# ComfyUI-Non-offset ComfyUI custom nodes focused on fixing pixel shift and spatial misalignment in the qwen-edit series edit models (e.g., qwen-edit-2509 and qwen-edit-2511). V1 targets qwen-edit-2509; V2 targets qwen-edit-2511. ## Examples

Prompt: change background to a seaside sunset
Prompt: wear a white down jacket, sunglasses, and pink pants

## 🚀 Key Features - CustomSmartResizePad nodes in two variants: - V1 (Standard): zero-padding, mainly for qwen-edit-2509. - V2 (Edge Replicate): recommended edge-pixel replication, mainly for qwen-edit-2511. - Why V2? - In qwen-edit-2509/2511 tests, traditional black borders caused positional drift (a few to tens of pixels). - Root causes: - Contrast interference: strong edges between pure black padding and the original image disturb positional embeddings. - Attention shift: large zero-valued regions dilute valid features during self-attention. - V2 fix: replicate padding keeps the padding region visually consistent with the original edges, improving coordinate capture and eliminating inpaint misalignment. ## 🛠️ Node ### Smart Resize & Pad V2 (Edge Pixel) - Resizes the input image proportionally to a target square (e.g., 1024x1024) and fills the short side using edge replication. #### Inputs - image: input image tensor. - target_size: target square size (default 1024, step of 8 supported). #### Outputs - padded_image: the filled image; the model sees a continuous frame instead of black borders. - mask: mask of the original region (white = original image, black = padding). It can guide models to focus on “real” pixels. ## 📂 Installation Go to your `ComfyUI/custom_nodes` directory: ```bash cd ComfyUI/custom_nodes/ ``` Clone this repo (or drop the `.py` files here): ```bash git clone https://github.com/starsFriday/ComfyUI-Non-offset.git ``` Restart ComfyUI. ## 📝 License MIT License ## Tips for qwen-edit For non-square images, keep V2 enabled. At 1024x1024 input, qwen-edit handles bounding boxes and point guidance most reliably. ## Contribution Found a better padding strategy? Please open an Issue or Pull Request!

Name	Type	Default	Description
image	IMAGE		—
target_size	INT	102464–8192	—

Name	Type	Description
padded_image	IMAGE	—
mask	MASK	—

Smart Resize & Pad (Black)

Inputs (2)

Outputs (2)