Kling 3.0 · Text to Video

Text to video is where these hosted models still embarrass local pipelines, because local video wants a start image, a stack of nodes and twenty minutes. Here you type a paragraph and get a clip. Kling 3.0 Text to Video runs from 10.08 credits per second - a five-second 720p take is about 50 credits, five cents - and it's the first node in this pack where the model will cut a multi-shot sequence for you.

What the node is

An API call with a nice panel. It builds a request from your widgets, takes a price quote, submits the task, polls it, and hands back an MP4 as a VIDEO. Kuaishou renders it. You never touch a checkpoint, a VAE, or a sampler, which is either the appeal or the dealbreaker depending on what you're doing.

Inputs that matter

prompt - scene, action, camera, and (with audio on) the dialogue in quotes and the sounds you want. Write it as beats if you plan to use shot_type: multi.

shot_type - single keeps one continuous shot. multi lets the model split your prompt into several shots by itself. customize follows your own list.

multi_prompt - a JSON text field: a list of objects, each with prompt and duration_seconds, in order. The durations must add up to duration_seconds or the request is rejected. Blank is fine, blank means "not sent", and malformed JSON stops the node with a "must be valid JSON" error before anything is charged.

duration_seconds - an integer 3 to 15, billed per second. The full range matters: you can cut a 7 or 11 second clip, which is far more edit-friendly than 5-or-10.

resolution - 720p, 1080p, 4k. Test at 720p, finish at 4k, watch the badge change as you do.

aspect_ratio - 16:9, 9:16, or 1:1. Note there's no 21:9 here; that's Kling 4.0 territory.

generate_audio - off by default and costs more per second. When it's on, the model generates voice, effects and ambience alongside the picture, which is the reason to use it. When it's off you're dubbing later.

guidance_scale - 0 to 1, default 0.5. Adherence dial, not CFG. negative_prompt - a short exclusion list.

Outputs

video (VIDEO) and task id (STRING). Wire the video to Save Video, or into a PopNow edit/extend node if this clip is a first pass. The task id is your handle in the API console.

Install

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

Restart, then paste a key from popnow.ai/api/console/keys into Settings > PopNow or export POPNOW_API_KEY. Restart a second time so the node list comes from your account's catalog rather than the bundled snapshot. Manager route: search PopNow, Install. There's genuinely nothing else - no requirements, no model files, no license gate. If you still have custom_nodes/popnow_api.py, delete it.

Where it gets expensive, and where it fails

Per-second billing plus a 15-second ceiling plus 4k is a combination that can quietly cost real money per take. Put a number in Settings > PopNow > Max credits per run on day one; runs quoted above it are refused before any credits are held, which turns an expensive mistake into a harmless dialog.

Cancelling in ComfyUI doesn't cancel the job. PopNow keeps generating, charges you if it succeeds, and the result appears in the API console. If you cancel, go look there before re-queueing.

Error codes you'll meet: 40201 empty wallet, 401 bad key, 40004 a settings combination the model won't serve, 400 naming a field for actual input mistakes (usually multi_prompt). content_rejected or generation_failed are refunded.

One honest note on the model itself: Kling 3.0's reception in the community was genuinely divided - there are threads calling it a downgrade from 2.6 and threads insisting it's the best video model you can call. It's very prompt-sensitive. If a take comes out dull, change the camera line before you buy a 4K one.