NS WaveSpeed Qwen Edit

Qwen-Image-Edit is one of the models that made mask-based inpainting feel old-fashioned - you describe the change in a sentence and it re-renders the image accordingly. This node is WaveSpeed's hosted version of that edit model, so the whole thing runs over the API and a finished IMAGE lands back in your graph.

It's part of the Symbiotica pack's neuralsins/WaveSpeed family. The KB's inpainting doc is blunt about the shift: instruction-editing models like Qwen-Image-Edit and Kontext took over most of the job that masked inpainting used to teach. This node is the closed-model version of that - Qwen-Image has open weights you could run locally, but WaveSpeed's hosted path means you don't need the GPU or the setup.

One thing worth calling out: the tooltip says it supports Chinese and English prompts. That's a genuinely useful property if you work in both, and it's a hint about the model's home turf.

How it works

image_url and prompt go in, it POSTs to WaveSpeed's /api/v3/wavespeed-ai/qwen-image/edit, and the edited image comes back through image (IMAGE). The controls:

• prompt - the edit instruction, in plain language, Chinese or English.
• image_url - the source image as a URL string. Standing family rule: hosted and reachable, not a local tensor.
• seed - -1 for random, otherwise fixed for reproducible results. Nice to have on an edit node.
• output_format - jpeg (default), png, or webp.

enable_sync_mode defaults on.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/symbiotica-ai/comfyui-nodes.git symbiotica
pip install -r symbiotica/requirements.txt

Restart or use ComfyUI Manager (search Symbiotica), and feed an NS WaveSpeed Client node with your WAVESPEED_API_KEY into client.

Common issues

• Whole-image edits drift. Qwen-Image-Edit changes what you asked for and sometimes what you didn't - the KB's inpainting doc notes the model tends to "shift other details not asked for" and degrade over successive edits. Keep prompts surgical, or pair it with masking locally if exactness matters.
• One image, one edit. There's no multi-reference input here - for that, the Plus variant takes up to three images.
• URLs, not tensors. You can't wire a local Load Image straight in.
• The Qwen-Image VAE reputation. If the output looks over-smoothed or soft, that's the model family's known decoder trade - it prioritizes text legibility over fine texture. Nothing the node can change, but it's why the result may look "cleaner" than you'd expect.

The honest take: this is the no-fuss edit node of the family - one image, one prompt, a seed, done. If your edits need reference images or style steering, the LoRA and Plus siblings are where this node graduates to.