Flux 3 Video Continuation

Every video model has a length ceiling, and 20 seconds is where FLUX 3 stops. This node is how you keep going: it takes an existing clip and generates a continuation that carries on from its final frames. Same subject, same scene, same camera language - the new segment starts where the old one ended, so you can stitch a sequence that reads as one long take.

It's from the same FLUX 3 family as the text and image-to-video nodes, landed in ComfyUI core in August 2026, and it shares their DNA: a partner node that runs on Black Forest Labs' servers through Comfy's API proxy. You supply a video and a prompt, it returns a new clip. The continuation costs more per second than the other FLUX 3 modes - continuing is harder than starting - so the price badge will be the first thing you notice.

How it works

Your source video is uploaded to BFL, where the model reads its final frames as the starting state for the new segment. The prompt steers what happens next: keep the camera drifting the same direction, have the subject turn and walk off, or cut the action in a new scene that shares the same look. Because the input is the actual footage, not a description of it, continuity is the whole point - the model knows exactly what it's continuing from.

Like the rest of the family it also generates synchronized audio by default, so the continuation lands with its own ambient sound rather than silence you have to paper over.

The inputs that matter

• video - the clip to continue. This is the whole trick; feed it your previous FLUX 3 output or any video you want extended.
• prompt - what the continuation should show. It's interpreted and expanded before generation, so a sentence or two beats a paragraph.
• duration - "auto" or 5 to 20 seconds, for the new segment (not the whole thing).
• resolution - 720p or 1080p; the continuation is a fresh render, not an upscale.
• generate_audio - on by default; off gives a silent continuation.
• safety_tolerance - moderation strictness. Video requests are capped at 2 regardless of what you set, same as the other image/video FLUX 3 nodes.
• seed - another "change to re-run" control; FLUX 3 picks its own seed and ignores yours.

Output is one VIDEO, ready to save from the player or pass downstream.

Gotchas

Continuations inherit the source footage's look, so garbage in really is garbage out - a low-quality or badly framed source gives the model little to anchor on. The per-second rate is higher than the other Flux 3 nodes, so this is the one to budget for: a 20-second 1080p continuation is a real chunk of credits. And as always with the cloud nodes, the work finishes on BFL's side even if you disconnect mid-poll, so don't assume an error means nothing ran.