Prompt & Dimensions from LLM

Prompt & Dimensions from LLM

A ComfyUI custom node created exclusively for the Qwen Image 2.1 Prompt Enhancer family. It consumes the JSON hint the model emits and turns it into a ready-to-use prompt plus an exact image resolution.

Description

Qwen Image 2.1 Prompt Enhancer is a prompt-rewriting model that converts a short image request (in any language) into a detailed English prompt and a target size. This node consumes that output and resolves the final width and height for the downstream image-generation node.

It is designed for both published variants:

• Text-to-Image — Qwen-Image-2.1-PE-T2I: emits {"rewritten_prompt": "...", "wh_ratio": "16:9"}.
• Image-to-Image — Qwen-Image-2.1-PE-I2I: emits {"rewritten_prompt": "...", "wh_ratio": "", "ratio_follow": "<image1>"}.

In both cases wh_ratio and ratio_follow are mutually exclusive — exactly one carries a value.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/BleynChannel/ComfyUI-QwenImage-PromptExtractor.git

How it works

1. Parses the JSON hint and validates the rewritten_prompt.
2. Resolves the size in one of two mutually exclusive ways:
   • wh_ratio + megapixels → computes width/height from the aspect ratio and target megapixels. The math matches ComfyUI's built-in Resolution Selector (1 MP = 1024 × 1024, rounded to the nearest multiple of multiple, default 8), so e.g. 1:1 @ 1 MP → 1024 × 1024.
   • ratio_follow → copies the exact pixel size of the referenced input image (ignoring megapixels).
3. Emits the final prompt, width, and height.

If both wh_ratio and ratio_follow are present, ratio_follow wins.

Inputs

| Input | Type | Description | | --- | --- | --- | | llm_prompt | String (multiline) | The JSON hint from the Qwen Image 2.1 PE model, e.g. {"rewritten_prompt": "...", "wh_ratio": "16:9"}. | | megapixels | Float | Target size in megapixels, 0.1–16.0. Used only with wh_ratio; ignored with ratio_follow. | | multiple | Int | Grid size to round to, 8–128 (step 4, default 8). Used only with wh_ratio. | | images | Autogrow (image) | Up to 10 reference images, keys image0…image9. ratio_follow may reference one of them. |

The node is not tied to a specific generation model: it works with any text-to-image model set by width/height (e.g. Empty Latent Image).

Outputs

| Output | Type | Description | | --- | --- | --- | | prompt | String | The rewritten prompt (rewritten_prompt). | | width | Int | Image width, in pixels. | | height | Int | Image height, in pixels. |

Resources

• 🤗 Qwen-Image-2.1-PE-T2I
• 🤗 Qwen-Image-2.1-PE-I2I
• 🐙 Qwen-Image-2.1 on GitHub
• 📑 Qwen Image 2.1 blog

Author

BleynChannel — [email protected]