Wan 2.7 Image to Video

Here's the thing to understand before anything else: the Wan you can run locally is not the Wan this node calls. Alibaba's open-weight line froze at Wan 2.2 in July 2025, and everything after - 2.5, 2.6, 2.7 - is API-only. This node is the 2.7 path, and it's a hosted API node built into ComfyUI core: it ships with ComfyUI, needs no install, and runs on Alibaba's servers through Comfy's proxy rather than your GPU.

The core input is simple. Drop a first_frame image in, and the output aspect ratio is derived from that image, so whatever shape your start frame is, the video follows. That's actually the nicest thing about it - no aspect-ratio dropdown to fight, your image is the canvas. last_frame is optional: give it one and the model animates a transition between the two frames, a feature that makes it genuinely useful for storyboarded sequences rather than just animating a still.

The model input is a dynamic combo - the top-level model is wan2.7-i2v, but clicking into it reveals sub-inputs: prompt, negative_prompt, resolution (720P or 1080P), and duration (2–15 seconds). That's a slightly hidden schema worth knowing about, because your prompt lives inside the model widget, not as a separate input on the node. Beginners routinely connect their text to a nonexistent prompt field and wonder why nothing happens.

Then there's audio, the sleeper feature. Wire an AUDIO input in (2–30 seconds) and Wan 2.7 will use it as a driving signal - lip-sync dialogue, beat-matched motion for a music clip. Leave it empty and the model auto-generates background music or sound effects to match the scene. Native audio on the open Wan? Never had it. On the hosted 2.7? Free, in the sense that it's the default behavior.

Other knobs are small but worth setting once: prompt_extend (default on) runs an AI rewrite of your prompt before generation - usually a win for quality, occasionally annoying when it "helpfully" embellishes, and worth toggling off if you have a carefully worded prompt. watermark defaults to false; flip it on if you want a visible AI-generated mark. seed is the standard "force a re-run" control, not a reproducibility guarantee.

Output is a single VIDEO, same as every other partner/video node, so it plugs into a Save/Preview or into downstream editing without fuss.

Two caveats. First, this is a paid, per-second API - the node shows a price badge in the UI and your Comfy account credits get billed per call, so test at 720p and short durations before committing to a 15-second 1080p render. Second, your input image gets uploaded to Alibaba's servers, and there's moderation on the other end. If a generation fails, it's usually the content filter, not your graph.

It landed in core with the Wan 2.7 batch in April 2026. If you want the latest Wan quality without a 24GB GPU, this is the cleanest way in.