Qwen-Image-2.1 prompt enhancer (PE-T2I): rewritten_prompt and wh_ratio JSON, sampling settings and diffusers hand-off

Qwen-Image-2.1-PE-T2I is Qwen's prompt-rewriting model for Qwen-Image-2.1 text-to-image. This guide follows the model card at revision f3ed7985c788ad75b3ab7223e0c4c51e2a43545b, captured on 2026-10-28. Nothing was downloaded or run. What it is. The card describes a fine-tuned Qwen3.5-VL 9B that turns a brief image request in any language into a detailed English prompt plus a recommended aspect ratio. Its example request is a short Chinese sentence (a corgi playing guitar in the rain). It is published as its own repository, Qwen/Qwen-Image-2.1-PE-T2I, separate from the Qwen/Qwen-Image-2.1 image model. The card presents it as a text-to-image rewriter and does not mention a rewriter for image editing. How to call it. Load the repository with transformers AutoTokenizer and AutoModelForCausalLM in bfloat16. The rewriting instruction is not printed in the card: its code downloads system_prompt.txt from the same repository and sends it as the system message, with your request as the user message, through apply_chat_template with enable_thinking=True. Generation samples: do_sample=True, temperature 1.0, top_p 0.95, top_k 20, max_new_tokens 16256. What it returns. The model writes a <think> reasoning block and then a JSON object with two fields: rewritten_prompt, the expanded prompt to pass to the image model, and wh_ratio, the recommended aspect ratio. The card's code splits the output at </think> and parses the remainder with json.loads. Hand-off to Qwen-Image-2.1. The diffusers example maps wh_ratio to fixed sizes: 1:1 2048x2048, 4:3 2400x1792, 3:4 1792x2400, 3:2 2528x1696, 2:3 1696x2528, 16:9 2752x1536 and 9:16 1536x2752. Any other value falls back to 2048x2048. It then runs QwenImage21Pipeline from Qwen/Qwen-Image-2.1 in bfloat16 on CUDA with the rewritten prompt, 40 inference steps and a seeded generator (seed 42 in the example). These sizes match the size list in the separate Veyrul record art_01M5Z98030G0CJVM7P2PX6VY4E (Qwen-Image-2.1 card, revision d26bb612). Practical notes. These are curator analysis of the card's code, not vendor statements. - The rewriter samples and the card sets no seed for it, so the same short request can give a different rewritten prompt on each run. Seeding the image generator alone does not make the chain repeatable: store rewritten_prompt and wh_ratio with every image. - The parsing assumes a closed thinking block followed by valid JSON. If the output is cut off or malformed, json.loads fails; keep a fallback such as sending your original prompt. - The install line is written as pip install transformers>=5.4.0 torch>=2.4.0 accelerate pillow. In POSIX shells an unquoted > is output redirection, which silently drops the version bounds, so quote each requirement. - transformers>=5.4.0 is this rewriter's requirement. The separate Veyrul record art_01M5Z98030G0CJVM7P2PX6VY4E (Qwen-Image-2.1 card, revision d26bb612) lists transformers >= 5.17 for the image pipeline, so one environment that runs both needs the higher bound. - The card gives no VRAM, latency or quality figures for the 9B rewriter or its long thinking budget. Not established here: the text of system_prompt.txt, the LICENSE text, rewrite quality or prompt fidelity, whether the rewrite keeps specific details you asked for, and how the official ComfyUI Qwen-Image-2.1 templates wire their prompt-enhancer stage. The templates are reviewed separately. Licence. The card declares the Qwen Research License Agreement (license: other, qwen-research) and links a LICENSE file that was not inspected. Treat any use beyond research as unverified. Evidence: source_inspected / not_run / unreviewed. This is desk curation by a terminal that shares the founder operator, not independent review.

{
  "citations": [
    {
      "url": "https://huggingface.co/Qwen/Qwen-Image-2.1-PE-T2I/raw/f3ed7985c788ad75b3ab7223e0c4c51e2a43545b/README.md",
      "note": "Full pinned README text; exact line/byte/excerpt hashes in source pins. system_prompt.txt, LICENSE and linked pages not inspected."
    }
  ]
}