Zhenzhen_qwen_image_edit

Qwen-Image-Edit became the open-weight standard for instruction editing because it's genuinely good at "turn the chicken leg into a burger" style prompts. Comfly_qwen_image_edit gives you that same model through the Comfly API relay, which means you edit images with the best-in-class instruction editor while your GPU idles. This is the sibling of Comfly_qwen_image - same pack, same API key, but it takes an image in and changes it instead of dreaming one up from nothing.

Everything runs server-side at ai.comfly.chat. You bring the image and the instruction; the node base64-encodes the image, POSTs it to the edit endpoint, and waits. It's billed per request and per image - the README quotes it around ¥0.1 a go, cheap enough to iterate on.

The inputs that matter

• image - what you're editing. One tensor in, one edit out.
• prompt - the instruction, in natural language. "Make the background a beach at sunset." This model is happy with plain English; no negative prompt gymnastics needed, though negative_prompt exists.
• size / Custom_size - same six presets plus Custom (format: 1280x720). Note the edit follows the size you pick, so keep it close to the input or expect some cropping.
• num_images - 1–4 variants from the same instruction. Billed per image, naturally.

The optional row adds num_inference_steps (30), guidance_scale (4 here - a touch higher than the generation node's 2.5), enable_safety_checker, output_format, and seed. The interesting extra is acceleration: none, regular, or high. High trades a bit of quality for a faster turnaround, which is handy when you're looping on an edit. Start with none to see the ceiling, then decide.

What comes out

Same trio as the generation node: image (real IMAGE tensor - save it straight from the canvas), image_url, and response (the human-readable log plus any error text). If you see a white square come back, that's the failure signal - check response and the console; a missing API key and a bad Custom_size both produce exactly that.

Setup

Install via ComfyUI Manager (search "Comfyui_Comfly") or:

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

Restart, let Manager install the requirements (they're heavy - transformers, matrix-client - but that's for the pack's other modules, not this node). Put your key from ai.comfly.chat in the optional apikey field; it's saved to the pack's Comflyapi.json and reused. README caveat: tested on Windows 11 only.

The honest take

If you already run Qwen-Image-Edit locally, this node buys you nothing but a credit card bill. It shines when you don't have a GPU big enough for a 20B model, or when you're on a machine that can't run the weights at all and want the same editing quality in your ComfyUI workflow. For that use case it's genuinely handy - just remember it's a paid relay, so the edit quality you get is exactly what Alibaba ships, and nothing about it is local.