ComfyUI-Non-offset

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

<table> <tr> <th align="left">Prompt: change background to a seaside sunset</th> <td align="center"><img width="220" src="https://github.com/user-attachments/assets/6ed8d71b-ba08-4157-81ac-30311c2eabe9" /></td> <td align="center"><img width="220" src="https://github.com/user-attachments/assets/15b0996a-f679-4a75-8120-acbae8a651de" /></td> <td align="center"><img width="220" src="https://github.com/user-attachments/assets/07f96708-319c-4ebd-9487-cf4e536d1d44" /></td> <td align="center"><img width="220" src="https://github.com/user-attachments/assets/995a13ef-d63e-4875-b509-d1029e4b6ece" /></td> </tr> <tr> <th align="left">Prompt: wear a white down jacket, sunglasses, and pink pants</th> <td align="center"><img width="220" src="https://github.com/user-attachments/assets/239ff036-a4d3-4591-834d-99596346930b" /></td> <td align="center"><img width="220" src="https://github.com/user-attachments/assets/7ec59d1c-8c12-47d9-989a-db8b4fecac31" /></td> <td align="center"><img width="220" src="https://github.com/user-attachments/assets/8dfb43d0-1f7b-4bd1-a89b-8547ddccc324" /></td> <td align="center"><img width="220" src="https://github.com/user-attachments/assets/f387b0ab-7d99-445f-a179-b784156d77db" /></td> </tr> </table>

🚀 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:

cd ComfyUI/custom_nodes/

Clone this repo (or drop the .py files here):

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!