Wan 2.7 Video Edit

There are two ways to "edit" AI video. One is regeneration with extra steps - hope the model re-renders close enough to what you had. The other is what this node does: take an actual video, give it instructions, and get a new video that keeps what you wanted and changes what you asked for. Wan 2.7's video edit model, served as a hosted API node built into ComfyUI core. No install, no local weights - 2.7 is API-only anyway, so this is the only way in.

The flow is: drop a video in (validated to 2–10 seconds), open the model dynamic combo, and write what you want changed. The sub-inputs are the real schema here: prompt takes editing instructions or style-transfer requests ("make it rain", "turn the day into neon-lit night", "restyle in the look of image1"). resolution is 720P or 1080P, ratio offers the standard five, defaulting to "don't change it" behavior - if you leave it alone it approximates the input video's aspect ratio. And duration is the one that surprises people: auto matches the input video, but a specific value truncates from the start of the video. Set 5 and you get the first 5 seconds re-edited, not a stretched version. That's the behavior to remember when your edit comes back shorter than expected.

You can also feed up to 4 reference_images into the combo. That's the style-transfer lane - bring a look reference, and the edit adopts it. Combined with a text instruction, you get "change the background to match image2" without regenerating the subject.

One widget on the node body itself is worth calling out: audio_setting, defaulting to auto. In auto, the model decides whether the edit calls for regenerated audio based on your prompt; switch to origin and it preserves the input video's original audio. For dialogue-heavy edits where the voice must survive, origin is usually the right call. watermark defaults off, seed is the standard re-run switch.

Output is a single VIDEO, ready for a Save/Preview or downstream nodes. The catch, as with everything in this family: it's a paid API, and the price badge makes the billing explicit - per second, and notably it counts input plus output time. A 10-second video edited to a 10-second result bills for 20 seconds of compute. That's the biggest practical surprise for first-time users, so keep clips short while you're iterating.

Wan 2.7's edit model landed in core with the rest of the 2.7 batch in April 2026. This is the node you reach for when your video is 90% right and you need to move it the last 10% - background swaps, style passes, weather changes - without the character drift that full regeneration brings.