Kling 4.0 · Reference to Video

This is the node for "I have a character, an outfit, a location and a piece of footage I like the camera from - make me a shot". Instead of one start frame you hand it sets of reference images, videos and audio clips, then address them by tag in the prompt: @Image1, @Video1, @Audio1. From about 32.4 credits per second, and reference video seconds are billed too, at a lower rate than output seconds.

The mechanism

Everything you connect gets uploaded to PopNow - images as PNG/JPEG (10 MiB ceiling each), video re-muxed to MP4 (90 MiB ceiling), audio converted to 16-bit WAV - and Kuaishou's model composes the shot from them. The node quotes first, submits with an idempotency key, polls with progress on the node, and returns the finished clip as a VIDEO. No local model, no IP-Adapter, no reference tricks to assemble by hand.

The three reference inputs are autogrow socket groups: connect one image to reference_image_urls and a second socket appears, up to ten. Numbering follows connection order, which is what @Image1, @Image2 refer to.

Inputs

prompt - say what each reference is and what the shot should do with it. Name them explicitly: "@Image1 walks into the frame wearing @Image2's jacket, camera moves like @Video1". Vague references are the number one reason this node produces something generic.

reference_image_urls - up to 10 images: characters, objects, outfits, places.

reference_video_urls - up to 5 clips, 30 seconds total, for motion, camera and pacing. Their seconds are billed too, at a lower rate, so a 20-second reference isn't free.

reference_audio_urls - up to 3 clips of 5–30 seconds each, for a voice or a sound.

keep_original_audio - off by default. Turn it on if you supplied a reference video and want its sound kept instead of generating new audio. Only used when there is a reference video.

duration_seconds - 3 to 30, per-second billing. aspect_ratio - auto or a fixed ratio. resolution - 720p or 1080p.

Outputs

video (VIDEO) and task id (STRING). The video behaves like any other ComfyUI video - Save Video, or another PopNow node downstream. The task id is what the API console and support use.

Install

Manager → search PopNow → Install → restart, or:

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

Then a key from popnow.ai/api/console/keys into Settings > PopNow, or POPNOW_API_KEY in the environment. Restart again so the catalog is read from your account. No model downloads, no dependencies, works on a machine with no GPU. Delete custom_nodes/popnow_api.py if it's still around from the single-file release.

What goes wrong

Uploads are cached by content hash for 12 hours per API key, so re-running with a tweaked prompt doesn't re-send your ten reference images. That's the good news. The bad news is the 90 MiB per-video cap - reference footage adds up fast when you connect five clips, and the error is a flat "larger than the 90 MiB upload limit".

Cost sneaks up in two places: the reference video seconds are billed on top of output seconds, and 30-second outputs at this rate are nearly a thousand credits a take. Put a ceiling in Settings > PopNow > Max credits per run and let the node refuse the expensive accidents for you.

Cancelling in ComfyUI stops the wait, not the job - the task keeps running on PopNow, is charged if it succeeds, and the result appears in the API console. 40201 means empty wallet, 401 a dead key, 400/40004 a field or combination the model won't take.

And a workflow note: this node pays off when your references are consistent. Ten images of the same character from different angles gives the model a consistent identity. Ten images from ten different artists gives it ten arguments.