OpenGradient (∇, ∇) (@OpenGradient): Introducing Extend Video: Make AI Videos Longer Than the Clip Limit

Introducing Extend Video: Make AI Videos Longer Than the Clip Limit Extend Video is now live in the DreamFort Video Studio. Pick a clip you already have, type what happens next, and the model keeps the shot going from where it ended. You get back one video: the original followed by the extension, joined with no cut. Every video model has a per-clip ceiling. Most of the lineup stops at 15 seconds, Seedance 2.5 at 30. That ceiling was the hard edge of what you could make in the studio, and the most common message we got for months was some version of "the take was going great, and then it ended." Extend is how you get past that edge. What it does Open Extend Video from the Post-production row, or press Extend on any clip's result page. Pick one of your recent clips, or upload a video from your device. Describe what happens next and hit Generate. The part that matters is how it continues. The model is handed the whole clip, not a screenshot of the last frame. It watches the motion, the subject, the camera move and the sound, and carries all of it forward. That's why the join doesn't read as a join: a camera drifting left keeps drifting left, the ambience stays, a person mid-gesture finishes the gesture. We tried the other way first. The obvious implementation is to grab the final frame and feed it to an image-to-video model, and for a week that's what the mode did. It looks fine as a still and wrong the moment it plays: the shot restarts from a photograph, so motion resets to zero at the seam and the audio drops out for a beat. We pulled it. Extend now only offers models that take a video as input, which today means Seedance 2.0 and Seedance 2.0 Mini. Fewer models in the picker, but the mode does what the name says. The clip's own model picks the continuation model where it can, and you can switch between the two. Each extension adds 5, 6, 8, 10, 12 or 15 seconds, at the quality you choose, with or without audio. There's no aspect-ratio picker on purpose: the extension always matches the clip it continues. One file, not a folder of parts When the extension lands, the studio joins the two clips into a single video in your browser, using the Movie Editor's renderer, at the source's size and shape. That joined file is the result: it's what plays, what you download, and what goes to your library, labelled as the combined clip. It's also what you extend next. Press Extend on the result and the whole joined video becomes the new source, so a chain of extensions grows one file instead of leaving you with six numbered fragments to line up later. The extension on its own is kept as a library clip too. If round three went sideways, you don't re-render rounds one and two. Go back to the previous combined file in your library and extend that again. Building a minute-long take, fifteen seconds at a time This is the use we built it for. Here's how to do it. Start with the best opening you can get. Everything downstream inherits the first clip's look, framing and pace, so spend your effort here. For the longest possible first leg, render it on Seedance 2.5 at 30 seconds. Any clip in your library can be extended, regardless of which model made it. Write each extension as the next beat, not a recap. The prompt box starts empty, and that's deliberate. Don't describe the scene again; the model has just watched it. Say what changes: "she reaches the end of the corridor and pushes the door open, light floods in." One beat per extension, the same rule that works for a fresh prompt. If you'd rather not write it yourself, the Enhance prompt toggle rewrites your idea as a continuation. It reads the clip's first and last frame and its original prompt, and keeps the subject, camera, light and sound consistent instead of restarting the shot. Repeat. Fifteen seconds, then thirty, forty-five, sixty. Four rounds gets you a one-minute continuous take, which not long ago wasn't something these models could make at all. Know where it drifts. Every extension is a fresh generation conditioned on the clip, and long chains accumulate small changes: a jacket colour shifts, a face softens. Two things help. Keep the prompt specific about what you want held steady, and check each round before you extend again, because a mistake in round two is in every round after it. When one slips, go back a step rather than fighting it forward. Other things it's good for The clip that ended a beat too early. The most common one. The render was perfect and it cut before the payoff: the glass never hits the floor, the hand never reaches the door. Six seconds of extension lands the beat you already paid for. Directing a scene one action at a time. The oldest prompting advice is "one action per clip", because three actions in one prompt fight over five seconds. Extend makes that advice stackable. She pours the coffee. Extend: she turns to the window. Extend: the cat lands on the table. Same camera, same light, one shot, and you approve each action before writing the next. Real footage that runs out. You can upload a video from your device, not just something generated in the studio. A product spin from your phone that stops a quarter-turn short, a drone shot that ends before it clears the ridge, b-roll two seconds shorter than the edit needs. Uploads stay in your browser until you press Generate; the file goes up only then, with its own progress bar, and the cap is 100 MB. Adding an ending to a Recipe. Recipes pin a format and a length. When the Award Ceremony winner needs one more beat, the walk off stage with the trophy, Extend takes the recipe's output like any other clip. Longer ads from a short hero shot. A 10-second product orbit becomes a 25-second spot: extend once for the pack shot, once more for the line on screen. Then send the combined file to the Upscaler or into the Movie Editor, both one click from the result page. What it costs The price shows in credits before you generate, same as everywhere in the studio (1,000 credits = $1). One detail worth knowing: because the model watches the clip it's continuing, the source's length is part of the bill alongside the new seconds, at the model's with-video rate. The source is counted up to 30 seconds, so a long chain doesn't get more expensive per round forever. Seedance 2.0 Mini is the cheaper of the two if you're blocking out a long sequence and want to spend the good credits on the final pass. Privacy, as usual Extend runs through the same anonymity layer as the rest of the studio. Requests are encrypted on your device, identity is split from content by Oblivious HTTP relays, and decryption happens only inside attested secure enclaves. No single party holds both who you are and what you made. That matters more for a feature that takes your own footage as input. A clip of your kid, your unreleased product, a face that hasn't agreed to end up in anyone's training set: that's the material you should least have to hand over under your own name to get more of it. Go make something long Extend Video is live now at dreamfort.ai, free to start, no card. Make the first clip in Text to Video, then press Extend. If you build something that runs past a minute, we want to see it. And if the chain drifts somewhere it shouldn't, tell us where; the join is the part we'll keep working on. Seedance is developed by ByteDance. DreamFort is not affiliated with ByteDance. Written By Adam Balogh, Co-founder of OpenGradient. Originally published on the DreamFort blog.