Writing Seedance 2.5 video edit prompts: removing is easy, replacing is hard

In September we ran a round of tests on Seedance 2.5's video edit: same source clip, one variable changed at a time, then a close look at what the model changed and what it left alone. Here are the conclusions first:

• Removing something works on the first try. Replacing something with something else failed every time. When asked to replace, the model adds the new thing to the frame and leaves the old one where it was.
• Once it actually changes one spot, it redraws the whole frame, not only the part you asked about.
• Boxes and arrows drawn on a reference image don't help, and they can end up rendered in the output.
• Timecodes work sometimes: they held in an 18-second clip and did nothing at all in a 4-second one.

What "video edit" means here

In the video panel's Task mode, pick Edit video, attach a 4–30 second source clip and write what you want changed. We add the edit instruction to the copy of the request we send upstream; your prompt is left exactly as you wrote it. Output length and aspect ratio follow the source.

Removing things: the wording that worked first time

Every run that succeeded followed these five rules:

1. One action per request. Don't remove, replace and move things in the same prompt.
2. Describe the location in words. "The red apple on the right side of the table", not a circle drawn on the image.
3. Say what should fill the gap. For example: "Fill in naturally with the existing wall, door frame and floor, with no ghosting, blur patches or human outlines." Without this the model tends to leave a smudge, or puts another person in.
4. List everything that must not change. If you only say what to change, the model changes other things too. In one run the person wasn't replaced, and extra people showed up instead.
5. Skip timecodes when the change applies to the whole clip.

An example written to those five rules:

Remove the red apple on the right side of the table. Fill the area naturally with the existing wood-grain tabletop, with no ghosting or blur. Keep the white porcelain cup on the left, the steam above it, the background and the lighting unchanged.

Replacing things: why it's hard

Replacing asks the model to do three things at once: remove the original, generate the new one, and match your reference image. In our tests it managed only the middle one. The new person was added to the frame and the original person stayed, so the scene simply gained people.

If you have to replace something, at minimum:

• write it as two explicit steps: "Remove the existing X, and put Y in its place";
• say where the new subject stands, how big it is and which way it faces.

We should be honest: we don't yet have a replacement that works reliably, so don't expect it on the first try.

Change one thing and the whole clip gets redrawn

We measured how close each result stayed to the source (PSNR: higher means closer, and above 40 dB the difference is essentially invisible):

How to read this: the model can leave a clip almost untouched. Once it really changes one spot, though, it redraws the whole frame, including the areas you explicitly asked it to keep. Wording doesn't matter here. The three phrasings landed within 0.2 dB of each other.

Three practical consequences:

• The more edits you stack, the more faces, costumes and textures drift. Each edit redraws the whole frame again.
• But if you put two changes in one request, the model usually does only the first. That happened in all three runs where we tried it. Several changes mean several runs, so do the most important one first.
• If you only want to change a short part of a long clip: an edit is billed on the full source length, and the full length gets redrawn. Cutting that part out as its own clip (at least 4 seconds) and editing only that should in principle cost less and leave the rest untouched. We haven't yet tested on a long clip whether the join looks seamless when you put it back.

Boxes and arrows: put them into words

A red box around a person or a yellow arrow pointing somewhere on the reference image means nothing to the model. The official reference guide warns about this too: don't rely on labels printed inside images, and when using a grid image, explicitly tell the model not to draw its borders, arrows and lines. In other words, the model treats them as picture content.

Describe the box instead: "the person deep in the corridor, smaller than the figures in the foreground, standing behind the smoke".

Timecodes: sometimes they work

We tried them twice and got opposite results:

• An 18-second clip, with "00:00-00:05 keep unchanged / 00:05-00:08 replace…": the change did land between 5 and 8 seconds, and the first 5 seconds barely moved.
• A 4-second clip (locked-off camera, a porcelain cup on the left and an apple on the right), same "remove the apple" instruction, only the timecode wording changed: second half only, both halves, or no timecodes at all. All three came out the same. The apple was gone from second 0, and "00:00-00:02 keep the original content unchanged" was ignored completely. As a control, when both halves said "keep unchanged" the apple stayed, which shows the check itself is sound.

The two tests differ in four ways: clip length, remove vs. replace, number of segments, and a moving person vs. a still object. So we can't yet say which one made the difference. Length is the top suspect, since 4 seconds is exactly the shortest source an edit accepts.

If you do use timecodes, write them as continuous ranges, as the official guide recommends (0-5s: …, 5-8s: …), covering the clip from start to end with no gaps.

Limits of this round

We couldn't fix the random seed in this round, so the same input gives a slightly different result every run. That's why each conclusion rests on a yes/no check, such as whether the object is still there, rather than on small numeric differences. Treat the similarity figures above as orders of magnitude only.