Qwen Image 2.1 Prompt Enhancer — ComfyUI Custom Node

ComfyUI nodes for the Qwen-Image-2.1 Prompt Enhancer (Qwen3-VL 8B fine-tunes).

Two nodes:

T2I Prompt Rewrite — expands a short text prompt into a detailed English description for text-to-image generation
Edit Prompt Rewrite — rewrites a vague edit instruction + source image(s) into a precise, actionable prompt
qwen3.5_9b_qwen_image_2.1_pe_t2i.int8_convrot.safetensors — 9.47 GB, text-to-image prompt rewriter https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/text_encoders/qwen3.5_9b_qwen_image_2.1_pe_t2i.int8_convrot.safetensors
qwen3.5_9b_qwen_image_2.1_pe_i2i.int8_convrot.safetensors — 9.47 GB, image-edit prompt rewriter https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/text_encoders/qwen3.5_9b_qwen_image_2.1_pe_i2i.int8_convrot.safetensors

Folder: ComfyUI/models/text_encoders/

Repository: github.com/benjiyaya/ComfyUI-Qwen-Image-2.1-Prompt-Enhancer

Architecture

The prompt enhancer models are Qwen3-VL 8B fine-tunes loaded as text encoders through ComfyUI's native CLIPLoader (type=qwen_image). Generation runs entirely on ComfyUI's built-in clip.tokenize() → clip.generate() → clip.decode() path — the same mechanism as the built-in TextGenerate node. No external server, no transformers pipeline.

Load CLIP (type=qwen_image)  →  CLIP
         ↓
   System prompt + user prompt (+ reference images)
         ↓
   Autoregressive generation (thinking mode on)
         ↓
   JSON answer parsed → positive_prompt, wh_ratio, ratio_follow

Setup

1. Download the PE checkpoints

Download the prompt-enhancer checkpoints from HuggingFace and place the .safetensors files in your ComfyUI/models/text_encoders/ directory:

Qwen/Qwen-Image-2.1-PE-T2I — for text-to-image prompt expansion
Qwen/Qwen-Image-2.1-PE-I2I — for image-editing instruction rewrite

2. Install this node

cd ComfyUI/custom_nodes/
git clone https://github.com/benjiyaya/ComfyUI-Qwen-Image-2.1-Prompt-Enhancer.git

3. Restart ComfyUI

The nodes appear under the Qwen Image category.

Usage

Add a Load CLIP node, pick the PE .safetensors file, set type to qwen_image.
Connect its CLIP output to the rewrite node's clip input.
Wire positive_prompt into your CLIP Text Encode / prompting node.
For Edit with multiple images, reference them as <image1>, <image2>, ... in the prompt.

Nodes

T2I Prompt Rewrite

Input	Type	Default	Description
clip	CLIP	—	PE checkpoint loaded via native CLIPLoader (`type=qwen_image`)
prompt	STRING	—	Short text prompt (any language)
temperature	FLOAT	1.0	Sampling temperature
top_p	FLOAT	0.95	Top-p sampling
top_k	INT	20	Top-k sampling
presence_penalty	FLOAT	1.5	Keep 1.5 for T2I
max_new_tokens	INT	16256	Max generation length
seed	INT	42	Random seed

Output	Description
positive_prompt	Rewritten detailed prompt
negative_prompt	Always empty (pipeline compatibility)
wh_ratio	Recommended aspect ratio (e.g. "16:9")
thinking	Model's reasoning trace
parse_ok	Whether the JSON answer parsed correctly

Edit Prompt Rewrite

Input	Type	Default	Description
clip	CLIP	—	PE checkpoint loaded via native CLIPLoader (`type=qwen_image`)
prompt	STRING	—	Edit instruction (use `<image1>`, `<image2>`... to reference images)
image_1 ... image_10	IMAGE	—	Optional reference images (batch tensors are split)
temperature	FLOAT	1.0	Sampling temperature
top_p	FLOAT	0.95	Top-p sampling
presence_penalty	FLOAT	0.0	Keep 0 for Edit
max_length	INT	24000	Max generation length
seed	INT	42	Random seed

Output	Description
positive_prompt	Rewritten precise edit instruction
negative_prompt	Always empty
wh_ratio	Canvas aspect ratio (if model chose one)
ratio_follow	Which input image to inherit the output ratio from
thinking	Model's reasoning trace
parse_ok	Whether the JSON answer parsed correctly

Workflow Example

Load CLIP (qwen_image) ─┬→ Qwen Image 2.1 — Edit Prompt Rewrite → CLIP Text Encode → KSampler
Load Image ─────────────┘                    ↓
                                      (positive_prompt)

Load CLIP (qwen_image) → Qwen Image 2.1 — T2I Prompt Rewrite ← "a corgi in the rain"
                                    ↓
                              (positive_prompt)

Hardware Requirements

VRAM: ~20GB for the PE model in bfloat16. A 24GB GPU is comfortable.
VRAM is managed by ComfyUI's model management — the model is loaded and evicted like any other text encoder.

Sampling Parameters

Parameter	T2I	Edit	Why
presence_penalty	1.5	0	T2I needs a high penalty to avoid repetition; edit should flow naturally
max length	16256	24000	Edit responses are longer (includes `ratio_follow` logic)
thinking	on	on	Both models were trained with `<think>` blocks

Optionally install json_repair to salvage slightly malformed JSON answers (pip install json-repair).

Files

File	Purpose
`__init__.py`	ComfyUI registration
`nodes/prompt_rewrite_nodes.py`	Node definitions
`prompts/system_prompt_t2i.txt`	T2I system prompt
`prompts/system_prompt_edit.txt`	Edit system prompt

Credits

Qwen-Image-2.1 by Alibaba/Qwen team
ComfyUI native integration pattern from TextGenerate

License

MIT — see LICENSE.

benjiyaya/ComfyUI-Qwen-Image-2.1-Prompt-Enhancer

Star growth

README