Your clip ends two seconds too early and you do not want to regenerate it and lose the take you liked. Extend it from its own last frame with a dedicated extension model, or continue from that frame with an image-to-video model, and write only what happens next.
You generated a clip you love, and it ends two seconds too soon. The reflex is to rerun the prompt with a longer duration and hope the new take keeps the same face, the same light and the same camera move. It rarely does. Extending the clip you already have is the better move: the model starts from the last frame of that footage and writes only the next few seconds, so the part you liked stays untouched.
A video model gives you a new roll of the dice every time you press generate. If the first take has the right subject, the right mood and the right motion, a longer rerun throws all three away and asks you to win the lottery again. Extension keeps the winning ticket. Your footage stays exactly as it is, and only the new seconds are generated, starting from where the picture stopped.
It also changes how you plan a scene. Short generations are easier to get right than long ones, so many creators build a sequence in pieces: a strong opening shot first, then a continuation for each beat that follows. Each piece can be judged on its own and thrown away without losing the others, although every extension still costs credits, which is worth remembering before you chain six of them.
Picasso IA lists 502 models, 28 of them in the video editing category, and the extension model lives there next to the other tools that work on footage you already have. You can browse the full catalog or the video editing collection to see what sits around it.
There are two routes, and they trade convenience for control. The first is a dedicated extension model: xai/grok-imagine-video-extension, listed as Grok Imagine Video Extension, takes your MP4 plus a sentence about what should happen next and generates the continuation from the last frame. The second is manual: export the final frame of your clip as a still, then feed it to an image-to-video model as the starting image and prompt the next beat yourself.
| Approach | What it starts from | Best for | Watch out for |
|---|---|---|---|
| Grok Imagine Video Extension | Your MP4, from its last frame | Adding seconds to a clip you already have | Format and length limits on the input clip |
| Last frame plus Wan 2.6 I2V | A still image of the final frame | Choosing duration and resolution yourself | The still loses motion, so speed can reset |
| Last frame plus Kling v2.6 | A start image, set as the first frame | Continuing with a cinematic look and audio on | Style can shift away from the first clip |
| Last frame plus Grok Imagine Video | An image in image-to-video mode | Keeping one model family across both clips | The aspect ratio must match the original |
| One longer generation from scratch | Only a text prompt | Planning a clip before the first take exists | You lose the take you already liked |
The still-frame route works with any image-to-video model, so the three models in the table are examples, not a closed list. Pick a model that accepts a start image and test with a short run first. The extension model is the shortest path; the still-frame route is the fallback when your file is refused, or when you want another model to write the next beat.
This is the dedicated-model workflow. The still-frame route follows the same shape, with the export step added in the middle.
If the extension model refuses your file, export the last frame, open an image-to-video model such as Wan 2.6 I2V or Kling v2.6, attach the still as the starting image and prompt the next beat. The result is a separate clip that begins where the first one ended, and you join the two yourself.
The extension prompt is not a scene description, it is a stage direction. The model already sees the last frame, so it knows the subject, the place and the light. What it cannot see is your intent, and that is the only thing the prompt needs to carry: what moves, in which direction, and what the camera does while it happens.
Compare two prompts for a clip of a lighthouse at dusk. A prompt that repeats the whole scene, a tall white lighthouse on a rocky coast under a purple sky, invites the model to reinvent the scene. A prompt that says the beam swings slowly toward the camera while the waves roll in gives it a single motion to deliver. Short, physical and specific beats long and decorative.
Four habits keep the continuation on track. Ask for one action per extension, because two actions in a few seconds usually means neither is done well. Name the camera, saying whether it holds still, drifts forward or pulls back, since an unspecified camera tends to wander. Keep the subject description short: repeating the clothes, the face and the setting gives the model room to change them. And match the pace of the original, because a slow clip followed by a fast continuation feels like two different videos.
Ask only for what is plausible from that frame: a person facing left cannot cleanly turn their back to the camera in two seconds.
Chaining is where most disappointments happen. Each extension starts from a frame the model generated, not one you filmed, so small errors in a hand, a logo or a patch of texture become the reference for the next round and get copied forward. After a few links the clip can look like a close relative of the original, not the original.
Extend once, review the seam, and only then extend again. If the second extension already looks different from the first, stop and rebuild the scene from a clean still instead of chaining a third.
Treat every extension as a checkpoint: save the last good frame and the prompt that worked, and if a link goes wrong, rerun that link, not the whole chain. For a recurring character, build the face with a reference image first, as described in consistent AI characters, so every new clip starts from the same identity. For a longer story built from many short pieces, AI short films and trailers covers how to plan the beats.
Sound needs its own pass: add background music or sound effects once the video is final.
Extension is a patch, not a magic wand, and it stops being useful in a few specific places:
If you fight the same seam a third time, the clip wants a new shot, not a longer one. Cut to another angle: viewers accept a cut far more readily than a face that quietly changes.
Yes, with limits. A dedicated extension model reads the last frame of your clip and generates new footage that continues from it, so the part you already have stays as it is. You can also export that last frame and continue it with an image-to-video model. In both cases you are generating only the new seconds, which is why the result can drift a little at the join.
The dedicated one is xai/grok-imagine-video-extension, listed as Grok Imagine Video Extension in the video editing category. You upload an MP4, write what should happen next and choose how many seconds to add. For the manual route, any image-to-video model that accepts a start image works, such as Wan 2.6 I2V or Kling v2.6. The catalog changes, so check the model pages for what each one accepts today.
You take the final frame of your clip as a still image, upload it as the starting image of an image-to-video model and write the next action. The model animates from that frame, so the new clip begins where the old one ended. The weak point is motion: a still carries no speed or direction, so you describe those in the prompt and join the two clips in an editor afterward.
You set the number of seconds you want to add in the extension model, and the model page lists the lengths it accepts for both the input clip and the continuation. Those limits can change, so check the page rather than trusting a number written here. For a much longer result, build from several short pieces and review each seam.
Mostly, because the model starts from the exact last frame and keeps the light, the subject position and the visual style. But it is generating, not copying, so small details can change: a hand, the folds of a jacket, lettering in the background. Review the seam at full size, and rerun with a shorter, more specific prompt if something shifted.
Yes. Every extension is a generation, and generations are paid in credits, with the cost depending on the model and the settings you choose. New accounts get free credits to try it. Chaining repeats the cost each time, so test with a short extension first. Current plans are listed on the pricing page, which is the only place worth trusting for numbers.
Often yes, as long as the file meets the extension model's input rules. It expects an MP4 in a supported codec and within a limited length, which many phone clips already satisfy, though some need trimming or converting first. The clip does not have to come from Picasso IA. If the file is refused, the last-frame route with an image-to-video model avoids the format restriction.
There is no magic number, because the limit is quality, not a counter. Each extension starts from a generated frame, so small errors accumulate and the subject slowly stops resembling the original. Extend once, review the seam, and continue only while the clip still looks like itself. When it starts to slide, go back to the last good frame and rerun from there instead of adding another link.
Only if the person is you, or the character is fictional. Do not extend footage of a real, identifiable person who has not agreed to it, and do not use a continuation to put words or actions into someone else's mouth. Extensions generate new behavior from a frame, so the new seconds show things the person never did. Use your own footage or invented characters.
Plan for that, because the extension model is built around the picture and the audio rarely carries over cleanly. Check the model page for what the output includes, then rebuild the soundtrack after the video is final: add music with a music model, or layer effects on top of the finished clip.
Ready to give a short clip its missing seconds? Open the Picasso IA toolkit, pick your best take and extend it from the last frame.
Choose your language
