ComfyUI-PepeUtils

ComfyUI-PepeUtils

Small utility nodes for ComfyUI.

Currently included:

• Anime PromptGen - generates anime prompt text with the FredZhang7 GPT-2 prompt generator or a compatible local GGUF file through Transformers.
• Load Image Cropped - loads an image and returns a cropped image + mask, with an interactive crop preview in the ComfyUI frontend.
• Pepe Paste Image - pastes a clipboard image into a selected node and keeps it only in ComfyUI's temporary storage.
• Pepe Resize Image - resizes, crops, pads, or pillarboxes images with automatic Lanczos upscale and Pepe Bicubic Sharper downscale selection.
• Pepe Scale Image By - scales images with a Photoshop Bicubic Sharper style approximation based on configurable cubic resampling.
• Stride Scale Size - computes width/height snapped to a chosen stride after scaling.

Node Screenshots

Installation

Option 1: Git clone

Clone this repository into your ComfyUI/custom_nodes directory:

git clone https://github.com/Pepehoschi/ComfyUI-PepeUtils.git

Option 2: Download ZIP

Download this repository as a ZIP, extract it, and place the folder here:

ComfyUI/custom_nodes/ComfyUI-PepeUtils

Then restart ComfyUI.

Requirements

No extra setup is currently documented beyond a normal ComfyUI installation.

This node relies on libraries that are typically already present in ComfyUI environments:

• torch
• numpy
• Pillow

The Anime PromptGen node also needs:

• transformers
• gguf when loading .gguf files through Transformers

Place local GGUF models in one of these folders:

ComfyUI/models/gguf
ComfyUI/models/llm_gguf
ComfyUI/models/LLM/GGUF

After restarting ComfyUI, files in that folder appear in the gguf_model dropdown. You can also paste an absolute .gguf path into gguf_file_path.

Included Nodes

Anime PromptGen

Category: utils/text

Inputs:

• prompt
• model_id_or_path
• tokenizer_id_or_path
• gguf_model
• gguf_file_path
• max_length
• num_return_sequences
• do_sample
• repetition_penalty
• temperature
• top_k
• early_stopping
• seed
• dtype
• device
• local_files_only
• strip_input_prompt

Outputs:

• prompts
• first_prompt

What it does:

• Wraps the Hugging Face text-generation pipeline.
• Exposes the generation arguments from the original outs = nlp(...) snippet as node widgets.
• Loads the normal Hugging Face model by default.
• Can load a compatible local .gguf model with Transformers using gguf_file.

Notes:

• Transformers loads GGUF checkpoints by dequantizing them into PyTorch weights, so memory use is closer to the selected dtype than to the GGUF file size.
• For GGUF workflows, the node first tries tokenizer metadata from the GGUF or adjacent files before falling back to tokenizer_id_or_path.
• If a custom node changes Hugging Face hub settings and you see requests to mirrors such as hf-mirror.com, enable local_files_only or set tokenizer_id_or_path to a local tokenizer folder.

Load Image Cropped

Category: image

Inputs:

• image

Outputs:

• image
• mask
• width
• height

What it does:

• Loads an input image from ComfyUI.
• Lets you define a crop rectangle.
• Returns the cropped image and cropped mask.
• Also returns the crop width and height as integers.

Frontend behavior:

• Includes an interactive crop preview.
• Internal crop coordinates are updated from the preview.
• Supports dragging a file onto the node to upload/select an image.
• Includes a Clear Crop context-menu action.

Notes:

• If the crop is invalid or empty, it falls back to the full image.
• Crop coordinates are clamped to the image bounds.

Pepe Paste Image

Category: image

Outputs:

• image
• mask

What it does:

• Select the node and press Ctrl+V to paste an image from the system clipboard.
• Alternatively, click Paste image from clipboard.
• Shows a preview and the pasted image dimensions.
• Uploads clipboard images only to ComfyUI/temp/pepeutils/paste-image.

Notes:

• Pasted files are removed by ComfyUI when it starts and when it shuts down normally.
• A saved workflow retains the temporary filename, not the image itself. Paste the image again after restarting ComfyUI.
• The clipboard button depends on browser clipboard permission. Ctrl+V remains available when direct clipboard access is unavailable.
• Exactly one Pepe Paste Image node must be selected for the Ctrl+V shortcut.

Pepe Scale Image By

Category: utils/image

Inputs:

• image
• scale_by
• snap_to_stride
• stride
• snap_mode (nearest, down, up)
• clamp_intermediate

Outputs:

• image
• width
• height

What it does:

• Scales a ComfyUI image batch by scale_by.
• Uses a NumPy separable resampler approximating Photoshop CS5.1 Bicubic Sharper behavior from Jason Summers' ResampleScope notes.
• For each axis, applies an integrated box prepass when scale is below 0.25, then a cubic pass with B=0, blur=1.05, and scale-dependent C.
• Processes X first, then Y, and clamps after cubic passes when clamp_intermediate is enabled.
• Uses a box blur of 0.33 and source-space phase 0.3999 for the large-downscale prepass.
• Uses cubic filter scale derived from the active cubic pass: direct downscales use scale; large downscales below 0.25 prebox to 4x target, so the final cubic pass uses 0.25.
• Equivalent cubic filter scale rule: max(scale, 0.25).

Notes:

• This is not guaranteed to be pixel-identical to Photoshop. Boundary handling, rounding, and exact clamping behavior are reverse-engineered approximations.
• Manual ring-pattern tests matched Photoshop closely from scale 0.3 upward with this derived math. Below 0.25, minor differences remain around Photoshop's exact box prepass/crop behavior.
• Very large images or batches may be slower than ComfyUI's built-in GPU scaling because this node runs the custom resampler on CPU.

Pepe Resize Image

Category: utils/image

Inputs:

• image
• width
• height
• keep_proportion (stretch, resize, pad, pad_edge, pad_edge_pixel, crop, pillarbox_blur, total_pixels)
• pad_color_r
• pad_color_g
• pad_color_b
• crop_position (center, top, bottom, left, right)
• divisible_by
• mask (optional)

Outputs:

• image
• width
• height
• mask

What it does:

• Provides KJNodes-style resize, pad, crop, pillarbox, and total-pixel geometry modes.
• Automatically uses Lanczos when enlarging.
• Automatically uses the existing Pepe Photoshop-style Bicubic Sharper path when reducing.
• Handles mixed-axis stretch by shrinking first, then enlarging.
• Applies the same geometry to optional masks with mask-safe bilinear interpolation.
• Aligns final dimensions downward to divisible_by.

Notes:

• Width or height may be 0; missing dimensions are derived from the source aspect ratio.
• pad_edge fills padding from average edge colors.
• pad_edge_pixel extends actual edge pixels.
• pillarbox_blur uses a blurred cover background plus a sharp fitted foreground.

Stride Scale Size

Category: utils/math

Inputs:

• image_width
• image_height
• rescale_by
• stride
• mode (down, up, nearest)
• side_selector (shortest, longest)
• clamp_rescale_min_1

Outputs:

• scaled_width
• scaled_height
• chosen_side

What it does:

• Scales an input size.
• Snaps the result to a stride.
• Returns the scaled width, height, and either the shortest or longest side.

Folder Structure

ComfyUI-PepeUtils/
├─ LICENSE
├─ __init__.py
├─ AnimePromptGen.py
├─ assets/
│  └─ nodes.png
├─ LoadImageCropped.py
├─ PasteImage.py
├─ PepeResizeImage.py
├─ PepeScaleImageBy.py
├─ StrideScaleSize.py
├─ examples/
│  └─ minimal_workflow.json
└─ web/
   ├─ load_image_cropped.js
   └─ paste_image.js

Example Workflow

A minimal example workflow is included at:

• examples/minimal_workflow.json

It places both included PepeUtils nodes into a small ComfyUI workflow:

• LoadImageCropped
• StrideScaleSize

Notes:

• Set the image filename in the workflow to a file that exists in your ComfyUI input folder.
• The two nodes are intentionally shown as simple standalone examples; StrideScaleSize uses numeric widget values rather than being wired from another node.

Development Notes

Files that should not be published as source artifacts are ignored in .gitignore, including:

• __pycache__/
• *.pyc
• .omx/

License

GNU GPL v3