Luma Dream MachineImage-to-videoStyle references
Luma Modify Video Guide: Restyle Footage, Keep the Motion
Modify Video in Luma Dream Machine takes a clip you already have, your own footage or a generation, and produces a new version that keeps the original motion and timing while changing its look, setting or subject according to your instruction. Pick a low adherence setting for look changes, a higher one for subject swaps, and describe the change rather than the whole scene.
Most video generation starts from nothing. Modify Video, added to Dream Machine in 2025, starts from footage. You give it a clip, a short instruction and a choice of how closely to follow the original, and it returns a video whose camera moves, timing and performance are preserved while the surface is transformed. It is the tool for turning a phone test shot into a stylised sequence, previewing a wardrobe or set change, or restyling a Ray2 generation without losing a motion you liked. Luma changes labels and options often, so if a control described here is named differently in your interface, use the one you see.
What Modify Video actually keeps
The model reads the motion in the source: how the camera moves, how subjects travel through the frame, the rhythm of the action. That skeleton is kept. What changes is everything built on top of it: materials, lighting, environment, style and, at stronger settings, the subject itself.
This means the source clip’s quality as motion matters far more than its quality as an image. A shaky, badly lit phone clip of someone walking through a kitchen is a perfectly good source if the walk and the camera move are what you want. Conversely, a beautiful clip with no motion gives Modify Video little to work with. Think of the source as a performance and camera plate, not as a picture.
Choosing the adherence strength
The strength or adherence control decides how far the output may depart from the source. The current interface offers a small set of levels rather than a slider, and the names may differ, but they map onto three jobs.
- Close adherence keeps composition, subjects and most detail. Use it for lighting, weather, time of day, colour grade and texture changes. "Make it golden hour", "turn the scene into a snowy evening", "give it the look of 16mm film".
- Medium adherence keeps the layout and motion but allows materials, costumes and backgrounds to change. "Replace the kitchen with a ship’s galley", "she is now wearing a long red coat".
- Loose adherence keeps the motion skeleton and little else. Use it to swap subjects wholesale or move into a different medium. "The man is now a bronze statue come to life", "render the whole clip as a watercolour animation".
Start closer than you think you need. If the change does not come through, step one level looser. Going loose first produces results that have drifted in ways you did not ask for, and you spend generations pulling them back.
Writing a modification prompt
A Modify Video prompt is an instruction, not a scene description. The model already knows the scene from the footage.
- State the change plainly. "Turn the daytime street into a rainy night with neon reflections."
- Name what must stay if it is at risk. "Keep the woman’s face and the camera move unchanged."
- Describe the target look in a few concrete words: palette, era, medium, grain. "Cool blues and magentas, 1980s film stock, light grain."
- One change per pass. If you want a new setting and a new costume and a new style, do them as three passes, closest first, so each can be judged on its own.
Avoid naming artists, films or brands as shorthand. Describe the qualities instead, which works better and keeps you clear of imitation. The general advice on describing a look in words is the same as for style references in image tools; see image prompts for the vocabulary.
Settings at a glance
| Setting | What it does | Start with |
|---|---|---|
| Source video | The clip whose motion is kept; your footage or a Dream Machine generation | A steady clip with clear subject and camera motion |
| Adherence / strength | How closely the output follows the source | Closest level for look changes, step looser as needed |
| Instruction | The change to make | One change, stated plainly, with what must stay |
| Reference image | Where offered, a still that shows the target look | Add one when words are not nailing the style |
| Resolution | Output size; upscale options for finals | Draft at the lower setting, upscale the keeper |
| Duration | Follows the source clip, within the model’s limits | Trim the source to the seconds you need first |
Preparing source footage
A little preparation makes the difference between a usable result and a muddy one.
- Trim first. Cut the source to the exact seconds you need, within the model’s duration limit, so no credits go on footage you will discard.
- Stabilise if the shake is not intentional. The model will faithfully keep a wobble.
- Keep the subject clear of the edges so the model does not have to invent what happens as they leave frame.
- Avoid fast cuts in the source. Modify Video treats the clip as one continuous motion; a cut inside it confuses the result.
- Match the aspect ratio to your delivery format, or plan to Reframe afterwards.
- Use only footage you have the right to transform. Your own shots and your own generations are safe ground.
For Dream Machine generations as sources, the workflow pairs well with keyframes: get a motion you like with a start image, then restyle it with Modify Video rather than re-prompting. The keyframes, loops and extend guide covers the first half of that.
Try it
Source: a 5-second handheld phone clip of a friend walking towards the camera along a plain corridor, daylight, 16:9.
Adherence: closest level.
Instruction: Turn the corridor into the interior of a 1970s passenger train at night, warm tungsten light, rain streaking the windows, slight film grain. Keep the person's face, clothing and walk exactly as they are.
When the setting change lands, run a second pass one level looser with "she is now wearing a long camel coat and carrying a leather case" to see how costume changes behave. On a third pass, go to the loosest level with "render as a hand-painted animation" and compare how much of the original walk survives each step.