Wan 2.7 Image

The Wan family was always secretly good at stills - people have been running single frames out of Wan 2.1 and 2.2 for image generation since 2025, and the community's summary was that video training gives the model a better feel for 3D space than a photo model has. Wan 2.7 Image is the same idea as a dedicated, closed-weight endpoint: text-to-image, plus editing when you give it an image.

Two things make it more than a novelty in ComfyUI. It generates up to four images per call (twelve in sequential mode), and it takes a bounding-box list so you can point at the part you want changed instead of describing it and hoping.

How it works

prompt is the only required field. input_urls is an optional IMAGE socket - feed it images and the node becomes an editor, leave it empty and it's text-to-image. The node uploads anything you connect, POSTs to KIE's marketplace /api/v1/jobs/createTask with model wan/2-7-image, polls /api/v1/jobs/recordInfo until the job resolves, then downloads the result image(s), archives them under ComfyUI/output/KIE-Results/images, and returns a ComfyUI IMAGE tensor.

Inputs that matter

• prompt - up to 5000 characters.
• input_urls - the source image(s) for editing. Leave it empty for pure text-to-image.
• n - how many images. Default 4; the range is 1-4 normally and 1-12 when enable_sequential is on. Note the default: run once and you pay for four. Set it to 1 while testing.
• resolution - 2K (default), 1K, 4K. The cost gradient is steep and non-linear.
• aspect_ratio - 1:1, 16:9, 4:3, 21:9, 3:4, 9:16, 8:1, 1:8. Only used when there's no image input.
• bbox_list - JSON, one inner list per input image, max two boxes each, each box [x1, y1, x2, y2]. This is how you say "change this region" precisely.
• enable_sequential - off by default; on, it does sequential/group image sets.
• thinking_mode - off. Only meaningful when enable_sequential is off and input_urls is empty; the schema says the frontend disables it otherwise.
• color_palette - JSON, 3-10 colours (8 recommended), only when sequential mode is off. A quiet way to force a consistent look across a set.
• watermark (off), seed, plus nsfw_checker, timeout_seconds, callback_url - the pack's standard trio.

Outputs

image is the deliverable - a normal IMAGE tensor, so SaveImage, PreviewImage, an upscaler, or a batch-splitter all work as usual. (Multiple results come back as a batch, which is what all_urls_json is for when you need to see what was actually returned.) Beside it: url, task_id, raw_json, and credits_consumed / credits_left, with the on-node receipt after each run.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
cd ComfyUI-KIE-Nodes-Next && pip install -r requirements.txt

Restart ComfyUI and look under KIE Next / Image / Wan. The pack ships no weights - requests, Pillow, numpy and PyYAML is it, since the model is entirely server-side.

Key, once: Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key; it validates first, then shows ✅ CONNECTED and a masked fingerprint. KIE_API_KEY for headless installs, and it takes priority. KIE is a pre-paid credit marketplace with no subscription.

Common issues

Four images when you wanted one. n defaults to 4. Set it explicitly.

The edit ignored your bounding box. bbox_list is JSON, and the outer list has to line up with input_urls - one entry per image, [x1,y1,x2,y2] per box, two boxes max. Malformed JSON is the usual culprit (the node will tell you it isn't valid JSON).

aspect_ratio does nothing. It only applies with no image input; an image sets the ratio.

4K takes a long time and costs a lot. Draft at 1K or 2K and only finish the winner at 4K.

thinking_mode seems ignored. By design - it only applies to plain text-to-image with sequential mode off.

Nothing saved to disk. It's an IMAGE output; wire it to a save node.