KIE Wan 2.7 Edit Pro (Image) [SL]

Funny thing about Wan: the open-weights family everyone actually runs locally froze at 2.2 back in 2025 - Wan 2.5, 2.6, and 2.7 shipped API-only, with no weights at all. So this node is your only door to the current Alibaba edit model: the "Pro" image editor that can take multiple input images and follow an edit instruction across them. You literally cannot run Wan 2.7 Image Edit Pro on your own GPU, because there's nothing to download.

What it does is multi-image editing, and that's the differentiator worth knowing. Where Seedream 4.5 Edit takes one image and patches it, Wan 2.7 Edit Pro accepts up to 9 input images and reconciles them - think character reference + pose reference + background plate, edited together per your instruction. The "thinking mode" toggle and watermark flags exist in the underlying API and are just turned off here, and every run rolls a fresh seed so you never get a cached repeat.

The inputs that matter

• images - required, up to 9. All of them upload and all participate.
• prompt - the edit instruction, up to 5000 characters.
• aspect_ratio - a strict enum: 1:1, 16:9, 4:3, 21:9, 3:4, 9:16, 8:1, 1:8. Anything else and the API rejects it before a task even starts.
• resolution - 1K or 2K. Note the missing 4K: KIE only allows 4K for non-sequential text-to-image, so an edit request at 4K fails with a specific error. Don't fight it.

Output is a single image tensor, straight into SaveImage.

How it works

Standard KIE job pipeline - upload all input images, create a task for wan/2-7-image-pro, poll until done, download, decode. The pack's source is unusually explicit about the quirks it works around: the aspect-ratio enum is validated up front because KIE rejects bad values with a confusing "not within the range of allowed options" error, and the seed is randomized internally so the node forces re-execution every run (IS_CHANGED returns NaN) - which means an unchanged graph never serves you a stale cached result.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/v3devv/ComfyUI-Kie-API-Serverless

or ComfyUI Manager search "Kie-API-Serverless", restart. No models, no extra dependencies, KIE key in the node input.

Where people get burned

The 4K misconception is the top one - people see an edit node and assume resolution goes as high as the image model, and it doesn't; use 2K as your ceiling. Second: multi-image input sounds like magic but the model still edits within the frames you give it, so mismatched reference images produce mush. Third: per-call cost with multiple uploads adds up - you're paying for uploads and a heavier edit on top of the render, so batch your references sensibly.