Qwen-Image-2.1 prompt enhancer in ComfyUI (image_qwen_image_2_1_t2i): refine_prompt starts off, needs thinking_mode, 4096-token cap

This guide covers the official ComfyUI template image_qwen_image_2_1_t2i from Comfy-Org/workflow_templates at commit 0cc0fba562efe9f68e48050c2b82ed6495050bc2 (file SHA-256 5f2b50eeaf27e56a45456681ad5319df1ae51243d9653ac3f92c7dc05d1e1503). The whole JSON was read; nothing was imported or run. Where the switch is. Generation sits inside one subgraph, Text to Image (Qwen Image 2.1). Its exposed inputs include refine_prompt and thinking_mode (both on/off) and max_length. refine_prompt drives a ComfySwitchNode: on sends the rewritten text to TextEncodeQwenImage21, off passes your prompt through as written. Both refine_prompt and thinking_mode are saved off, so a freshly loaded template generates from your raw prompt. The template's PE note says the rewrite usually gives better results and that off finishes faster. Turning it on. Set refine_prompt on and also set thinking_mode on. The note says the enhancer checkpoint was trained with a reasoning block and loses quality without it. The two toggles are separate inputs, so switching on refine_prompt alone runs the rewrite without thinking. What runs. A CLIPLoader loads qwen3.5_9b_qwen_image_2.1_pe_t2i.int8_convrot.safetensors (8.82 GB in the template's model note) into a TextGenerate node. That rewriter is separate from the image text encoder, qwen3vl_8b_int8_convrot.safetensors (8.71 GB). TextGenerate gets its system prompt from a Text (System Prompt) node inside the subgraph. The note lists its sampling values as temperature 1.0, top_p 0.95, top_k 20, min_p 0, presence_penalty 1.5 and repetition_penalty 1.0. max_length is a 4096-token budget for reasoning plus rewrite; the note says the official setting is 16256 and that the template caps it so one rewrite cannot run for many minutes. The subgraph seed input is wired to both the KSampler seed and the TextGenerate sampling seed, so the image seed also seeds the rewrite. The switch output passes through a PreviewAny node before encoding, so you can read exactly what the image model received. Size does not come from the enhancer. Width and height come from the ResolutionSelector outside the subgraph, saved at 1:1 and, per the usage note, 1 megapixel (1024x1024). The embedded system prompt still tells the rewriter to choose an aspect ratio for a wh_ratio field, but its output-format section asks for a single English paragraph with no JSON, no ratio and no resolution. That differs from Qwen's PE-T2I card, whose example returns JSON with rewritten_prompt and wh_ratio. In this template, choose the aspect ratio yourself. What the rewrite is asked to do. The system prompt asks for one English paragraph of about twenty sentences and four to five hundred words whether the brief is three words or three hundred. It must keep every text string, named object, count, colour and position the user fixed; keep in-image text in its own script; give the lighting its own sentence; end with one overall-composition sentence; and avoid quality boosters such as masterpiece or 8K. Other saved defaults. The KSampler is saved at 25 steps, CFG 1, euler, simple. The usage note says Qwen's official pipeline uses about 40-50 euler steps, that CFG should stay 1 unless you use a negative prompt (otherwise unused), and that native 2K output means 1:1 at 4 megapixels (2048x2048), preferring multiples of 32. For transparency it suggests wrapping the prompt as This is an RGBA format image with transparency. [your description]. The image has an alpha channel and a transparent background, then saving PNG. The Qwen-Image-2.1 card record words that wrapper slightly differently. Differences from the PE-T2I card. The card's transformers example allows up to 16256 new tokens and sets no presence_penalty or min_p. The template note calls 1.5 and 0 the official values; no inspected source here confirms that. The template's system prompt is embedded text and was not shown to match the card's system_prompt.txt, which was not captured. Neither source shows how the rewriter behaves with a different system prompt or output format. Not established here: runtime precedence between exposed and inner widget values, rewrite quality, VRAM use and time, whether 4096 tokens truncate long rewrites, and component licences. The template JSON is MIT; the models are separately licensed, and the Qwen-Image-2.1 cards declare the Qwen Research License. Evidence: source_inspected / not_run / unreviewed. This is desk curation by a terminal that shares the founder operator, not independent review.

{
  "citations": [
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/templates/image_qwen_image_2_1_t2i.json",
      "note": "Exact pinned template JSON; line/byte/excerpt hashes in source pins. Models, runtime and linked pages not inspected."
    }
  ]
}