Nodes/ComfyUI_MieNodes/Qwen-Image-2.1 Prompt Generator ๐Ÿ‘
ComfyUI Node

Qwen-Image-2.1 Prompt Generator ๐Ÿ‘

A ComfyUI node in ๐Ÿ‘ MieNodes/๐Ÿ‘ Prompt Generator with 9 inputs and 1 output.

By MieMieeeeeยทCreated 2 years agoยทUpdated a day agoยท 262
Qwen-Image-2.1 Prompt Generator ๐Ÿ‘
  • llm_service_connector
  • reference_images
  • enhanced_prompt
โ—„modet2i - ๆ–‡็”Ÿๅ›พ(้ป˜่ฎค)โ–บ
โ—„user_inputโ–บ
โ—„seed0โ–บ
โ—„caption_modecache_memory_disk - ็ผ“ๅญ˜:ๅ†…ๅญ˜+็ฃ็›˜(ๆŽจ่)โ–บ
โ—„temperature0.70โ–บ
โ—„max_tokens16384โ–บ
โ—„timeout300โ–บ
Category๐Ÿ‘ MieNodes/๐Ÿ‘ Prompt Generator

Inputs (9)

NameTypeDefaultDescription
llm_service_connectorLLMServiceConnectorโ€”
modeCOMBOt2i - ๆ–‡็”Ÿๅ›พ(้ป˜่ฎค)Mode / ๆจกๅผ: t2i ๆ–‡็”Ÿๅ›พ(้ป˜่ฎค) โ€” observer English paragraph. t2i_rgba ้€ๆ˜Žๆ–‡็”Ÿๅ›พ โ€” same plus forced RGBA bookend sandwich. edit ๅ•ๅ›พ็ผ–่พ‘ โ€” single-image edit instruction (rewrite body is always English). multi_ref ๅคšๅ‚่€ƒ็ผ–่พ‘ โ€” multi-reference edit instruction (<image1>โ€ฆ<imageN>; rewrite body always English). remove_bg ๆŠ ๅ›พ โ€” fixed BG-remove block, no LLM call.
user_inputSTRINGRough idea / ่‰็จฟ: a sentence, an outline, a few lines, or a free paragraph. The rewriter expands it into the prompt the Qwen-Image-2.1 sampler consumes. Output is always English; verbatim text inside straight double quotes in your input is preserved as-is. / ๆ‰ฉๅ†™ๆญฃๆ–‡ไธ€ๅพ‹่‹ฑๆ–‡๏ผ›็›ดๅŒๅผ•ๅทๅ†…็š„ๅญ—็ฌฆไธฒไฟ็•™ๅŽŸ่ฏญ็งไธŽ้€ๅญ—ใ€‚
seedINT00โ€“18446744073709550000Seed / ็งๅญ forwarded to the LLM call. 0 lets the connector pick a fresh seed.
reference_imagesoptIMAGEOptional IMAGE batch / ๅฏ้€‰ IMAGE ๆ‰นๆฌก. For mode=edit / multi_ref ๅ•ๅ›พ็ผ–่พ‘ / ๅคšๅ‚่€ƒ็ผ–่พ‘, the node captions each image (<image1>โ€ฆ<imageN>) with the bundled _caption_image.txt system prompt and injects the captions into the rewrite user message as locked identity / wardrobe / palette DATA. Pixels are visible to the captioner LLM but NOT to the rewrite LLM. / ๆ”นๅ†™ LLM ็œ‹ไธๅˆฐๅ›พใ€‚ Recommended ceiling: 10 (model slot limit); beyond that the node logs a warning but still passes the full count. IMPORTANT / ๆณจๆ„: ๅคš่ง†ๅ›พ่ฎพๅฎš่กจๆ˜ฏ็‰ˆๅผใ€‚ๅ››่บซๆจชๆŽ’ๅ–‚่ฟ›ๅŽป๏ผŒๆˆๅ›พไผšๅคๅˆถ่ฟ™ๆŽ’ไบบ๏ผŒๆฏไธชไบบไธ€ไธชๅงฟๅŠฟใ€‚ๅ…ˆ่ฃๆˆๅ•ไบบ๏ผˆๆญฃ้ขๆˆ–ๅ››ๅˆ†ไน‹ไธ‰ไพง้ข๏ผ‰ๅ†ๆŽฅ reference_images๏ผŒ่พ“ๅ‡บ็”จ็ซ–ๅน…ใ€‚็ผ–่พ‘่Š‚็‚น็œ‹่ง็š„ไปๆ˜ฏไฝ ๆŽฅไธŠ็š„้‚ฃๅผ ๅ›พ๏ผ›่ฟ™ไธช่Š‚็‚นๅชๅ†™ๆ็คบ่ฏใ€‚ A multi-view character sheet is a layout: four standing views in, four poses out. Crop to one person before wiring the image. Wire any sampler / loader output that emits an IMAGE batch โ€” e.g. LoadImageBatch or a custom collector.
caption_modeoptCOMBOcache_memory_disk - ็ผ“ๅญ˜:ๅ†…ๅญ˜+็ฃ็›˜(ๆŽจ่)Caption cache strategy / ๆ ‡ๆณจ็ผ“ๅญ˜็ญ–็•ฅ (only used when reference_images is connected and mode in {edit, multi_ref}). cache_memory_disk (recommended) ็ผ“ๅญ˜:ๅ†…ๅญ˜+็ฃ็›˜ โ€” in-memory + persistent on-disk cache under <ComfyUI output>/mien_nodes/caption_cache/. cache_memory_only ็ผ“ๅญ˜:ไป…ๅ†…ๅญ˜ โ€” RAM-only cache, lost on restart. no_cache ็ฆ็”จ็ผ“ๅญ˜ โ€” always bypass cache. force_recaption_once ๆœฌๆฌกๅผบๅˆถ้‡ๆ‰“ๆ ‡ โ€” force fresh captions now (e.g. after editing _caption_image.txt).
temperatureoptFLOAT0.700โ€“2Sampling temperature / ้‡‡ๆ ทๆธฉๅบฆ. 0.0โ€“2.0, step 0.05. Lower = strict adherence to system prompt; higher = more surface variation.
max_tokensoptINT1638464โ€“32768Token budget / Token ไธŠ้™. 64โ€“32768. Reasoning models count their thinking against this cap; 16384 is the safe default for the whole plugin.
timeoutoptCOMBO300Per-call timeout (seconds) / ๅ•ๆฌก่ฐƒ็”จ่ถ…ๆ—ถ(็ง’).

Outputs (1)

NameTypeDescription
enhanced_promptSTRINGโ€”