This repository provides improved Qwen Image Edit custom nodes. The main goal is to fix the editing offset issue caused by ComfyUI’s built-in node, and to provide more flexible scaling and cropping logic.

Main nodes:

TextEncodeQwenImageEditAdv: Handles image + prompt encoding only, without scaling
QwenImageEditScale: Provides configurable scaling and cropping logic tailored for Qwen Image Edit models
QwenImageEditSimpleScale: Simplified version that automatically chooses the most suitable resolution

In comparative testing, this implementation eliminates offset issues and does not require multiple re-rolls for stable results.

Background

ComfyUI’s built-in TextEncodeQwenImageEdit node contains hard-coded scaling logic. However, this scaling approach is not fully aligned with the official Qwen Image Edit requirements and is difficult to debug.

This repository separates scaling from encoding, and provides more reasonable scaling strategies for easier testing and fine-tuning.

You can use ./workflows/demo_compare.json to quickly compare the official workflow with the custom nodes, and inspect consistency via the comparer.

Nodes & Parameters

1. TextEncodeQwenImageEditAdv

Works almost the same as the built-in TextEncodeQwenImageEdit:

Inputs: clip, prompt, optional vae and image
Outputs: conditioning and latent
No scaling logic included → fully transparent and controllable

2. QwenImageEditScale

Provides highly configurable scaling and cropping logic.

Parameters:

upscale_method: Resizing method (area / bicubic / bilinear / nearest-exact / lanczos) — usually fine to leave default
ratio_strategy: Aspect ratio handling strategy
- disabled: No ratio adjustment
- closest: Auto-crop to the closest officially supported ratio
- W:H: Force crop to a specific ratio (e.g. 3:2)
target_megapixels:
- Controls total pixels after scaling
- Default 1.0 → roughly 1024x1024
- Valid range: 0.3M ~ 1.4M
- Recommended values: 1, 0.65, 0.95, 0.9, 2
- Larger → more freedom, but less faithful to the prompt
- Smaller → stronger consistency, less offset
alignment: Step size for dimension alignment (default 32, official training value). Recommended values: 32, 16, 56, 14, 8

Outputs:

IMAGE: Resized image
width / height: Actual output resolution
ratio: Cropped aspect ratio string (e.g. 3:2)

3. QwenImageEditSimpleScale

A simplified version of QwenImageEditScale:

Automatically selects the most suitable scaling ratio based on input size
Ensures resolution falls within Qwen Image Edit’s acceptable range (~1M pixels)
Only requires two settings:
- resolution: Target resolution (default 1024) → in practice, almost only 1024 works properly due to model training
- alignment: Alignment step (default 32) → rarely needs to be changed, but you can experiment with 32, 56, 16, 8

Use cases:

For general use → QwenImageEditSimpleScale is recommended
For manual fine-tuning → use QwenImageEditScale
You can also use any other resize node outside this package, but it requires more manual adjustment

Recommended LoRAs

Combining the following LoRAs can further improve consistency and low-resolution performance:

DiffSynth-Studio/Qwen-Image-Edit-Lowres-Fix
QwenEdit Consistence Edit Lora by xiaozhijason

README_en.md Unescape Escape

Qwen Image Edit Adv - ComfyUI Nodes