Wan 2.7 Video Continuation

Video models have a hard ceiling: they make a few seconds, then stop. This node is the fix for that within the Wan family - you feed it an existing clip and Wan 2.7 continues it, generating whatever comes next and stitching it seamlessly onto the end. It's the "one more scene, please" node, and it's the correct mental model for working around the clip-length wall without doing a manual chain-and-restitch loop.

The one thing to keep front of mind: Wan 2.7 is API-only. No weights ever shipped for it - Alibaba's numbered Wan series has been closed since 2.5 - and the local community's reception was muted, with evaluations putting it behind open Wan 2.2 on character-heavy motion. What this node gives you is that closed 2.7's continuation behavior, not something you can run on your own GPU. If you want local long-form Wan, you're chaining Wan 2.2 through extension workflows instead. Both approaches are legitimate; this one just doesn't need the VRAM.

What you set

• first_clip - the video to continue, between 2 and 10 seconds. The output's aspect ratio is inherited from this clip, so there's no ratio dropdown to get wrong.
• model - a dynamic combo holding wan2.7-i2v plus its own nested options: prompt, negative_prompt, resolution (720P/1080P), and duration (2–15 seconds total output). The tooltip explains the logic clearly: the model fills the remaining time after the input clip, so a 5-second clip with a 10-second duration gets 5 seconds of new footage.
• last_frame - the optional steering wheel. Drop in an image and the continuation transitions toward it. This is the difference between "keep going" and "keep going, and end up looking like this."

The usual API-node extras are here too: seed for reproducible takes, prompt_extend (AI prompt enhancement, on by default), and watermark (off by default).

How it runs

Standard cloud flow: ComfyUI uploads your clip - and the last frame if you set one - to the Wan API through the Comfy proxy, the task runs on Alibaba's servers, and the finished, extended video is downloaded back into the graph as a VIDEO object. Needs a Comfy account with credits, no local model, no GPU. The price badge reflects that you're billed for the input clip and the output duration, which is the honest way to think about the cost: every second you add is billable, and the source seconds you already had get re-encoded into the result.

Practical notes

Because the first clip is 2–10 seconds and the output is 2–15, this is a build-in-chunks tool: extend a short clip by a few seconds, feed the result back in, extend again. Each hop is one paid generation, so plan the take in a text editor before you spend. And if you care about matching motion at the seam, the last_frame input is your friend - a clip that ends in one place and a continuation that starts somewhere unrelated is how you get the "morphing mush" failure. Point it at a frame and the handoff gets dramatically more sane. For everything else, the negative prompt and resolution are the two levers that move quality.