manual renumbering.
v2.18.0LLM (TJ), Text Encode Qwen Image 2.1 (TJ), Qwen Image 2.1 prompt formats, and additional RTX VSR sizing choices were documented.These add more prompt, model-workflow, and output-size options.
v2.17.0Skin Retouch (TJ) was added.You can make targeted skin adjustments without using a generative edit for that step.
v2.16.0Several causes of empty or truncated GGUF-generated MiniMax H3 briefs were fixed.Prompt-writing results should be less likely to stop before the useful answer.
v2.15.0–v2.15.1Separate RTX processing nodes, KREA2 UNET/CLIP GGUF loaders, and Video Resize (TJ) were added.You can apply these processing and sizing steps individually.
v2.14.0–v2.14.1LTX 2.5 text-encoder GGUF support was added, followed by a fix for a related tensor crash.LTX 2.5 workflows gained documented loader support and a stability fix.
v2.12.0–v2.13.0MiniMax H3 one-take functions were separated into canvas nodes.You can build that pipeline from individual nodes rather than relying on an all-in-one studio node.
v2.11.0Multi Switch, Batch to MinimaxH3, PromptDB gallery and library improvements, and an Enhanced KSampler Krea2 compatibility fix were documented.These changes improve branch control, optional reference inputs, record management, and sampling compatibility.
v2.9.2Shortcut Launcher’s local-open endpoint was restricted to local, same-origin requests.This addresses the documented risk of an unwanted remote request opening a local file or program.
v2.7.0LLM-node VRAM handling was improved.The author reports better behavior when rerunning vision tasks and freeing GPU memory after use.

The README also describes improvements to Multi Image Loader, Queue Loop, Resolution, Smart Show, and LoRA Analyzer without assigning every change a version in the supplied text.

Troubleshooting ComfyUI-TJ_NODE

A local LLM is very slow, or a newer model does not work.

  1. Check whether the workflow is using the GGUF/llama.cpp backend.
  2. Review LLAMA_GPU_SETUP.md. The author notes that a basic llama-cpp-python installation may be CPU-only.
  3. Check that the model you selected is compatible with your backend. The provided context does not establish one universal model that works for every LLM task.

My Enhanced KSampler controls do not seem to affect the result.

  1. Check that enhance_enabled is on.
  2. Set enhance_arch to the architecture of the loaded model: Krea2, Klein, or Z-Image.
  3. Look for an architecture-mismatch warning. When the selection does not match, the node falls back to standard behavior.
  4. If a Krea2 enhancement began failing after a ComfyUI update, consult the v2.11.0 fix described in CHANGELOG.md.

An optional Batch to MinimaxH3 output causes an error.

  1. Count the images in the batch.
  2. Check whether the failing connection comes from an output beyond that count.
  3. Connect these optional outputs to nodes that accept an empty slot. A required-image input such as PreviewImage cannot use the empty None value.

PromptDBSave did not add a row.

  1. Read the status displayed beneath the node; it distinguishes a successful save, a skipped save, and an error.
  2. Confirm that the node is set to SAVE, not BYPASS.
  3. If json_input is connected, check that it contains settings rather than an empty value or {}.
  4. Check which inputs are connected directly. Connected values take priority over automatically extracted settings.

A reference is not appearing where I expected.

  1. Check the asset’s ID, alias, or name in the Reference Asset Library.
  2. Make sure the prompt uses the intended @ID, @alias, or @asset-name reference.
  3. In Scene Maker, list assets you already have in existing_references, one per line, so they are treated as existing rather than suggested as missing.
  4. Check the documented attachment limits: nine images, three videos, and three audio items.

A Send button does not appear, or the receiving group has no value.

  1. Give the Send Point a unique, nonempty point_name; an empty name does not create the button.
  2. Click the corresponding button on Send to copy the result. This transfer is separate from queue execution.
  3. Check the Send Point output toggle if the value is present but is not reaching later nodes.
  4. Bear in mind that types the workflow cannot serialize are retained only for the current session.

Frequently asked questions

  • Do I need every feature to organize my workflow? No. Start with the nodes that solve a specific problem, such as a long connection, repeated reference setup, or lost generation settings.
  • Does Multi Switch save work compared with bypassing a branch? Its documented behavior is to remove the unselected upstream branch from the queued prompt, so that branch does not execute.
  • Does PromptDB automatically know all my generation settings? It attempts to extract settings from the connected workflow. Direct inputs take priority; settings it cannot find may retain the node’s default values, so check important records.
  • Can I reopen a workflow and use a Send Point’s previous result? Supported stored values can be restored with the workflow. Non-serializable values, including some latent data, are session-only.
  • Is TQD Score Estimate a general image-quality rating tool? The documented purpose is experimental scoring for a Krea2 LoRA training dataset. Its structure and detail scores serve that workflow.

Learn More about ComfyUI-TJ_NODE

  • CHANGELOG.md — check exact version tags, fixes, and release details before relying on a feature described as pre-release or in testing.
  • LLAMA_GPU_SETUP.md — the author’s guidance for GPU-accelerated local GGUF/llama.cpp prompt workflows.
  • THIRD_PARTY_LICENSES.md — licensing information for incorporated third-party work.
  • comfyUI-Realtime-Lora — the project the author credits as inspiration for the LoRA Analyzer idea; the README says the analyzer itself was independently implemented.
  • musubi-tuner-gui, tj-custom branch — related Krea2 LoRA training work using datasets made with the experimental TQD scoring node.

ComfyUI-TJ_NODE Related Nodes

RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.

ComfyUI-TJ_NODE detailed guide | ComfyUI