Deterministic styler-pipeline nodes designed to stay reproducible across runs.
AI-assisted style selection that queries an LLM by category and returns ranked style candidates with scores.
Manual style browsing and selection through the Browser workflow with category navigation.
Per-category Randomize mode in Browse to choose a different style from selected categories on each workflow run.
Dynamic Styler that applies style safely to existing conditioning.
Classic dropdown-based Advanced Styler node for category-by-category control in the graph.
Compatible with ControlNet workflows, including OpenPose-guided use cases.

Installation

Requirements

ComfyUI (recent build)
Python 3.10+

Steps

Clone this repository into ComfyUI/custom_nodes/.
Restart ComfyUI.
Confirm nodes appear under Styler Pipeline/.

Nodes

Styler Pipeline

At a glance:

Main node for everyday styling with the Edit panel.
Deterministic and reproducible because selections are stored in internal JSON.

Inputs:

positive (CONDITIONING, required)
negative (CONDITIONING, required)
clip (CLIP, required to apply styles)
strength (FLOAT, default 1.0)
redundancy (INT, default 1)
selected_styles_json (STRING, internal UI state)

Outputs:

positive (CONDITIONING)
negative (CONDITIONING)

Behavior notes:

Uses selected styles to encode additional style conditioning, then merges it into existing conditioning.
Click Edit to manage category/style picks in one panel and write them into internal JSON.

Strength and Redundancy Guidelines

strength controls how strongly the selected styles steer the generation. Different checkpoints/models are not equally influenceable: some apply styles strongly at low strength, while others are more resistant.

If a model is resistant, increasing strength can help. But beyond a point, it usually hurts output quality; around ~1.3+ is commonly where degradation becomes noticeable because it is effectively "shouting" the instruction at KSampler.

redundancy literally repeats the selected styles multiple times to increase their weight. This can improve style adherence, but increasing redundancy too much can harm composition.

Safe starting point: strength = 1.0, redundancy = 1.
Typical tuning: increase strength gradually in small steps first.
In most cases, keep redundancy at 2 or below.

AI Styler module: Describe the look you want, and AI Styler asks an LLM to automatically suggest the best matching styles by category. It supports major API providers (OpenAI, Anthropic, Groq, Gemini, Hugging Face), and it also supports Ollama (Local) so you can run in offline/no-internet environments. In the image below, you are looking at the AI Styler tab opened from Edit, where prompt-driven suggestions are generated and applied.

Browser module: If you prefer not to use AI Styler, the Browse module lets you choose styles manually and keep more control over selected styles. In the image below, you are looking at the Browser tab in the same panel, where categories and styles are selected manually.

Use the Randomize chip in any category to pick a random style from that category each time the workflow is queued.

Note: For reproducibility, generated images store the resolved style in their embedded workflow metadata. When you reload a workflow from a generated image, Randomize will no longer be active for that category; the workflow will instead use the exact style that was randomly chosen for that generation. Selecting a concrete style also disables Randomize for that category.

Editor module: Editor lets you view styles loaded from the category JSON files (data/*.json). Editing tools are currently under construction and will be available soon (limited AI token budget at the moment).

Note

Because selected styles are saved inside node data, the same workflow stays reproducible even if you add/remove categories and styles defined in the JSON style files, as long as you keep the styles you originally selected.

Styler Pipeline (Single)

Apply one style at a time by manually choosing category and style.

Inputs:

positive (CONDITIONING, required)
negative (CONDITIONING, required)
category (STRING/dropdown, required)
style (STRING/dropdown, required)
clip (CLIP, required to apply styles)
strength (FLOAT, default 1.0)
redundancy (INT, default 1)

Outputs:

positive (CONDITIONING)
negative (CONDITIONING)
style (STRING)

Styler Pipeline (By Index) + Index Iterator

Use this pair for deterministic style sweeps, avoiding manual selection of styles by using an incrementing index to apply styles one by one from a selected category. Styler Pipeline (By Index) applies one style from a selected category using style_index, and Index Iterator provides an incrementing index each run.

Inputs:

Styler Pipeline (By Index): positive, negative, category, style_index, clip, strength, redundancy, prepend_timestamp.
Index Iterator: reset, start.

Outputs:

Styler Pipeline (By Index): positive, negative, style.
Index Iterator: index (INT).

Usage: Connect your positive and negative conditioning and connect clip correctly, then select a category in Styler Pipeline (By Index) and feed its style_index from Index Iterator's index output. On each workflow run, Index Iterator increments from the configured start value so the next style in that category is applied automatically. This is useful for quickly testing many styles without manually changing selections each time before sending the resulting conditioning to downstream nodes such as KSampler.

Advanced Styler Pipeline

Classic menu-based styler with direct dropdowns for each JSON category.

At a glance:

Good when you want category-by-category dropdown control in the graph.
Explicitly appends style conditioning to your current positive/negative paths.
Faster to scan than opening the panel when you already know your category picks.

Inputs:

positive (CONDITIONING, required)
negative (CONDITIONING, required)
clip (CLIP, optional input, required to apply style encoding)
strength (FLOAT, default 1.0)
redundancy (INT, default 1)
Style dropdowns loaded from data/*.json

Outputs:

positive (CONDITIONING)
negative (CONDITIONING)

Usage: Connect incoming positive and negative conditioning to this node, connect clip, and choose the style dropdowns you want from each category to layer your look. The node augments your existing conditioning rather than replacing it, so adjust strength and redundancy as needed for balance. Connect the output positive and negative conditioning to downstream nodes such as KSampler for generation.

LLM Setup

AI Styler uses the Provider and Model you choose in the UI. Open Edit and use the AI Styler tab to select a Provider first, then select a Model for that provider.

Cloud API providers

Cloud API providers (OpenAI, Anthropic, Google Gemini, Hugging Face, Groq, etc.) are queried via their API. Select the provider and model you want to use in the AI Styler tab, then paste your API key or token in the token field before running suggestions. Before using a cloud provider, click Refresh to fetch the latest model list.

Provider notes (subject to provider policy and may change):

Hugging Face — offers models with free-tier access depending on the model and provider.
Groq — often offers a free tier; check current policy.
OpenAI, Google Gemini, Anthropic — typically require billing enabled for API usage.

Warning

OpenAI API could not be tested because billing could not be activated using prepaid cards. If you encounter an error while using OpenAI, please open a GitHub issue with detailed error information so it can be fixed as soon as possible.

The API key or token is used for the current run only and is not saved by the plugin; but you can save in your browser's Password Manager by using the Save token button provided.

Ollama models (Local + Cloud)

Ollama is a free desktop app that lets you run LLMs completely offline on your own hardware. Once signed in to a free Ollama account, you can also use Ollama Cloud models without downloading them locally.

Tip

Ollama never requires an API key — not for local models and not for cloud models. Cloud models only require a free Ollama account sign-in inside the Ollama app.

Getting Ollama models to show up:

After installing Ollama, the AI Styler may list no models until you activate one in the Ollama app:

Open the Ollama desktop app and keep it running (minimize is fine; do not close it).
In the Ollama app, select the model you want to use:
- Local model: choose a model to download to your machine. gemma3:4b is a good starter — lighter and faster than most.
- Cloud model: sign in to your free Ollama account inside the app, then select a cloud model.
Send any short message in the Ollama app (e.g., "test") to activate the selected model.
Return to AI Styler and click Refresh; the model should now appear in the model dropdown.

Warning

It's strongly recommended not querying local Ollama models while a ComfyUI workflow is running. Doing so can severely overload shared GPU/CPU resources, make your system very slow and unstable. Whenever possible, prefer a cloud provider, which is typically faster and more efficient. If you still want local Ollama, start with a small model like gemma3:4b before trying larger ones.

Troubleshooting (Ollama local):

No local models listed:
- Send any message to an Ollama local model in the Ollama app to initialize it.
- Confirm Ollama is running and reachable at http://127.0.0.1:11434.
Status shows "Not connected":
- Restart Ollama, then reopen AI Styler.
- Check that local firewall/security software is not blocking localhost port 11434.
Ollama not running:
- Start the app (Windows/macOS) or run ollama serve (Linux).

AI Prompts

Keep prompts short and specific. Describe the visual direction, not a full story.

What to include

Genre/style: sci-fi, noir, anime, fantasy, etc.
Mood: tense, cozy, melancholic, energetic.
Lighting: soft, practical, cinematic rim light, harsh noon sun.
Time of day: dawn, golden hour, night, overcast afternoon.
Environment: alley, spaceship interior, forest, classroom, rooftop.

What to avoid

Over-long prompts with too many competing ideas.
Contradictory directions in one sentence (for example: "dark night scene with bright midday sun").

Using the returned suggestions

Start by keeping 1-2 strong categories that best match your goal.
Generate/test, then refine with a small number of extra categories.
Avoid stacking conflicting categories at once; add changes incrementally.

Advanced JSON

For advanced users only. JSON editing is currently the only way to modify styles; a visual Editor UI is planned for a future version. Built-in prompts were AI-refined but not exhaustively tested — some may need small manual tweaks.

To keep your own styles safe during plugin updates, place personal JSON files here:

ComfyUI/user/default/styler_pipeline/

Put one or more *.json files in that folder, using the same JSON shape shown below.
To add styles to an existing category, use the same filename as the bundled category, such as mood.json.
To create a new category, use a new filename, such as cinematic_portrait.json.
Bundled styles load first, then your personal styles are added. If your file contains a style with the same name in the same category, your version is used.

Because this folder is stored under ComfyUI/user/default/ instead of the plugin's bundled data/ folder, your custom styles are not replaced when the plugin is updated. For compatibility, files already placed under the earlier ComfyUI/user/styler_pipeline/*.json path are still loaded for now.

Reproducibility note: Existing workflows stay reproducible as long as the style entries they reference are not renamed or removed. If a style used by an older workflow is renamed or deleted, that workflow will no longer find its style definition and will not reproduce the same result.

Keep custom *.json style files consistent so styler nodes stay predictable.

JSON shape

[
  {
    "name": "style name",
    "prompt": "style description, {prompt}, token1, token2, token3",
    "negative_prompt": ""
  }
]

Required keys per item:

name (string)
prompt (string)
negative_prompt (string, can be empty)

Practical rules

Prefer concrete visual language over abstract quality tags.
Keep prompts concise and visually descriptive.
Keep naming user-friendly and easy to browse.
Maintain strict valid JSON (no comments, no trailing commas).
Avoid words that models commonly interpret as physical objects. Some nouns trigger literal object rendering even when the intent is a color or hairstyle. For example, amber-toned can cause the model to draw amber stones instead of a warm golden color; crown braids can cause a literal crown to appear. The safest fix is to remove the triggering word entirely and describe the intent with different vocabulary — e.g., instead of "amber-toned" use "warm golden hue"; instead of "crown braids" use "intricate braided updo".

Tip

If a style prompt causes an unexpected object to appear in outputs, a literal trigger word is likely the cause. Common examples: amber-toned (renders amber stones) and crown braids (renders a literal crown).

Support

Why Your Support Matters

This plugin is developed and maintained independently, with regular use of paid AI agents to speed up debugging, testing, and quality-of-life improvements. If you find it useful, financial support helps keep development moving steadily.

Your contribution helps:

Fund AI tooling for faster fixes and new features
Cover ongoing maintenance and compatibility work across ComfyUI updates
Prevent development slowdowns when usage limits are reached

Tip

Not donating? A GitHub star ⭐ still helps a lot by improving visibility and helping more users

💙 Support This Project

Prefer scanning? Show QR codes

Ko-fi

PayPal

USDC (Arbitrum) ⚠️

Show USDC address

0xe36a336fC6cc9Daae657b4A380dA492AB9601e73

Warning

Send USDC on Arbitrum One only. Transfers sent on any other network will not arrive and may be permanently lost.

Gallery

Sample Workflow

Click the image below to open the full workflow example: You can also drag and drop this workflow image into ComfyUI to open/import it. This example workflow uses ControlNet for OpenPose via an OpenPose Studio node.

Sample Images

Note

All demo images below use the same model, the same LoRA, the same base prompt, and the same seed. The only difference is the styles applied by the Styler Pipeline node.

Image	Styles used
	- Baseline: Styler not applied - Generation settings (shared): - Resolution: `1024×1344` - Seed: `717891937617865` - Steps: `25` - CFG: `4` - Sampler: `dpmpp_2m_sde` - Scheduler: `karras` - Denoise: `1.0` - Checkpoint: `yiffInHell_yihXXXTended.safetensors` - LoRA: `inuyasha_ilxl.safetensors` - ControlNet: `illustriousXL_v10.safetensors`
	- aesthetic: `Enchanted Forest` - atmosphere: `Neon/Bioluminescent Glow` - environment: `Nature/bamboo forest` - filter: `BlueHour` - lighting: `Bioluminescent Organic` - mood: `Enchanted` - timeofday: `Twilight` - face: `Raised Eyebrow` - hair: `Color combo silver and cyan` - clothing_style: `Iridescent` - depth: `Soft Focus` - clothing: `Specialty/fantasy outfit`
	- aesthetic: `Rustic` - atmosphere: `Melancholic/Cold Overcast` - environment: `Historical/medieval village` - filter: `BlueHour` - lighting: `Overcast Diffusion` - mood: `Bleak` - timeofday: `Midday` - face: `Serious` - hair: `Silver white hair` - clothing_style: `Denim Fabric` - depth: `Deep Focus` - clothing: `Historical/viking raider`
	- aesthetic: `Dark Fantasy` - atmosphere: `Dark/Night Ambient` - environment: `Outdoor/temple hill overlook` - filter: `Soft` - lighting: `Soft General` - mood: `Meditative` - timeofday: `Midnight` - face: `Worried` - hair: `Long wavy hair` - depth: `Ultra Sharp` - rendering: `Semi-Realistic` - clothing: `Medieval/monk robe`
	- aesthetic: `Cyberpunk` - atmosphere: `Dark/Night Ambient` - environment: `Asian/japanese neon alley` - filter: `Neon` - lighting: `Multi-Source Complex` - mood: `Gloomy` - timeofday: `Midnight` - face: `Skeptical` - hair: `High ponytail` - clothing_style: `Neon Accents` - depth: `Selective Focus` - rendering: `Anime Style` - clothing: `SciFi/cyberpunk streetwear`

Best practices for reliable results:

Styler influence varies by model; some models are easier to steer than others. If a model does not cooperate with styles, slightly increase strength or redundancy to raise Styler influence.
The positive prompt (CONDITIONING) usually has more weight than the Styler node. Your prompt should not contradict the desired styles, or the Styler effect will be reduced.
For SDXL, Pony, and Illustrious, ControlNet's OpenPose is typically a guideline rather than a strict rule and can be overridden by the prompt. If the prompt contradicts your applied pose, ControlNet may be ignored or produce inconsistent composition. Reinforcing the pose in the prompt is generally a good idea.
Use camera_angles carefully so it does not conflict with your prompt or ControlNet. This is the most sensitive category and is often ignored when used incorrectly because it drives composition more than style.

Styler Iterator Workflow

Extensions required: comfyui-openpose-studio

You can load this image into ComfyUI to extract/open the workflow. This workflow iterates sequentially through styles in a category on each run, so you can test different styles without changing values manually. Due to a technical limitation, the generated image cannot include the iterated style name in its own workflow; use the style output from the Styler Pipeline (By Index) node as part of the filename, otherwise it is very difficult to identify which style was applied. The iterator workflow cannot persist the used index or the applied style name back into the workflow.

Conditioning Areas Workflow (Experimental)

The Styler Pipeline node is not only compatible with ControlNet workflows, but is also 100% compatible with the Conditioning Pipeline Area nodes from comfyui-lora-pipeline. This setup enables per-area styling, so you can apply different styles to different areas of the image by connecting Styler nodes into that pipeline. Those nodes also allow multiple LoRAs without blending their styles, because they encapsulate ComfyUI native Cond Pair Set Props logic while not exposing hooks, and using areas rather than masks.

Extensions required: comfyui-openpose-studio, comfyui-lora-pipeline
Experimental: fine-tunning this multi-LoRA multi-area with ControlNet workflow is more complex, and running it is considerably slower than regular workflows.

Per-area styles and consistent poses can be straightforward, but final image quality depends on many factors and is not detailed here. For more details, read the comfyui-lora-pipeline README.

See this post for a complete workflow combining multiple conditioning areas, OpenPose, ControlNet and Styler all used together.

Contributing

Core principles

Keep pull requests focused and minimal.
Avoid broad refactors unless discussed first.
Preserve existing architecture and rationale.

AI-assisted changes

If using an AI coding assistant, require it to read and follow AGENTS.md before making changes.

Acceptance criteria

One clear problem or improvement per PR.
Localized, reviewable diffs.
Clear explanation of why the change is needed.

License

MIT License - see LICENSE for full text.

Last Updated: 2026-02-13 Maintained by: andreszs Status: Active Development

docs/README.md Unescape Escape

English | Deutsch | Español | Français | Português | Русский | 日本語 | 한국어 | 中文 | 繁體中文

ComfyUI Styler Pipeline ✨

Table of Contents

Features

Installation

Requirements

Steps

Nodes

Styler Pipeline

Strength and Redundancy Guidelines

Styler Pipeline (Single)

Styler Pipeline (By Index) + Index Iterator

Advanced Styler Pipeline

LLM Setup

Cloud API providers

Ollama models (Local + Cloud)

AI Prompts

What to include

What to avoid

Using the returned suggestions

Advanced JSON

JSON shape

Practical rules

Support

Why Your Support Matters

💙 Support This Project

Ko-fi

PayPal

USDC (Arbitrum only ⚠️)

Gallery

Sample Workflow

Sample Images

Styler Iterator Workflow

Conditioning Areas Workflow (Experimental)

Contributing

Core principles

AI-assisted changes

Acceptance criteria

License

docs/README.md