LTX-2 19B · Video Extend

Continuation, not regeneration

Every video model in this pack can make a five-second clip. Far fewer can take the clip you already have and add five more seconds that actually belong to it. That's what this node does: you feed it a VIDEO, describe what happens next, and it returns your clip with a newly generated segment joined onto the end. The last frame of the source opens the new segment, so the join isn't a jump cut - it's the model picking up where you stopped.

This is the escape hatch when twenty seconds isn't enough. Chain two or three extends and you can build a minute-long piece out of one original composition, while each call only bills for the piece it generated.

The mechanism

Your clip gets remuxed to MP4 without re-encoding and uploaded. PopNow takes the final frame as the seed for the new segment, generates it at whatever resolution and duration you specify, and stitches it onto the front. The result comes back as a normal VIDEO - save it, then feed that into another extend node if you want to keep going.

Two limits worth memorising. The source clip can be MP4, MOV or WebM, up to 60 seconds. And the source is not billed - only the new segment is, per second, which is what makes chaining sane. Resolution is the price lever: roughly $0.014/s at 480p, $0.019/s at 720p, $0.029/s at 1080p.

Inputs that matter

• video_url - the clip to continue. The name says URL; you connect a ComfyUI VIDEO and the node handles the upload.
• prompt - what happens next, not a description of what's already there. Action, camera move, and sound. The tooltip nudges you to put the audio on its own line, same as the text-to-video node, and the same advice holds.
• resolution - and here's the trap: it sets the sharpness of the new segment only. Mismatch it against your source and you get a visible step in quality at the join. Match what the original was generated at, not what you wish it had been.
• duration_seconds - 5 to 20, length of the new part alone.
• seed - fixed by default; re-queueing reuses instead of re-billing.

Outputs: video (the whole thing, new segment included) and task id.

Install

Manager > search PopNow > Install, or:

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

Restart, then add a key from popnow.ai/api/console/keys in Settings > PopNow (or set POPNOW_API_KEY). The pack ships no dependencies - it builds on what ComfyUI already includes - so there's nothing to pip install and no model to download, which is the point: the 19B never touches your disk.

Where it bites

The 90 MiB upload ceiling. Long, high-resolution source clips get rejected before generation even starts. If your 1080p 60-second clip won't upload, re-export it shorter or smaller; the model can only extend what it can receive.

Continuity drift. Extends inherit the last frame, not the whole composition. Two or three joins in, characters have wandered, lighting has shifted, and the camera has developed opinions. It's the nature of autoregressive extension, not a bug in the node - plan for the number of joins your shot can survive.

Cost creep. Each extend is cheap on its own and expensive in aggregate. Six chained calls at 1080p is a couple of minutes of video at feature-short prices. Set Settings > PopNow > Max credits per run and keep an eye on the API console's log, which lists every task you've paid for.

Cancel ≠ refund. Stopping the wait in ComfyUI doesn't stop the job on PopNow. It finishes, and it's charged if it succeeds. If a run is going long and you've changed your mind, let it land and look at the result - you've already bought it.