Kling Text to Video

This is the flagship. Text in, video out, no local video model, no VRAM math, no CUDA OOM - Kling's kling-v3 renders your prompt into a 5-to-15-second clip with sound and hands you the result as frames, a file, and an audio track. If you came to this pack for one node, it's this one. Where the local video ecosystem (Wan, LTX, and friends - see the KB's video docs) is a slog of model downloads and speed-LoRA tuning, this is the opposite pole: describe a scene and walk away.

Mechanically it's a cloud task, and the pack does all the boring work: your auth signs a JWT, the node posts the task to Kling's Singapore endpoint, then polls until the clip renders. The retry logic handles transient server errors with backoff, and IS_CHANGED means re-queueing re-renders instead of serving a cached clip.

Inputs that matter:

• prompt - the whole ballgame. It supports @image1 and @video1 references (assets you've uploaded), so you can anchor the scene to a specific subject: "@image1 walks into a neon bar". Describe motion, camera, and atmosphere.
• negative_prompt - what to avoid. Kling honors this better than you'd expect; use it for "blurry, extra fingers, warped faces."
• model_name - kling-v3 (default, latest), kling-v2-5-turbo (fast), kling-v2-6, kling-v2-master (cinematic), kling-v1-6. v3 is the default for good reason; v2-master is the one to try when you want a more filmic look.
• duration - 5/10/15s. Seconds are the most expensive thing on Kling's price list.
• aspect_ratio, shot_type (natural/wide_angle/medium_shot/close_up), mode (pro/std), sound (on by default - Kling generates ambient audio natively), cfg_scale (0–1, adherence vs creativity).
• camera_control - optional KLING_CAMERA from the Camera Control node for a forced dolly/pan/zoom.

Outputs: video (frames as IMAGE), video_file, audio (AUDIO - yes, the generated sound comes out as real audio), url, task_id.

Installing it

Part of ComfyUI-API-Toolkit. Manager: search "API Toolkit". Manual:

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

Restart ComfyUI. Then wire in the Kling AI Authentication node (access + secret key from app.klingai.com) and top up credits.

Gotchas

The KB's API-node essay nails the honest framing: Kling has no open weights, so this is the legit case for a cloud node - the community's objection to API defaults doesn't apply to a model you literally can't run locally. That said, cost is real and it stacks: 15s pro + sound is a chunk of credits per queue, so draft at 5s/std. The video tensor output holds every frame in RAM - for long clips, take video_file or url instead (or the Fast Video Saver). Content policy (error 1302) rejects at the source with no bypass, which is the permanent trade of a hosted model. And when Kling's server is slammed you may see error 1200 - the pack retries it for you with a 60s wait, so don't panic-kill the queue.