How to Edit a Finished Video with AI Using Marked Frames

To change a finished clip with AI, you no longer need to describe in words where exactly the object is. You pause the clip, circle the object with a box right on the frame and write what to do with it. The model receives both the marked picture and the text, and changes exactly what you circled.

This technique solves the main pain of text-based video editing: the phrase "remove the bag from the table on the right" works as long as there is one bag and one table in the frame. After that the guessing begins.

TL;DR

• What it is. Upload a clip, put a box, an arrow or a label on the right frame, write the edit, and get a new clip.
• Where it lives. The Seedance 2.5 Video Editor action in the Actions section.
• Input. MP4 or MOV, up to 48 MB, from 2 to 30 seconds long. Up to 30 marked frames per request.
• Price. From 64 tokens per clip; the total depends on the duration and the resolution, and the page shows the exact amount before the run.
• How it works. The model rebuilds the clip anew following your markup; the result stays in your gallery.

Why markup on the frame works better than a description

A text prompt for video is an address without a street. The model knows what to look for but not where. In a clip with two people, the phrase "change the jacket to a coat" is a lottery.

A marked frame removes the ambiguity: you give the model a picture where the right object is already circled and refer to it in the text. What works next is not a search by description but pointing with a finger.

The technique relies on the ability of video-to-video models to accept not only the source clip but also a set of reference images. In Seedance this is the reference mode: up to thirty hint pictures plus the clip itself.

Don't do this: "remove everything extra and make it beautiful". The model does not know what is extra for you and will return its own version of beauty. Circle a specific object and write what to replace it with.

What it looks like step by step

1. Upload the clip. MP4 or MOV, up to 48 MB, from 2 to 30 seconds. Trim a long video in advance.
2. Find the frame. A strip of thumbnails under the canvas; the left and right arrows step through time, with Shift exactly one second.
3. Mark it up. A box, an arrow, a brush, a caption, a numbered marker or an eraser. The tools switch with the letters R, A, B, T, P, E.
4. Apply the frame. The markup is burned into the picture, the frame goes into the reference set, and a chip like @image-1 is added to the prompt.
5. Describe the edit and run the generation.

You can apply several frames, for different moments of the clip. Each gets its own number, and in the text you refer to exactly that one.

What the editor itself has

The canvas is not a picture with a box but a working surface.

• Six tools. The box selects an area, the arrow sets the direction of motion, the brush outlines a complex contour, the caption writes text right on the frame, the numbered marker sets a point for step-by-step instructions, the eraser removes what is extra.
• Editing what you drew. A shape can be selected, dragged, resized by the corners and deleted with the Delete key. The arrow and the brush stroke can also be rotated; the box and the caption have no rotation on purpose: the model accepts them only along the axes, and a rotated shape would come back unrotated.
• Undo and redo. Ctrl+Z and Ctrl+Shift+Z, with history for the whole drawing session.
• Zoom and pan. The mouse wheel zooms toward the cursor, and holding the space bar temporarily turns any tool into a "hand". The keys +, − and 0 zoom in, zoom out and fit.
• A timeline with thumbnails. Twelve frames under the canvas, with ticks where you have already marked something: you see what is done and do not have to mark twice.

Returning to a frame to edit it

We keep the vector of the markup, so a click on a preview in the feed brings the markup back to the canvas: the clip jumps to the right timecode, the shapes are back in their places, and you can draw more or erase what is extra.

The edit replaces the frame in its own place: the number and the order are kept. This matters more than it seems: the model addresses pictures by their position, and if the frame moved to the end of the list, your phrase "@image-1 remove the earring" would start referring to someone else's picture.

If there is unapplied markup on the canvas, the editor will not overwrite it: it will first ask you to apply or reset it.

Settings visible before the run

On the right are the model parameters, and they change before generation, not after:

• Aspect ratio: from vertical 9:16 to cinematic 21:9, plus "auto" from the source.
• Quality: 480p, 720p, 1080p.
• Duration: from 4 to 30 seconds.
• Bitrate and native sound: the model can generate a soundtrack itself.

The cost is recalculated on every change and shown right on the button. This is not cosmetic: for this model the resolution surcharge goes per second, and switching from 480p to 1080p changes the total many times over. The finished clip appears right there on the page, with a download link: you do not have to look for it in the gallery.

The prompt is assembled automatically: your text, the markup legend ("the red box is the object to replace") and a continuous timing scale. The latter is a requirement of the model: the segments must go without gaps or overlaps, otherwise it picks at random which moment an instruction belongs to.

A prompt you can copy

The wording that works most consistently: a reference to the frame, then the action, then what to keep.

@image-1 replace the brown bag on the table with a black backpack of the same size.
@image-2 remove the reflection in the glass on the right.
Keep everything else (people, lighting, camera movement and the order of events) unchanged.

The last line is not decoration. Without an explicit "leave the rest alone", the model considers itself entitled to rethink the whole scene: the light changes, a different angle arrives, the people in the frame get different clothes.

Tips for a clean result

• Use edit verbs. Write "replace", "remove", "add". The words "continue" and "extend" switch the model into clip continuation mode, where it picks the format and duration itself. The page warns about such words before the run.
• Keep the markup out of the result. The model sometimes takes your red box for part of the scene. A thinner line and a color that contrasts with the background help; on plain surfaces this happens more often.
• Send only the fragment you need. The seconds of the input clip are billed too, so trim the clip to the fragment with the edit.

How much it costs

The cost is per clip and depends on the duration and the resolution: the longer and the higher the quality, the more. The starting point for this model is from 64 tokens, and the page shows the exact amount on the button before the run, before any charge.

The saving tip is simple: first run the edit at 480p on a short fragment, make sure the model understood the task, and only then repeat it at the resolution you need.

Further reading

• AI video enhancement: if the clip needs not changing but sharpening and a higher resolution.
• Seedance 2.5: three video modes: a breakdown of the model itself, its modes and token costs.
• What Seedance can do: the whole model family.

The model is developed by ByteDance: the Seedance family is described on the official ByteDance Seed page.

Frequently asked questions

Does the source clip stay the same? Yes. The model does not edit your file but assembles a new clip. The source stays with you.

Can I mark several moments in one clip? Yes, up to thirty frames per request. Each gets its own number, and in the prompt you refer to the right one with @image-2. If you delete a frame from the feed, the numbers of the others shift automatically, together with their mentions in the text.

Why is my red box visible in the finished clip? The model sometimes takes the markup for part of the scene. A thinner line and a color that contrasts with the background help; on plain surfaces it happens more often.

My clip is longer than 30 seconds: what should I do? Trim it to the fragment you need before uploading.

How is this different from a text prompt for video? Text describes the object, markup points at it. The difference shows where the frame has several similar items: in words you have to tell them apart by description, with a box in one move.

In short

Markup on the frame is a way to show the model with your finger instead of explaining in words. It works wherever the object is visible and can be circled: replace an item, remove what is extra, change a detail of clothing.

You can try it on the Video Editor page: a clip, a box, a phrase. To run it, you need to create an account: new users get welcome tokens to start.