NH-Nodes for ComfyUI

Production-focused custom nodes for ComfyUI workflows: image loading, tiling, compositing, mask utilities, smart resize, logic routing, prompt/text tooling, batch helpers, indexed model loaders, garment/portrait segmentation, and VTON preprocessing.

• Author: jetthuang.com
• Hugging Face: nhathoangfoto
• Support: PayPal

---

Why NH-Nodes

NH-Nodes is built for practical workflow construction. It focuses on reusable building blocks that remove repetitive graph work:

• Load images with filename/path metadata.
• Split images into tiles, upscale or process tiles, then stitch them back.
• Build image grids, layer composites, comparisons, and labeled previews.
• Resize images and masks with predictable aspect-ratio behavior.
• Route workflow branches with boolean logic and generic data switches.
• Generate prompts from templates, regex, lists, schedules, and text utilities.
• Pick indexed LoRA/diffusion models for larger configurable workflows.
• Prepare VTON masks, agnostic images, DensePose/parsing outputs, and related masks.

---

Installation

ComfyUI Manager

Search for NH-Nodes, install, then restart ComfyUI.

Comfy CLI

comfy node install nh-nodes

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/jetthuangai/NH-Nodes.git
cd NH-Nodes
pip install -r requirements.txt

Restart ComfyUI after installation.

---

Featured Workflows

Tile, Process, Untile

Use this when an image is too large for a model, or when every tile needs the same enhancement/upscale step.

Load Image -> Image Tile (NH) -> upscale/process tiles -> Image Untile (NH)

Image Tile (NH) supports:

• original_ratio: derive tile dimensions from the original image aspect ratio.
• custom: use explicit tile width and height.
• square: force square tiles, for example 1024 x 1024.
• Automatic overlap calculation.
• Edge padding so square/custom tile outputs keep the requested tile size.

Image Untile (NH) uses tile_data from the tile node to stitch results back. It also handles tiles that were resized or upscaled, such as 2x or 3x per tile.

Grid Contact Sheet

Image batch -> Image Grid Composite (NH) -> Preview Image

Use manual cell size or first-image cell size. Resize modes include stretch, pad, and fill.

Metadata-Aware Image Load

Load Image Info (NH) -> image / mask / file_name / path

Useful for naming outputs, logging source images, or routing by filename.

Smart Resolution Resize

Load Image -> NH Smart Ratio Image Resize -> (your pipeline) -> NH Remove Resize Padding

NH Smart Ratio Image Resize picks the closest model preset itself from model_family + resolution_level, so it does not need NH Smart Resolution Picker upstream. Resize with cover crop, contain pad, or stretch behavior.

In contain_pad mode it also emits resize_data describing the content region, padding, and scale factor. Feed that into NH Remove Resize Padding at the end of the pipeline to crop the padded border back off (content_region) or to restore the original input size (original_size).

Use NH Smart Resolution Picker when you need target dimensions and a latent for text-to-image, instead of resizing an existing image.

Garment and Portrait Segmentation

Load Image -> Garment Segment (NH) [part=upper_garment] -> mask / cutout_rgba / mask_image
Load Image -> Portrait Segment (NH) [hair + face] -> mask / cutout_rgba / mask_image

SegFormer-based masks at input resolution. cutout_rgba is the input image with the mask as alpha (RGB is left untouched outside the mask); mask_image is the mask as an RGB preview. Dresses and jumpsuits belong to upper_garment. Both nodes share mask_expand, mask_blur, and fill_holes. Models are auto-downloaded to models/NH-Nodes/ on first use (existing models/face_parsing or models/RMBG/segformer_fashion copies are reused).

---

Node Catalog

Image Loading and Saving

Image Tiling, Grid, and Composition

Smart Resolution

Vision Segmentation

Mask Tools

Logic and Flow Control

Text and Prompt Tools

Lists and Batch Helpers

Indexed Loaders

Workflow Utilities

VTON Preprocessing

---

Menu Structure

NH-Nodes/
  Image/
    Load/save helpers
    Tile/untile
    Grid and layer composition
    Resize, compare, label
  Mask/
    Morphology, properties, bbox, aspect matching
  Vision/
    Garment and portrait segmentation
  Logic/
    Compare, gates, switches, branching, math, random choice
  Text/
    String, regex, split, concatenate, templates, scheduling
  Batch/
    Image batch, list helpers, any-list split
  Resolution/
    Smart resolution picker, ratio resize, preset resize, padding removal
  Loaders/
    Indexed LoRA, diffusion model, VAE, and CLIP loaders
  Utils/
    Sliders, booleans, universal pipe
  VTON/
    VTON Ultimate Processor

---

Requirements

numpy
Pillow
scipy
opencv-python
scikit-image>=0.18.0
onnxruntime>=1.8.0
transformers>=4.20.0
tqdm>=4.62.0
huggingface_hub>=0.16.0

Most image/text/logic nodes use common ComfyUI dependencies. VTON and Vision nodes download model assets on first use (transformers SegFormer weights for Vision: ~190 MB fashion, ~340 MB face).

---

Development Notes

• Layout: core/ holds shared helpers and model backends (no nodes); nodes/<category>/ mirrors the ComfyUI menu, and every module there that defines NODE_CLASS_MAPPINGS is registered automatically by __init__.py - drop a file in the right folder and restart. VTON's vendored preprocessing lives in nodes/vton/preprocess/.
• Tests: python -m pytest custom_nodes/NH-Nodes/tests -q from the ComfyUI root.
• Restart ComfyUI after installing or updating nodes.
• If a node appears cached in the browser, refresh the ComfyUI page after restart.
• Image Tile (NH) and Image Untile (NH) must pass tile_data directly through the workflow so reconstruction uses the original tile layout.

---

License

MIT. Free for personal and commercial use.

If NH-Nodes helps your workflow, consider starring the repository: github.com/jetthuangai/NH-Nodes