Reference, Extend, or Edit? Choosing an AI Video Workflow

AI video tools increasingly let us upload existing media instead of starting with text alone. But attaching a video creates a new problem: what exactly should the model do with it?

Depending on the tool, an uploaded clip might serve as:

• A visual reference for an entirely new generation
• The starting point for a continuation
• The source footage that should be modified

These workflows may look similar in the interface, but they solve different problems. Choosing the wrong one can produce a technically valid result that completely misses your intention.

This article presents a practical framework for deciding when to Reference, Extend, or Edit an AI video.

The Three Workflows Solve Different Problems

The easiest way to distinguish the modes is to ask one question:

Do I want a new video, more video, or a changed video?

That produces a simple mapping:

Desired result	Workflow
Create a new clip inspired by existing material	Reference
Continue what happens after an existing clip	Extend
Change something inside the existing clip	Edit

Car example: Reference creates a new scene, Extend continues the action, and Edit changes one element.

This distinction sounds obvious, but it becomes less clear when a request contains multiple intentions.

For example:

Keep the same character and visual style, move the scene to a rooftop, and continue the action.

Should the model create a new rooftop scene based on the character reference? Should it extend the uploaded scene until the character reaches a rooftop? Or should it transform the existing background?

Before writing the prompt, decide which part of the request is most important.

Reference: Create Something New From Existing Material

Use Reference when the uploaded video should guide a new generation without becoming its timeline.

The source material can communicate:

• Character appearance
• Product design
• Lighting
• Color palette
• Camera movement
• Animation style
• Motion rhythm
• Overall atmosphere

The resulting video does not need to begin where the reference ends. It can use a different location, composition, action, or narrative situation.

Example

Imagine you have a five-second video of a silver running shoe rotating under studio lighting.

You want a new scene showing the same type of shoe landing on a rain-soaked street at night. The source video provides information about the product's shape, materials, and visual treatment, but its existing timeline is irrelevant.

That is a Reference task.

A useful prompt might be:

Create a new cinematic night scene guided by the attached product reference. A silver running shoe lands on wet pavement, sending small droplets outward. Preserve the metallic mesh texture, dark sole, and reflective side details. Low tracking camera, blue storefront lighting, realistic material response.

The prompt describes the new result while explaining which properties should survive from the reference.

When Reference Is the Wrong Choice

Do not use Reference when the next frame must connect precisely to the final frame of the uploaded video.

A reference can communicate visual identity, but it does not necessarily preserve the exact position, movement, or timing at the end of the source clip.

If temporal continuity matters most, use Extend.

Extend: Continue From the End of a Video

Use Extend when the existing clip is already correct and you want to generate what happens next.

The final frames of the source video become the transition point for the continuation. This makes Extend useful for:

• Completing an unfinished action
• Holding a product shot for longer
• Adding an ending to a scene
• Continuing a camera movement
• Creating extra room for a title or call to action
• Lengthening a clip without rebuilding its beginning

Example

Suppose the shoe in the previous example lands successfully, but the original clip ends immediately after impact.

You want the camera to follow the runner for several more seconds before slowing into a final product frame.

That is an Extend task.

The prompt should focus on what happens after the current ending:

Continue from the final frame. The runner accelerates forward as the camera tracks beside the shoe at ground level. Water splashes naturally with each step. After several strides, transition into a slower side profile and finish on a clean frame with space in the upper-right corner.

Notice that the prompt does not redesign the beginning of the scene. It assumes the existing video remains unchanged.

When Extend Is the Wrong Choice

Extend cannot repair a mistake that already appears in the source footage.

If the product changes color halfway through the original clip, continuing the video will not remove that error. The model may even carry the inconsistency into the new section.

Fix the source first or use Edit if the workflow supports the required modification.

Edit: Change the Existing Video

Use Edit when the uploaded clip contains the timeline you want, but something inside it needs to change.

Typical Edit tasks include:

• Replacing a background
• Adjusting the visual style
• Changing weather or lighting
• Removing or replacing an object
• Modifying clothing or product color
• Reworking part of the motion
• Creating a variation of an existing shot

Unlike Reference, Edit is not primarily creating a separate scene. Unlike Extend, it is not adding material after the ending. It is transforming the attached clip itself.

Example

Suppose the shoe animation and camera movement are correct, but the background is a plain studio floor. You want to preserve the motion while moving the scene into a rainy city environment.

That is an Edit task.

A clearer prompt would be:

Preserve the shoe, its rotation, the timing, and the existing camera path. Replace the studio environment with a rain-soaked city street at night. Add realistic reflections beneath the shoe and soft blue storefront lighting. Do not change the shoe's color, proportions, materials, or movement.

For editing tasks, preservation instructions matter as much as transformation instructions.

A useful structure is:

Keep A, B, and C. Change D. Do not alter E or F.

This reduces ambiguity about which parts of the input are constraints.

A Practical Decision Framework

Before generating, answer these questions in order.

1. Must the original timeline remain visible?

• No: Choose Reference.
• Yes: Continue to the next question.

2. Should the output happen after the existing ending?

• Yes: Choose Extend.
• No: Continue to the next question.

3. Should something inside the current footage change?

• Yes: Choose Edit.
• No: Reconsider whether a video attachment is necessary.

The decision can also be summarized like this:

Do you need a new clip?
└── Yes → Reference

Do you need more footage after the ending?
└── Yes → Extend

Do you need to modify the current footage?
└── Yes → Edit

If your request seems to require two modes, split it into two generations.

For example:

1. Edit the original clip to correct the environment.
2. Extend the corrected result to create a longer ending.

Trying to perform both operations in one ambiguous instruction usually reduces control.

Person example: Reference creates a new setting, Extend continues the next moment, and Edit changes only the jacket.

Aspect Ratio and Duration Matter

The selected workflow can also affect which output settings remain flexible.

A new Reference generation may allow you to choose a different aspect ratio because it is creating a new composition.

Extend and Edit usually need to preserve more properties from the attached video. Changing the aspect ratio during these operations could require inventing visual information outside the original frame or cropping important content.

Editing may also stay close to the duration of the source because the operation modifies its existing timeline rather than adding a separate sequence.

The exact constraints vary by tool, so check them before preparing a prompt. One interface that separates these behaviors is the Seedance 2.5 AI video generator, where Reference, Extend, and Edit are presented as distinct operations after a video is attached.

Common Workflow Mistakes

Using Reference When You Need Frame-Level Continuity

The new result may resemble the source but begin with a different composition. Use Extend when the transition must feel continuous.

Using Extend to Repair Existing Footage

Extension adds to the ending. It does not automatically correct problems in earlier frames.

Using Edit With a Prompt That Changes Everything

If the prompt replaces the character, environment, camera, lighting, and action, there may be too little left to preserve. In that case, Reference is probably a better fit.

Failing to State What Must Remain Unchanged

"Make this more cinematic" leaves the preservation rules undefined. The model may change details you considered essential.

Name those details explicitly.

Combining Multiple Production Stages

A request such as "replace the background, continue the action, add a new product, and create a different ending" contains several jobs.

Separate them into an editing step and an extension step so each result can be reviewed independently.

A Pre-Generation Checklist

Before submitting, confirm the following:

• Am I requesting a new clip, a continuation, or a modification?
• Which properties from the source must remain unchanged?
• Does the result need to connect to the source's final frame?
• Am I trying to fix something that already exists?
• Does the operation need to preserve the original aspect ratio?
• Would two separate generations provide more control?
• Have I described both the desired change and the preservation constraints?

A short decision process can prevent several failed generations.

Final Thoughts

Reference, Extend, and Edit are not interchangeable labels for "use this video." They represent three different relationships between the source material and the output.

Use Reference when the source should inspire something new.

Use Extend when the existing ending should become the beginning of what happens next.

Use Edit when the timeline is useful but something inside it needs to change.

Choosing the workflow before writing the prompt makes the rest of the generation process much easier. It clarifies what the model should preserve, what it should invent, and where the output should begin.

---

Disclosure: I am affiliated with Seedance2AI.io. The product is included as one example of an interface that separates Reference, Extend, and Edit workflows.

AI assistance disclosure: This article was created with AI assistance and reviewed by the author before publication.