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 > Nukun_ComfyUI_Nodes

ComfyUI Extension: Nukun_ComfyUI_Nodes

Repo Name

Nukun_ComfyUI_Nodes

Author
onekosl (Account age: 1091 days)
Nodes
View all nodes(1)
Latest Updated
2026-08-29
Github Stars
0.0K

How to Install Nukun_ComfyUI_Nodes

Install this extension via the ComfyUI Manager by searching for Nukun_ComfyUI_Nodes
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter Nukun_ComfyUI_Nodes 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

Nukun_ComfyUI_Nodes Description

Nukun_ComfyUI_Nodes adds ComfyUI nodes for image, video, and audio workflows, including model cycling, prompt refinement, regional conditioning, noise sampling, high-resolution fixes, and media tools.

Nukun_ComfyUI_Nodes Introduction

Nukun_ComfyUI_Nodes is a collection of ComfyUI nodes for building image, video, and audio workflows. It brings together checkpoint loaders, prompt-writing helpers, regional masks, conditioning tools, noise samplers, model patches, and high-resolution refinement tools.

If you make many variations of an image, the cycler nodes can move through models, prompts, or noise profiles for you. If you want different parts of an image to follow different instructions, the regional nodes help you define those areas. If you are experimenting with a look, the sampler and conditioning nodes give you more ways to adjust it while keeping useful reports and seeds for comparison.

The author also includes tools for local Ollama-assisted writing and captioning, ACE-Step song workflows, and Wan 2.2 video creation. You can use the parts that suit your workflow; the optional integrations are only needed for their particular nodes.

How Nukun_ComfyUI_Nodes Works

ComfyUI workflows pass information between connected nodes. This extension adds nodes that prepare that information, transform it, or make repeated runs easier to manage:

  • Loaders choose a checkpoint or diffusion model and pass its model, text encoder, VAE, and useful filename information to later nodes.
  • Text helpers assemble prompts from your writing, vocabulary files, or a local Ollama model.
  • Conditioning nodes turn prompt instructions into information the image model uses during generation. Some help balance or inspect that information; others apply it to particular image regions.
  • Masks mark where regional instructions should apply, much like selecting an area in an image editor.
  • Samplers guide the generation process. Their noise settings change the starting pattern from which an image develops.
  • Workflow helpers keep track of settings, frames, and segments when a project involves several passes or video extensions.

For example, you could cycle through several checkpoints using the same prompt, save each result under a model-specific name, and compare them. Or you could write one full-scene prompt, add separate instructions for the left and right sides, preview the masks, and send the resulting conditioning to a sampler.

Some nodes are intended for experimentation. When testing one, keep the prompt and seed consistent where possible and change one setting at a time. That makes it easier to tell which change affected the result.

Nukun_ComfyUI_Nodes Features

Model and checkpoint cycling

Checkpoint Cycler Loader loads a selected checkpoint and outputs its MODEL, CLIP, and VAE, along with its name and folder. Its checkpoint selector supports ComfyUI’s after-generation controls, so you can increment, decrement, or randomize the selection between queued images.

Checkpoint + VAE Cycler Loader adds a VAE choice. Select checkpoint to use the VAE included with the checkpoint, or choose an external VAE without changing the model and CLIP outputs.

Checkpoint Pair Cycler Loader tests pairs of checkpoints from one selected subfolder. It includes same-model pairs and both orders of mixed pairs, so A × B and B × A are separate tests. An inclusive model-index range lets you divide a large comparison into smaller runs. Its combined model name is useful in save prefixes.

The 4-Prompt Model Cycler Loader and 4-Prompt Checkpoint Cycler Loader run four prompt positions for each model or checkpoint in an exact folder. The model version keeps CLIP and VAE fixed; the checkpoint version loads the selected checkpoint’s CLIP and embedded VAE. Their seed modes let you keep a seed fixed, advance it once per four-prompt group, or choose a new seed for each group through the ComfyUI frontend. Empty prompt fields still occupy positions in the cycle.

Vocabulary and prompt building

Incrementing Int to String turns an incrementable number into text. Its minimum and maximum values define a wrapping range, which is useful for labels or other numbered workflow inputs.

Random Vocab String List selects words from a vocabulary file. Its seeded shuffle-bag behavior helps you move through non-overlapping groups of words before reshuffling. You can choose how many words to output and append them to an existing string through chain. Bundled vocabularies cover subjects, places, actions, composition terms, art styles, and other categories. For a single style direction, select resources/visual_art_styles.csv and set amount = 1.

Multi Vocab String List combines four vocabulary selections. Each slot has its own word count and index, so you can vary, for example, a subject while holding a setting or style selection steady.

MiniMax H3 Prompt Builder organizes text into up to six video-prompt sections: Scene, Character, Action, Camera, Visual Style, and Audio. Each section can combine your fixed description with sampled vocabulary phrases; amount = 0 leaves it text-only. Empty sections are omitted. When you provide spoken_dialogue, the node includes the exact line in both the action and audio directions. See the vocabulary and H3 prompt guide for examples.

Local Ollama writing and captioning

Ollama Prompt Refiner turns connected source words and instructions into prompts for a selected target profile. Available profiles are pony_v6, illustrious, pony_v7, z_image, anima, wan2_2_video, and krea2. You can supply a style_anchor for details that should survive the rewrite. Supported profiles also accept left, right, top, and bottom ideas as composition guidance rather than exact geometric boundaries.

The refiner provides positive and negative prompts, split prompt components, a processing report, and JSON outputs for planning or review stages. prompt_mode = strict stays close to the source; creative permits compatible elaboration while retaining key requirements. More involved pipeline modes add planning and, optionally, review, but require additional Ollama requests. Fallback settings determine whether a failed response stops the workflow or preserves usable source material. Older saved workflows may need their former profile-specific output connections moved to the current positive and negative outputs.

Ollama Video Prompt Refiner accepts separate Scene, Character, Action, Camera, Visual Style, and Audio source fields. It can prepare a structured MiniMax H3 prompt or a continuous Wan 2.2 visual prompt. Its faithful, balanced, and cinematic creativity choices range from restrained editing to stronger production direction. It preserves supplied dialogue and reports processing decisions. For a vocabulary-driven H3 workflow, connect the matching section outputs of the MiniMax H3 Prompt Builder to this node.

Ollama Vision Captioner describes the first image in a ComfyUI image input and outputs a caption, tags, a text seed for a refiner, a report, and detail-pass text for HiResFix. Choose natural_caption for readable sentences, danbooru_tags for tag-style output, pony_source for concise factual words, or refiner_seed for richer input to the Prompt Refiner. It requires a vision-capable Ollama model; a text-only model cannot reliably describe the supplied image.

These Ollama nodes use a local Ollama service. The model selected in Ollama supplies the writing or vision capability; the extension does not include an Ollama language or vision model of its own.

Conditioning and regional composition

T5 Equal-Length Prompt Balancer encodes positive and negative prompts at a shared effective length without cutting off a longer prompt. It reports the raw token counts and chosen length. Its default target is 1024; on a low-memory system, target = 0 avoids padding to that fixed target.

T5 Sculpt Equal-Length Prompt Balancer adds experimental token-vector adjustments to the equal-length approach. CLIP Sculpt Text Encode provides related sculpting for SD1/SDXL CLIP workflows. Start with the ordinary equal-length encoder or ordinary CLIP encoding as a comparison, then increase sculpting intensity only if the results are useful to you.

Conditioning Analyzer passes conditioning through unchanged and reports properties such as shape, token count, and unusual numeric values. Conditioning Adjust applies selected, repeatable presets, including soft_balance, literal_detail, contrast_pop, and negative_tamer. Its neutral_report_only setting is useful when you want information without an adjustment.

Regional Prompt Encoder combines a base prompt with two or three regional prompts and supplies separate regional conditionings, a base conditioning, and a conditioning for a later detail pass. Regional Sculpt Prompt Encoder performs similar prompt assembly with experimental sculpting on the regional and HiRes outputs, while leaving the base conditioning unsculpted.

For placement, choose the tool that matches your composition:

  • Split Masks makes horizontal or vertical two- or three-part masks. overlap expands neighboring areas around a boundary.
  • Regional Rect Masks makes freely positioned rectangular masks using proportional x, y, w, and h controls. soft_edge blurs their edges, and the node includes a draggable rectangle editor.
  • Native Regional Split Conditioning and Native Regional Rect Conditioning combine regional conditioning and masks using ComfyUI’s core conditioning features.
  • Regional Split Regions outputs regions compatible with an A8R8 Attention Couple workflow. The node itself does not require A8R8.
  • DenseDiffusion Split Apply and DenseDiffusion Rect Apply provide optional DenseDiffusion-based regional workflows. The rectangular version is useful when areas should not run across the full width or height of an image.

Preview regional masks before a long generation. In a two-region setup, the third mask is empty and third-region inputs are ignored.

Noise and sampling

Advanced Noise Sampler combines initial-noise generation with advanced sampling. Its gaussian noise, auto device, and noise_strength = 1.0 provide a ComfyUI-like baseline. Other profiles offer different starting patterns for experiments; stronger profiles are generally easier to judge at reduced strength or in a later pass. Setting add_noise off uses zero initial noise.

Illustrious Noise Sampler and Pony V7 Noise Sampler provide composite-noise choices oriented toward those workflows. The Illustrious node offers modes such as balanced, texture, composition, and wild. The Pony v7 node includes gentler first- and second-stage profiles.

Universal Noise Sampler collects the basic and composite profiles in one node and is the author’s preferred combined sampler for new Nukun workflows. Universal Noise Sampler Advanced adds step-range controls for partial passes and handoffs. If continuing an earlier pass without adding fresh noise, start the next range at the preceding range’s ending step rather than one step later.

Noise Profile Cycler steps through selected groups of profiles for systematic comparisons. UNet Block Noise Patch is a separate experimental tool that adds controlled noise within model block groups; its default strengths are zero.

SPEED Sampler offers an alternative sampler that begins at a lower latent resolution and expands during denoising. It may reduce runtime in suitable workflows, but inspect the output for artifacts. The node is based on the SPEED implementation and does not require a separate ComfyUI-SPEED package.

All Nukun sampler nodes offer a preview_method choice. default follows ComfyUI’s queue setting, latent2rgb provides lightweight step previews, taesd uses compatible preview assets when available, and none disables previews.

High-resolution refinement, audio, and video

HiResFix Tiled combines model upscaling and tiled image refinement in one step. It uses the supplied upscale model’s native scale and requires ComfyUI_UltimateSDUpscale. Tiled HiRes Fix Advanced requires ComfyUI_TiledKSampler.

For music workflows, ACE Song Variation Director uses local Ollama to develop a variation plan from ACE-Step tags and sectioned lyrics. ACE Song Timeline Conditioning maps song sections onto time-regional ACE-Step 1.5 conditioning and produces a timeline report. Audio Timeline Mixer 5 places and mixes up to five audio inputs with controls for timing, gain, mute, fades, and output peaks.

For Wan 2.2 video, Wan 2.2 Video Settings turns duration and frame rate into a valid frame count and offers resolution presets. Wan 2.2 TI2V Latent helps switch between text-to-video and image-to-video setups without rewiring those paths. Wan 2.2 Run Manifest records prompts, seeds, dimensions, and sampling settings for a run. Continuation helpers plan extensions, record segment information, store numbered frames, reload the current end frame, and assemble a saved frame sequence into an MP4.

Examples of Wan 2.2 start, extension, and final-assembly workflows are in examples/wan22/. The Wan 2.2 workflow analysis and Anima-to-Wan workflow guide provide more detail on the documented workflows.

Nukun_ComfyUI_Nodes Models

This extension mainly provides nodes and settings, not a bundled collection of image-generation models. The relevant model choices depend on the workflow:

Model or profile choiceWhat it is used for
Checkpoints and diffusion modelsChosen by the loader and cycler nodes for image-generation runs and comparisons.
External VAEOptionally replaces a checkpoint loader’s VAE output while retaining its model and CLIP outputs.
Ollama language modelWrites or refines prompts in the Ollama prompt and video nodes; it must be available to your local Ollama service.
Ollama vision modelDescribes an input image for Ollama Vision Captioner. Choose a model with image-understanding support.
Ollama Prompt Refiner target profilespony_v6, illustrious, pony_v7, z_image, anima, wan2_2_video, and krea2 shape the prompt for different downstream workflows; they are output formats, not models included with the extension.
MiniMax H3 and Wan 2.2 video profilesTell Ollama Video Prompt Refiner whether to produce six structured H3 sections or one continuous Wan visual shot.
Upscale modelSupplied to HiResFix Tiled to enlarge an image before tiled refinement.

The Ollama nodes list models reported by the selected local Ollama service. The original documentation names autoren-darkidol-llama-3-1-8b:latest as the Prompt Refiner’s default and user-v4/joycaption-beta as the Vision Captioner’s default; their availability depends on what is present in your Ollama setup.

Troubleshooting Nukun_ComfyUI_Nodes

A workflow says a node is missing.

  1. Identify the missing node by name.
  2. If it is a DenseDiffusion Apply node, check whether the optional comfyui_densediffusion package is available.
  3. For HiResFix Tiled, check ComfyUI_UltimateSDUpscale; for Tiled HiRes Fix Advanced, check ComfyUI_TiledKSampler.
  4. If an older Prompt Refiner workflow loads but its prompt outputs are disconnected, reconnect them to the current positive and negative outputs.

Ollama prompt generation fails or takes too long.

  1. Check that your local Ollama service is reachable at the URL shown in the node.
  2. Confirm that the selected model is available at that Ollama endpoint.
  3. Inspect the node’s report output for the failed stage, validation issue, or fallback used.
  4. For a simpler test, use a single-stage pipeline before adding planning or review. Those additional stages make more requests and take longer.
  5. Select a fallback mode based on your needs: strict stops on invalid output, while supported adaptive or continue paths can retain usable source material.

A vision caption does not describe the image. Confirm that you selected a vision-capable Ollama model, not a text-only one. If you supplied a batch, note that the Vision Captioner captions only the first image and mentions additional images in its report.

Regional prompts affect the wrong area. Preview the masks first. For rectangles, adjust x, y, w, and h in the node’s editor; for splits, check the orientation and boundary settings. Use soft_edge or overlap when a hard transition is undesirable.

A sampling experiment changes more than expected. Return to gaussian, auto, and strength 1.0 for a baseline. Keep the seed and other generation settings consistent while trying another profile. If using ranged passes, check the handoff steps and whether later passes add new noise.

T5 balancing uses more memory than expected. Try target = 0 on an equal-length balancer. This uses the longer actual prompt length instead of padding both prompts to the default 1024-token target.

Frequently asked questions

Do I need every optional custom-node package? No. The optional packages listed above are needed only for their respective integration nodes.

Does Regional Split Regions require A8R8? Not by itself. Its region output is intended for workflows that connect to A8R8 Attention Couple.

Are noise profiles the same as image models? No. They change the sampler’s initial noise; your selected checkpoint or diffusion model still performs the generation.

Can I compare settings fairly? Use the same prompt, model, and seed where the workflow allows it, then change one setting at a time. The cyclers, seed outputs, reports, and run manifests can help you keep track of those tests.

Learn More about Nukun_ComfyUI_Nodes

  • Vocabulary String List Guide: practical guidance for the vocabulary nodes and a complete MiniMax H3 prompt example.
  • Wan 2.2 Workflow Analysis: details on the Wan toolkit and workflow comparisons.
  • Anima-to-Wan Workflow Guide: controls and documented smoke-test details for the Anima keyframe-to-video workflow.
  • examples/wan22/: example workflows for starting, extending, and assembling segmented Wan 2.2 video.
  • SPEED project: background on the sampling approach used by the SPEED Sampler node.

Nukun_ComfyUI_Nodes 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.

Nukun_ComfyUI_Nodes detailed guide | ComfyUI