Fal Text-to-Video

Text-to-video is the purest version of the cloud-render deal: type a prompt, get a clip, never find out how many gigs the checkpoint would have cost your card. Fal Text-to-Video is the node that does it from ComfyUI, one of eleven in the ComfyUI-Fal-Gateway pack. The dropdown pulls live from fal's catalog - Veo 3.1 T2V, Kling V3 T2V, Seedance T2V, Wan 2.6 T2V, and over a hundred more - so the closed frontier is one selection away from your canvas.

The honest framing before you get excited: T2V is the hardest mode to control. You're asking a model to invent a subject, its motion, and a coherent scene from a sentence. It's why image-to-video took over as the community favorite for consistency work. Where T2V genuinely shines is open-ended scenes - establishing shots, abstract motion, "a city built of glass at dawn" - where you don't have a reference to anchor to and local models would cost you an afternoon of VRAM thrashing for worse output.

Inputs and outputs

Just two required inputs, the cleanest node in the pack:

• model_id - the dropdown. [blackforestlabs] Flux 3 Text to Video, [bytedance] Seedance 2.5, [alibaba] Wan Text to Video, Kling, Veo - provider-prefixed and type-ahead searchable.
• prompt - multiline. Write motion, camera, and subject together; these models reward describing what happens, not listing attributes.

Four outputs:

• frames - the clip as an IMAGE batch. Wire into VHS_VideoCombine or SaveVideo to save.
• video_url - the source mp4 on fal's CDN.
• audio - AUDIO socket, extracted from the mp4. Comes back None for silent clips, which is safe downstream.
• info - JSON string of the full result (seed, timing, safety flags).

How it works

Same machinery as every node in the pack: the job is submitted to the fal endpoint for the model you picked, the cloud renders, and the node downloads the mp4, decodes it into frames with OpenCV, and pulls out the audio track. No weights on your disk, one API call, one bill. There's no image upload here - the payload is prompt-only, which is why this is the node people use to test whether a closed video model is even worth their money before wiring up the I2V pipeline.

Install

Not in the Manager registry yet, so clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/modbender/ComfyUI-Fal-Gateway
pip install -r requirements.txt
export FAL_KEY=your_fal_key

Restart ComfyUI and the pack appears under Fal-Gateway. Dependencies are the usual modest set - fal-client, aiohttp, opencv-python, Pillow, numpy, pydantic.

Gotchas

• Watch the cost badge. The title bar colors the estimated price green → red as you change widgets, and video is the reddest category on fal. One clip can outprice an afternoon of local Wan renders - which is the fair trade for a model you literally cannot run locally.
• Cold catalog on fresh nodes - you may briefly see <no models available> while the first fetch runs. Right-click → "Fal-Gateway: refresh catalog cache", and restart to surface brand-new models.
• No key, no run. "FAL_KEY not set" is the error, fal.ai/dashboard/keys is the fix.

If you already run Wan locally, this node isn't a replacement - it's the door to the models that never shipped open weights. For everything else, keep your local sampler on the job.