Zhenzhen_qwen_image

Qwen-Image is a 20-billion-parameter model. Running it locally means finding a card with the VRAM for it, downloading the weights, and fighting with text-rendering quirks on a daily driver. Comfly_qwen_image skips all of that: it sends your prompt to the Comfly API relay and hands you back a real IMAGE tensor on the same canvas. Your GPU stays asleep - this is an API node, not a local one - and the only thing it costs is a few cents and an API key.

This node belongs to Comfyui_Comfly, a pack where every node is a relay to ai.comfly.chat, the author's paid API reseller. If you want the local experience, Qwen-Image is Apache 2.0 and runs fine at home; this node is for the days you want a fast generation without touching the weights. It is not a replacement for a full local setup, just a shortcut.

What you actually set

The prompt is the whole job. size gives you six sensible presets plus Custom, in which case Custom_size takes a literal 1280x720 string - get the format right or it bails with a white image and a message in the response output. num_images runs 1–4 and bills per image, so "generate four" means "pay for four." Leave model alone, it only has one choice.

The rest live under optional, and defaults are fine for a first run:

• guidance_scale (2.5 default) - how hard the model sticks to your prompt
• num_inference_steps (30) - more steps, more time, small quality gain past ~30
• enable_safety_checker (on) - turn off only if you know what you're asking for
• negative_prompt - Qwen-image takes it, though it's less decisive than in SD
• output_format - png or jpeg
• seed - set it if you want reproducible results

What comes out

Three outputs: image (the real IMAGE tensor - wire it to a Save Image node and it just works), image_url (a direct link to the hosted file), and response, a formatted text summary of what was generated. If something fails, response is where the error text lands - check it before you blame the node.

Installing it

The pack goes in through ComfyUI Manager (search "Comfyui_Comfly") or:

cd ComfyUI/custom_nodes
git clone https://github.com/ainewsto/Comfyui_Comfly

Restart ComfyUI. Manager will offer to install the pack's requirements - it lists a long haul (transformers, matrix-client, psutil), but that's for the pack's other AI-chat modules, not this node. Then grab an API key from ai.comfly.chat and paste it into the optional api_key field. Do it once and it's saved to the pack's Comflyapi.json, so future nodes won't ask.

Where people get burned

Missing or wrong API key is the classic failure: the node returns a blank white image and only whispers the reason into response and the console. A white square is the "check your key" signal. Also note the README says the pack is only tested on Windows 11, and the API is paid - it can't run offline, and if the relay is down, so are you. The English community barely knows this pack; it lives mostly in the author's QQ group, so when something's off, the README (in Chinese) is your best documentation.