Seedance 2.5 Video Editing and Extend: The 3 Modes Everyone Confuses

I wanted to change one thing in a video I'd already generated. One jacket. That's it.

What came back was a different room, a different camera move, and a person who no longer looked like the person I started with.

I did that three more times before I understood what was actually wrong. I wasn't writing a bad prompt. I was writing a reference prompt and expecting an edit result. Those are two different modes, and Seedance treats them completely differently.

ByteDance's official practice guide has a section on exactly this confusion. It says, in plain terms, that reference, edit, and extend get mixed up constantly — and that the mix-up is what makes results unpredictable. Once I read that section, my hit rate went up more than any prompt-writing trick ever did.

Here's the whole thing, including the part buried in the guide's fine print: what editing silently locks that you can't override.

The Mistake That Ruins Most Edits

There's a simple test that tells you which mode you're in:

Do you already have a finished video that you want to keep?

• No — you're generating from scratch, using images, video, or audio as inspiration. That's reference.
• Yes, and I want to change something inside it — that's edit.
• Yes, and I want more of it before or after — that's extend.

That sounds obvious written down. It isn't obvious at 1am when you're on your ninth attempt. What I was doing was feeding my finished video in as if it were inspiration, then describing the whole scene again from scratch. The model did exactly what I asked: it made a new video inspired by mine. Of course the room changed. I'd asked it to invent a room.

Reference vs Edit vs Extend: Three Modes, Three Prompt Shapes

Each mode has its own prompt shape. Get the shape right and most problems disappear.

Edit and extend look identical on paper. The difference is in the verb: edit rewrites inside the existing runtime, extend adds runtime before or after it.

What reference mode actually covers

Reference isn't just "here's a photo of my character." The official guide breaks it into eight sub-abilities, and knowing they exist changes what you think to ask for:

The white-model one is worth calling out because it's genuinely useful and almost never mentioned. Feed in a rough grey 3D animation, feed in a material reference image, and get a finished render that follows your blocking exactly:

Keep the camera movement, duration, composition, shot size, spatial
relationships, object positions, model structure, and motion paths in
@video1 unchanged. Use @image1 as the reference for material, lighting,
color, and overall mood. Replace the grey-model material in @video1 with
realistic material close to @image1, adding natural light and shadow,
contact shadows, ambient light, reflections, highlights, and depth, for an
overall cinematic render.

Motion reference is the other underrated one. This whole prompt is enough to move a climbing performance onto a different body entirely:

Transfer the motion in @video1 to a gorilla climbing a cliff face.
Do not change the camera rhythm or the movement.

The 3-Step Edit Formula

The official guide gives three steps for writing an edit. They're short enough to memorize:

1. Point @video1 at the original.
2. Name what changes and what it becomes. If you're replacing a subject, attach @image2 for the new appearance. If you're adding or removing something, describe the object and where it is.
3. Lock the rest. Add an explicit "everything else stays unchanged."

Step 3 is the one people skip, and it's the one that saved my jacket. Without it the model treats the untouched regions as fair game.

Here's the same edit written badly and written well:

❌  Change the man's jacket to a leather one and make it look better.

✅  Replace the clothing of the man in dark clothes in @video1 with the
    clothing in @image2. Everything else in the frame stays unchanged.

The second one names the target precisely ("the man in dark clothes," not "the man"), supplies the new appearance as a reference, and locks the rest. Three sentences, three jobs.

5 Edit Commands You Can Copy

All five are adapted from ByteDance's official guide. Notice how short they are — editing prompts are supposed to be short. The long, detailed prompt belongs in reference mode.

Remove a subject

Video edit: remove everyone from @video1 except the lead.

Or, for cleaning up a shot:

Remove the drone and the foreground rail from @video1. Fill the removed
areas naturally. Everything else stays unchanged.

This is the fastest way to fix a boom mic, a crew reflection, a rig, or a watermark you don't own.

Add a subject

Keep the people, environment, and composition in @video1 unchanged.
An energy bow appears in the character's hands.

Replace a scene, keep the performance

Replace the setting in @video1 with a medieval stone castle terrace —
castle walls, wind, mist, distant ridgelines, flat stone ground @image1.
Replace the clothing of the man in dark clothes with @image2, and the man
in light clothes with @image3. Keep the choreography exactly as it is;
do not change the original rhythm.

Relight the whole film

Change the lighting in @video1 from harsh midday overhead light to low
side light at sunset.

Name the current state and the target state. "Make the lighting better" gets you nothing, because the model has no idea what you dislike.

Restyle without touching the blocking

Keep the camera movement and rhythm of @video1. Change the visual style
to black-and-white comic, then Japanese anime, then American hardcore
comic, then ink-wash animation.

Swap the audio only

Keep the picture and camera movement in @video1 unchanged. Replace the
background music with @audio1.

Editing Just One Slice of the Timeline

You don't have to edit the whole clip. Seedance 2.5 accepts timestamps inside an edit instruction:

Replace the setting in @video1 between 0:02 and 0:05 with a medieval stone
castle terrace. The character's movement and rhythm stay unchanged.

This is the difference between a targeted fix and a full re-roll. If a single beat in a 30-second clip is wrong, fix the beat.

How to Extend a Clip: Forward, Backward, and In Between

Extend keeps the picture, pacing, and style of the original and writes new footage around it. Three directions:

Extend forward — start from the last frame and continue:

Starting from the last frame of @video1, write 6 more seconds: the lead
walks out of frame, the sky darkens, street lamps come on one after
another. Picture and audio join seamlessly.

Extend backward — build the lead-in:

Treating the first frame of @video1 as the end point, add 4 seconds of
lead-in: the lead approaching from a distance before entering frame.
Match the visual style of the first frame.

Bridge two clips — generate the missing transition:

Generate a 3-second transition between @video1 and @video2, seamlessly
matching picture, color, and camera movement.

And you can chain extensions. The official guide walks through a three-link chain: a bee lands on a flower, then the flower goes to seed, then the seed falls into soil. Each step extends the previous output:

Step 1 — Extend @video1 by 5 seconds: a bee flies in and lands on the
flower; a macro close-up shows pollen on its legs and abdomen; it takes
off and the camera follows it to another flower of the same species.

Step 2 — Extend @video2 by 5 seconds, time-lapse feel: the yellow petals
wither, the center swells, and a green pod grows, colors, and fills out.

Step 3 — Extend @video3 by 5 seconds, time-lapse: day and night flicker
past, the ripe pod splits open, and a seed slides out; the camera follows
it through the air and into damp soil @image1.

Three prompts, one continuous sequence, no editor.

What Editing Locks That You Can't Override

This is flagged in the official guide's own "pay attention" callout, and it explains a whole category of "the tool ignored my settings" complaints.

When you use edit, extend, or first/last-frame generation, the model locks some generation parameters based on your input material. You cannot set them yourself.

Editing:

• Output aspect ratio is locked strictly to the original. Your edit will not stretch or crop the frame.
• Output duration matches the original, near enough. Slight differences can appear because of how the model processes input, but the body of the video stays intact.

Extending:

• Output aspect ratio is locked strictly to the original, so the new footage joins the old at the same frame size.
• Duration of the new segment is yours to set — say it explicitly in the prompt ("extend by 6 seconds").

First-frame and first/last-frame generation:

• Output aspect ratio follows the first-frame image.
• If the last frame has a different aspect ratio from the first, the last frame gets stretched to match — which is exactly how you get that weird warp at the end of a clip. Feed in first and last frames with matching dimensions.
• Duration is yours to set.

That last bullet is worth taping to your monitor. Mismatched first/last frame dimensions are a silent, self-inflicted distortion, and nothing in the interface warns you.

Seedance 2.5 Video Editing FAQ

Can Seedance 2.5 remove an object from a video? Yes. Feed the clip in as @video1, name the object, and ask for the gap to be filled naturally. Add "everything else stays unchanged" so the model doesn't wander.

Can I edit only part of a video? Yes. Include a timestamp range in the instruction — for example, "between 0:02 and 0:05" — and the rest of the timeline is left alone.

Will editing change my aspect ratio or length? No. Editing locks the output aspect ratio to the original and keeps duration effectively the same. That's automatic and not adjustable.

How long can an extension be? You choose, and you should say so in the prompt. Unlike editing, extension duration is a parameter you control.

Can I replace the music without touching the picture? Yes. Audio editing is its own sub-ability — supply @audio1 and lock the picture and camera movement.

What's the difference between reference and edit? Reference generates a new video inspired by your material. Edit rewrites inside a video you already have and keeps everything you didn't mention. Using the wrong one is the single most common cause of unusable results.

The Bottom Line

Almost every failed Seedance edit I've made was a mode error, not a prompt error.

Pick the mode first. No finished video, reference. Finished video you're changing, edit. Finished video you're growing, extend.

Keep edit prompts short and locked. Point at the original, name the one thing that changes, then say everything else stays unchanged. Three sentences beats three paragraphs.

Remember what's locked. Editing fixes your aspect ratio and duration to the original. First/last-frame generation stretches a mismatched last frame. Neither one warns you.

Seedance 2.5 opens on August 7, 2026, but the three-mode discipline works on Seedance 2.0 today — you can generate and edit with Seedance 2.0 on Tryonr right now, watermark-free.

Next: the prompt formula itself, in the Seedance 2.5 prompt guide. Or see exactly what changed from 2.0 in Seedance 2.5 vs 2.0. The model overview lives on the Seedance 2.5 page.

Prompts in this guide are adapted from ByteDance's official Seedance 2.5 practice guide.