Qwen Image 2.1 Prompt Enhance Subgraphs
Two basic Prompt Enhance subgraphs for Qwen Image 2.1 that can be copied and pasted into an existing ComfyUI workflow.
This includes separate versions for T2I and I2I.
T2I
Use this for normal text-to-image generation.
It takes your original prompt, sends it through the Qwen Image 2.1 T2I Prompt Enhance model, and outputs the rewritten prompt.
No image input is required.
I2I
Use this for image-to-image/editing workflows.
This version also accepts an image input, allowing the Prompt Enhance model to use the contents of the source image when rewriting the editing prompt.
Controls
Prompt Enhance
Turn this on to use the rewritten PE prompt. Turn it off to bypass Prompt Enhance and return your original prompt instead.
Seed
The seed belongs to the Prompt Enhance text generation, not your image sampler.
Changing the seed can give you a different rewritten prompt or a slightly different interpretation of the same request. If you like the result from a particular seed, keep that seed.
Required PE models
T2Iqwen3.5_9b_qwen_image_2.1_pe_t2i.int8_convrot.safetensors
I2Iqwen3.5_9b_qwen_image_2.1_pe_i2i.int8_convrot.safetensors
My subgraphs currently expect these inside a QwenImage_2.1 folder. If your models are stored somewhere else, simply change the model selected in the loader.
These are intentionally basic utility subgraphs, not complete generation workflows. Copy the appropriate one into your existing Qwen Image 2.1 workflow and connect the resulting string to your normal prompt encoding path.
Description
Initial release.
Includes separate T2I and I2I Prompt Enhance subgraphs for Qwen Image 2.1.
Uses ComfyUI core nodes only
Exposes Prompt Enhance on/off toggle
Exposes Prompt Enhance seed
T2I version enhances text prompts without an image input
I2I version uses the source image as context when rewriting edit prompts
Changing the PE seed can produce a different rewritten prompt/interpretation
FAQ
Comments (1)
Quick note for the I2I PE subgraph:
If you’re feeding multiple reference images into it, they need to be resized/padded to matching dimensions before batching. Once the images are the same size, batch them together and feed that batch into the I2I PE image input.
If the image sizes don’t match, the batch will fail or behave unpredictably. The T2I version does not need this since it has no image input.