OneThingAI Wanxiang Image

There's a fun irony in this node. The Wanxiang name belongs to Alibaba - the lab whose open releases (Wan 2.1/2.2 video, Qwen-Image, Z-Image) are the community's default local models, all Apache 2.0. OnethingAIImageWanxiang ("OneThingAI Wanxiang Image") is the other Wanxiang: the closed API tier, served through OneThing AI. Four models, one key, no weights. If you're on a machine with no GPU and you want a taste of the Alibaba family, this is the on-ramp; if you have a GPU, the open weights are probably still the better path.

How it works. Standard OpenAI-compatible client: POST to OneThing AI's /v1/images/generations with your Bearer key, get base64 JSON back, and it becomes a regular IMAGE tensor in your graph. The interesting bit is the model gating. This node's four models are wanx2.1-t2i-plus, wanx2.1-t2i-turbo, wanx2.0-t2i-turbo, and wan2.5-image-preview - and reference_image is enabled for exactly one of them. The t2i models are text-to-image, and the pack silently filters out any reference image you wire to them. If you want image-to-image, you must pick wan2.5-image-preview.

Inputs that matter. The model dropdown is the decision point, as above. After that:

• api_key and prompt - the essentials
• image_size - presets (1536x1024 / 1024x1024 / 1024x1536) or 自定义, with custom_width/custom_height (512–4096, step 64)
• retries / timeout - 3 and 120 by default
• reference_image (optional) - only meaningful on wan2.5-image-preview
• extra (optional) - JSON string merged into the request payload, for API params the node doesn't surface

Output is a single IMAGE.

Install. One clone, one restart - nothing heavy:

cd ComfyUI/custom_nodes
git clone https://github.com/OneThingAI/ComfyUI_Onething_Image.git
pip install Pillow requests

Restart ComfyUI, or search "ComfyUI OneThing AI Node" in ComfyUI Manager. No model downloads, no GPU requirements, because the inference happens on OneThing AI's servers.

Troubleshooting. The main gotcha is the silent reference-image drop: wire a ref image into a t2i model and you'll get a perfectly good image that just ignores it. Check the model dropdown before you chase your tail. Beyond that, it's the pack-wide stuff - 401/403 for a bad key or empty credit, a hung queue usually means the 120s timeout is too tight (push it toward 180), and the node re-runs on every queue even if nothing changed, so every pass is a billed generation.

Worth repeating the honest framing: the Wanxiang API tier is where Alibaba charges you for what its open weights give away. If you have the hardware, Wan 2.1/2.2 or Z-Image locally is the more community-tested route. This node is for when you don't - and it does that job cleanly.