Alibaba WAN 2.7 Text-to-Video

WAN 2.7 Text-to-Video is the text-only entry point to Alibaba's Wan video family inside the ERPK Collection. No image input, just a prompt, and after a few minutes of polling you get a video_url STRING back. It's the purest "describe a clip, get a clip" node in the pack's WaveSpeed section.

Here's the thing you need to understand before you get excited: Wan 2.7 has no open weights. The Wan family went API-only after 2.2 - the KB panel is blunt that "2.5, 2.6 and 2.7 have no weights" - so everything this node touches is a billed WaveSpeed call. If you came looking for the open Wan you can run locally, that train left at 2.2. What this node gives you instead is the current hosted Wan without you writing any API code, which is still a perfectly good reason to reach for it.

The inputs

• prompt - the video description
• negative_prompt - elements to exclude (a genuinely useful field for Wan, which can wander into unwanted territory)
• audio - optional audio URL to guide generation
• duration - 2 to 15 seconds, default 5
• aspect_ratio - default 16:9, with the usual portrait/square options
• resolution - output resolution dropdown
• enable_prompt_expansion - auto-enriches your prompt before generation
• seed - reproducibility, -1 for random
• client - optional WaveSpeed client, falls through to ERPK Settings

The single output is video_url (STRING), wired into Preview Anything like every other video node in this pack. No IMAGE output, no tensor - it's URL or nothing.

Install and setup

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk && pip install -r requirements.txt

Or "ERPK Custom Nodes" from ComfyUI Manager, restart, and put a WaveSpeed API key in ERPK Settings (right-click canvas). The pack polls every 10 seconds with a 15-minute timeout; a long, high-res clip sits at the slow end of that.

Troubleshooting and where it fits

The cheapest wins are on the node itself: turn on enable_prompt_expansion before you start hand-writing longer prompts, and actually use negative_prompt - Wan responds well to being told what to avoid ("no morphing, no flicker, no extra people"). Watch the duration range (2–15s) and keep test runs short; a 15-second clip is a long generation and a bigger bill.

For the model choice, the honest guidance from the pack's lineup: Wan is the continuity workhorse, Seedance is the reference-and-audio specialist. If your use case is "turn this sentence into a usable b-roll clip," this node is the simplest on-ramp - one prompt field, one negative field, one output, and Preview Anything does the rest. Just remember you're renting the model, and treat every run like the metered call it is.