Qwen has released two prompt enhancers for the Qwen-Image-2.1 image generation system, fine-tuned on the Qwen3.5-VL 9B model. Qwen-Image-2.1-PE-T2I expands a short request in any language into a detailed English description and guesses the aspect ratio of the future frame, while PE-I2I rewrites instructions for editing with consideration of input images, preserving lighting, tone, and unmentioned details. Both models return the result strictly in JSON, and the Comfy-Org team has already prepared an int8-quantized text encoder for ComfyUI, reducing VRAM requirements.

image
image
image

What happened

Qwen placed two fine-tuned copies of the Qwen3.5-VL 9B model on Hugging Face, which work not as a chatbot, but as a request preparation stage within a generative pipeline. After an internal reasoning block, each model returns a strictly structured response: a rewritten_prompt field plus wh_ratio for the text-to-image scenario or ratio_follow for the editing scenario. Demand was noticeable from the first day: the PE-T2I and PE-I2I versions were downloaded 4604 and 876 times respectively. In parallel, the Comfy-Org team, whose repack is led by Kijai, released files for ComfyUI, including an int8-quantized version of the text encoder under the name qwen3.5_9b_qwen_image_2.1_pe_i2i.int8_convrot.safetensors.

Context

Manual prompt creation is considered a weak point in open-source image generation: the final quality heavily depends on the user's ability to describe the scene, usually in English. Previously, this skill remained manual: auxiliary models returned free text that could not be reliably parsed and checked automatically. A different pattern is applied here: a vision-language model is fine-tuned for a narrow task and a strict output format, effectively becoming a prompt compiler, and the machine-readable response makes it a full-fledged link in the pipeline, suitable for automated testing. Distribution of heavy text models through quantized repacks by Comfy-Org is already a familiar way in the ComfyUI ecosystem to install such components on consumer cards.

Why this matters for the industry

For the industry, the model size is not important, but the contract: strict JSON with fixed fields allows embedding automatic prompt enhancement as an intermediate layer in products and agentic pipelines without manual text processing. Deployment is described specifically: a 9B parameter model fits on one GPU, and the stack is fixed on vLLM 0.19.1 and transformers 5.4.0, which simplifies deployment planning. This makes generative pipelines cheaper and faster for startups and simultaneously devalues manual prompt enhancement as a separate competence or service. The team gets a ready-made example of how specialization in a "fine-tuned VLM for a narrow machine-readable contract" is implemented within a creative pipeline.

Why this matters for users

Everything necessary is already available for download: PE-T2I and PE-I2I are on Hugging Face, and the int8_convrot encoder is in the Comfy-Org repack. To run it, one graphics card is enough: about 20 GB of weights in bf16, comfortable on a card with 40 GB of VRAM, on 24 GB the model starts with the --max-model-len 12000 parameter. The official prompt_rewrite repository has scripts for a vLLM server and batch processing, and a ready-made workflow is attached to the upscale. An important caveat: the prompt in this workflow will have to be adjusted for specific images, otherwise editing may produce artifacts.

What is still unknown / limitations

The sources do not contain a single quality metric for rewriting: the preservation of lighting, tone, and unmentioned details in PE-I2I remains a developer claim, and the accuracy of predicting the wh_ratio aspect ratio — the most testable part of the contract — is not measured anywhere. There is also no data on quality degradation after int8 quantization of the encoder. Finally, it is unknown whether a separate prompt enhancer will remain a permanent stage of the pipeline: if base models learn to natively accept short requests in any language, this layer will turn out to be a transitional stage.

Sources

Author

Look at AI, editorial team