Wan 2.5 Preview Image-to-Video (fal)

Here's the thing that makes this node worth understanding before you use it: you cannot run Wan 2.5 on your own machine. The open Wan line - the one everyone builds LoRAs and fine-tunes on - froze at Wan 2.2 back in July 2025. Alibaba took the numbered series behind an API after that, and 2.5, 2.6 and 2.7 never shipped weights. So if you want to feed a still image to the newer Wan and get motion out, an API bridge like this is basically the only door. That's the whole reason this node exists.

What it actually does

Nothing here runs locally. You give it a starting image and a prompt, the node packages that up, ships it to fal.ai, fal runs Wan 2.5 on its own hardware, and hands you back a finished clip. No VRAM, no 14B checkpoint download, no waiting for a two-expert MoE to load. In exchange your image and prompt leave your machine and every run costs real money billed by fal.

If you're already comfortable running Wan 2.2 locally, the honest question is whether the newer preview is worth paying for versus just staying on the open weights you own. For a lot of work it isn't - 2.2 with a good fine-tune is still excellent. But 2.5 is genuinely a step up on prompt adherence and motion coherence, and you don't need a rented GPU to try it.

The inputs that matter

• image (required): your first frame. Wan animates outward from this, so a clean, well-composed still gets you a lot.
• prompt (required): describe the motion and the scene. Write like you're directing a shot, not tagging a booru.
• resolution: 480p, 720p, or 1080p (default 1080p). Higher costs more and takes longer.
• duration: 5 or 10 seconds. Ten seconds is roughly double the bill.
• variations: 1 to 10 clips from one submission. Handy for rolling options, brutal on cost - each one is a separate paid generation.

Two more worth knowing: negative_prompt comes pre-filled with a sensible quality-guard string, and enable_prompt_expansion (on by default) lets fal rewrite and pad your prompt. Expansion helps a terse prompt but will change your wording, so turn it off when you want exactly what you typed. There's also a seed for reproducibility.

The output is a STRING list - video URLs, not a playable clip inside ComfyUI. To actually get the file, wire it into the pack's Fal Save Media from URL node, which downloads the result into your output/ folder.

Installing it

Easiest path: open ComfyUI Manager, search ComfyUI-fal-API, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
cd ComfyUI-fal-API
pip install -r requirements.txt

Restart ComfyUI. There's no model to drop in models/ - the real dependencies are the fal-client package (pulled in by requirements.txt) and a fal API key. Copy config.ini.example to config.ini inside the pack folder and paste your key under [API], or export FAL_KEY before launching. No key, no video.

Where people get burned

The number-one issue is the missing key. After that, watch the cost: variations up to 10 times duration of 10s at 1080p is a real tab, and it's easy to trigger without noticing. The pack caches identical calls so an exact re-run is free, but change any input and it's a fresh charge - so lock a seed while you dial in the prompt.

If the node loads but every run errors, read the message: this pack surfaces fal's actual error text (content policy, validation, quota), and it usually names the parameter to fix. And if you're on ComfyUI Windows Portable and get an import error, install the client into the embedded interpreter directly: .\python_embeded\python.exe -m pip install fal-client.