wan / v2.7 / fal / text-to-video

This is the plain vanilla of the Wan 2.7 family in the Civitai pack: a text prompt goes in, a video comes out, and nothing about the actual generation happens on your machine. Wan 2.7 is API-only - Alibaba hasn't shipped open weights since Wan 2.2 - so this node is Civitai's way of letting you call that API from inside ComfyUI, with the job running on their fleet via fal and billed in Buzz.

Reach for it when you want the convenience of a text-to-video step in a bigger workflow without a GPU that can breathe. The honest framing: Wan 2.7 is not the model the community is excited about (its launch reception ranged from muted to "Wan 2.2 looks better than this"), so don't expect this node to be your forever generator. It's a no-install, no-download way to get a current-generation video model into a graph that's already going to the cloud for other steps.

How it works

Same pack skeleton as every other Civitai recipe node. The node packs your prompt (plus whatever optional inputs you set) into a videoGen workflow, submits it to orchestration.civitai.com, long-polls the status until the job finishes, and converts the returned MP4 into ComfyUI's native VIDEO type, persisted to the temp folder. A small status line on the node shows the workflow id and per-job cost. It's a cloud call in a node costume - your prompt leaves your machine, and generation is gated by your Civitai account balance.

The inputs that matter

Just two are required: prompt and cfg_scale (default 4). Everything that makes the output watchable lives in the optional drawer:

• duration - 1 to 30 seconds, default 5. Start short; long clips are where text-to-video starts to drift.
• resolution - 720p or 1080p (default). 720p is a fine place to iterate before spending the Buzz on 1080.
• aspect_ratio - 16:9, 9:16, 1:1, 4:3, or 3:4. Set it for the platform you're posting to.
• negative_prompt - yes, Wan takes one, and it helps with common failure modes.
• seed (with control-after-generate), frame_rate (default 24), steps (10–50), loras (wire a Civitai LoRA Selector), enable_safety_checker, and enable_prompt_expansion (default on - Civitai rewrites your prompt first, which usually helps but can move the output off your intent).
• audio_url - a WAV/MP3 (3–30s, up to 15 MB) that the job generates the video with. Nice when you want sound without a separate audio pass.

Outputs: video (the clip), workflow_id, and raw_json (the full workflow response for debugging and cost inspection).

Install and auth

One install for the whole pack. In ComfyUI Manager: Custom Nodes Manager → search Civitai Comfy Nodes → Install → restart. Or:

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

The pack resolves credentials in order: a Civitai Auth node, the CIVITAI_API_TOKEN env var, a stored API key/OAuth from the Civitai sidebar, then interactive login. Headless installs should set the env var.

Where people get burned

Pure text-to-video is the least controllable of the Wan modes - no image to anchor it, so composition and character are up to the model's mood. That's why the image-to-video and reference-to-video siblings in this same family tend to get more love. Budget your expectations (and your Buzz) accordingly: iterate at 720p with a fixed seed, and treat the first pass as a draft.