Qwen Releases PE-T2I and PE-I2I Prompt Enhancers for Qwen-Image-2.1

Qwen has released two prompt enhancers fine-tuned on Qwen3.5-VL 9B for the Qwen-Image-2.1 image generator: PE-T2I expands short prompts in any language into detailed English descriptions, while PE-I2I rewrites editing instructions based on 1..N input images, returning the result strictly in JSON.

Compiled by Sergey KostenchukPublished 2026-09-21Updated 2026-09-22

2026-09-21 Multimodal HuggingFace

🎨 Qwen Releases Prompt Enhancers for Qwen-Image-2.1

Two enhancers based on Qwen3.5-VL 9B: PE-T2I expands short prompts in any language into detailed English descriptions and guesses the aspect ratio, while PE-I2I rewrites instructions for editing based on 1..N input images, preserving light, tone, and unmentioned details. The result is strictly JSON (rewritten_prompt, wh_ratio/ratio_follow).

🌍 Manual prompting — a weak point in open-source generation — is addressed with ready-made models featuring a strict JSON contract, and the Comfy-Org repack has already released an int8 encoder for ComfyUI, reducing VRAM requirements.

👤 PE-T2I and PE-I2I are already on Hugging Face — 4604 and 876 downloads on the first day. There is a "creative upscale" workflow by reverentelusarca, but the author warns: the model "kills realism," and grain and artifacts are possible.

Source 1: https://huggingface.co/Qwen/Qwen-Image-2.1-PE-T2I Source 2: https://huggingface.co/Qwen/Qwen-Image-2.1-PE-I2I

Sources