wan / v2.7 / fal / edit-video

The name is doing a lot of work here, so let's be blunt about what this node is and isn't. There is no local Wan 2.7. Alibaba stopped shipping open weights after Wan 2.2, and 2.7 - like 2.5 and 2.6 before it - is API-only. This node is Civitai's thin client for that API: it takes your prompt and a video, sends both to Civitai's orchestration fleet via fal, and drops the edited clip back into your graph as a native VIDEO. No model download, no VRAM, no checkpoint hunting. Just Buzz.

If that sounds like cheating to your local-first instincts, it kind of is - and that's the point. This is the node you reach for when you want an actual video-to-video edit (change the scene, restyle it, follow a new instruction) and you don't have the hardware to run Bernini or a VACE pipeline. It's also the node the community is mostly lukewarm about: Wan 2.7's reception has been muted at best, with people comparing it unfavorably to Wan 2.2 and LTX 2, so don't expect a miracle model wrapped in a convenient node.

How it works

Every node in this pack shares one skeleton. You fill in the widgets, hit run, and a background thread submits a workflow to orchestration.civitai.com, long-polls until it's done, and converts the returned file into ComfyUI's VIDEO type (persisted to the temp folder). The node shows a small status line with the workflow id and the cost, and it's billed per job in Buzz. It's a cloud call wearing a node costume - your input video and prompt leave your machine, full stop.

The inputs that matter

Two are required, one is the whole job:

• video_url - the clip you're editing. Per the tooltip: MP4/MOV, 2–10 seconds, max 100 MB. This is also the most common way to fail, so check the length and size before you blame the model.
• prompt - the edit instruction. This is an editing model, so describe the change, not the whole scene: "turn the sky to sunset" rather than "a beach at sunset."
• cfg_scale - defaults to 4. Start there.

The optional set is where the mode lives: source_image (a URL, data URL, or base64), reference_image for reference-based editing, duration (1–30s, default 5), frame_rate (default 24), resolution (720p or 1080p), seed, steps (10–50), loras (wire a Civitai LoRA Selector), aspect_ratio (falls back to the input video's), and audio_setting - "auto" or "origin" to keep the original audio. enable_safety_checker defaults to false.

Install and auth

This pack is official Civitai, published to the Comfy Registry, so the easy path is:

1. ComfyUI Manager → Custom Nodes Manager → search Civitai Comfy Nodes → Install → restart.
2. Or clone manually:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt   # just requests

Auth resolves in order: a connected Civitai Auth node, the CIVITAI_API_TOKEN env var, a stored key/OAuth from the Civitai sidebar, or a browser login. Headless? Use the env var.

Where people get burned

Wan 2.7 has no open weights, and the SEO-spam ecosystem has already spun up fake "Wan 2.7 local download" pages - don't fall for them. Beyond that, the real costs are money and patience: video jobs are the expensive end of the catalog, and your inputs go to Civitai's servers, filtered by whatever moderation runs there. Expect edits that need a retry or two to nail the instruction - editing is where this model family is weakest.