Kling 2.1 Master · Text to Video

Here's the whole node in one sentence: type a scene description, pick a frame and a length, pay $0.216 per second at 1080p, get a silent clip back. Five seconds is $1.08, ten is $2.16.

The honest question is whether text-to-video at the Master tier is the right purchase. Usually it isn't, and it's worth understanding why. Kling's motion quality is real, but text-to-video throws away the one thing that makes these models reliable in a pipeline: a reference frame to hold. You can't control the composition, so you spend a $1+ clip discovering the model put the subject in the wrong third of the frame. The people who get value out of this node use it for shots where the place matters more than a specific framing - a storm over a city, traffic at night, a forest in fog - and then cut it into a timeline, not for anything with a character who has to hit a mark.

The alternative route is the one most working graphs take: generate a still you're happy with (any of the image nodes in this pack, or your local model), then animate it in the image-to-video node. It costs slightly more per attempt in some setups and much less in wasted iterations, because you're throwing away bad stills at a few cents instead of bad clips at a dollar.

What's on the node

• prompt - scene, action, camera. With no reference frame to anchor it, spend your words on the camera: "slow dolly in, 35mm, golden hour" steers more reliably than adjectives about vibe.
• aspect_ratio - 16:9, 9:16 or 1:1. Pick before you prompt; the frame changes the composition the model plans.
• duration_seconds - 5 or 10, billed per second.
• negative_prompt - a short list of what to avoid. Works on this generation.
• guidance_scale - 0 to 1, default 0.5. This is a literalness dial, not CFG: higher follows the prompt more closely, lower gives more freedom. Cranking it up generally gets you a stiffer clip rather than a better one.

Outputs: video, a ComfyUI VIDEO (Save Video, a preview node, or a video-stitch node - don't plug it into image nodes) and task id (STRING, the run's handle in the SpicyAPI console).

There's no audio. This generation is silent; sound arrived with 2.5 Turbo and the 2.6 tier, and if you want sound you should price that node instead.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart ComfyUI, open Settings > SpicyAPI, paste a key from spicyapi.ai/console/keys, restart once more so the model list is fetched from your account. Manager search: "SpicyAPI". Server or cloud box: set SPICY_API_KEY. The pack needs no models and no GPU - it's an HTTP client with widgets - so the whole install is a git clone and a key.

Before you spend anything

Set Max cost per run. It's the one setting in this pack that changes behaviour rather than convenience: the node quotes every job first and aborts, uncharged, if the quote is above your ceiling. At Master's rate a couple of accidental queues is real money.

Know that cancelling doesn't cancel. Interrupting a run stops your wait. The job keeps going on SpicyAPI, is charged if it succeeds, and its result stays in your console for up to 14 days. You can also set seeds to randomize (control after generate) when you genuinely want a fresh roll - the default is fixed, precisely so that re-queueing doesn't silently charge you again.

Expect the standard errors: No SpicyAPI key is set if the settings step was missed, 40310 if your account email isn't verified, 40004 when the option combination you picked isn't served, and a proxy/firewall message if HTTPS_PROXY is set but broken.

Restart for new models. The node list is fetched at startup. A model that isn't in your search results isn't missing - it just wasn't in the catalogue the last time ComfyUI booted.