Zhenzhen_kling_text2video

This is the workhorse of the Kling nodes in the Comfly pack: text in, video out, and - unusually for an API wrapper - a real set of creative controls instead of a prompt box and a prayer. You get the model pick, a duration, a mode, a camera-movement menu, and Kling's signature imagination slider. If you're going to generate video through a proxy, this is the node that feels like driving rather than being a passenger.

Kling's been a recurring name in the closed-video conversation for a while - the model the community kept measuring against - and this node hands you its API without a Kuaishou account. Paid per call, cloud inference, GPU idle. And the README's warning applies double here: some of these models (the kling-v2-master tier especially) are expensive, so read the price list before you make master your default.

How it works

Prompt and settings go to the Kling endpoint via the proxy; the node polls until the clip renders (Kling runs five to six minutes per generation, so this is a "run it, do something else" node). Video comes back as a ComflyVideoAdapter output that plugs into ComfyUI's Save Video node, plus video_url, task_id, video_id, and response.

The inputs that matter

• model_name - kling-v2-1-master, kling-v2-master, kling-v1-6, kling-v1-5, kling-v1. The v2 masters are the quality/price top tier; kling-v1-6 is the reliable all-rounder. Start there.
• prompt / negative_prompt - the usual pair; Kling reads both.
• imagination - 0–1, default 0.5. This is Kling's "how far from the literal prompt do I wander" dial. Low = faithful, sometimes stiff; high = creative, sometimes off-brief. It's the single most useful control here, and the one people forget exists.
• mode - std / pro. Pro costs more for visibly better motion quality.
• duration - 5 or 10 seconds, billed per tier.
• aspect_ratio - 1:1, 16:9, 9:16.
• num_videos - 1 to 4 takes per call. All four get billed, so don't default to 4 "for options."
• camera - the fun one: none, horizontal, vertical, zoom, vertical_shake, horizontal_shake, rotate, and the master_* moves (down_zoom, zoom_up, right_rotate_zoom, left_rotate_zoom). Pick a move instead of writing "camera pans right" and hoping.
• camera_value - -10 to 10 (default 0), the intensity of that camera move.
• seed - lock a take.

Outputs: video (VIDEO), video_url (STRING), task_id (STRING), video_id (STRING), response (STRING).

Install

Manager → Comfyui_Comfly, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ainewsto/Comfyui_Comfly

Restart. Standard requirements.txt - no downloads, no VRAM. Key from ai.comfly.chat in Comflyapi.json or the node's api_key.

Where people get burned

Cost control, in order: num_videos 4, duration 10, mode pro, model master. Any one of those is fine; all four at once is a wallet event. And the imagination slider cuts both ways - leave it at 0.5 for a while, because turning it to 0.9 and getting a clip that ignores your prompt is not a bug, it's the knob doing its job. Save the video_id/task_id outputs if there's any chance you'll lip-sync the clip later.