Upscaling Media in Luma: A Complete Guide to Higher-Resolution Video and Images
Upscaling is how you prepare an asset for higher-resolution final delivery.
In Luma, however, video and image upscaling are different workflows.
Video has a dedicated upscaler that increases the resolution of an existing video to 1080p or 4K. It is designed as a finishing step after the creative work is complete.
Images do not use the same dedicated upscaler. Instead, you can:
- Generate the image at a higher resolution from the start.
- Regenerate an existing image at a higher resolution with an image model such as Nano Banana Pro.
- Use Split Grid to separate and upscale individual images from a grid.
Understanding that distinction makes it much easier to choose the right workflow.
Upscaling video
To upscale a video, select it on the board and tell the Agent the resolution you want.
For example:
Upscale this video to 1080p
Upscale this video to 4K
You can also right-click the video and choose Upscale from the context menu.
Luma creates a new higher-resolution version, leaving the original source untouched.
There are two output resolutions to choose from.
1080p is a strong choice for social, web, monitors, and general HD delivery.
4K is better suited to large displays, premium presentations, and high-resolution client delivery.
The right choice depends on where the asset is going. If the final video will primarily be watched online, 1080p may be all you need. If it needs to hold up on a large screen or as a premium final asset, 4K gives you the additional resolution.
Upscale at the end of your workflow
The most important rule for video is:
Upscale last.
Upscaling is a delivery step, not a working step.
Complete your creative workflow before increasing the resolution:
generate → modify/color/VFX → reframe → trim → stitch → add audio → UPSCALE → download
If you upscale first and then continue reframing, modifying, coloring, trimming, or otherwise processing the video, those later operations can resample or re-encode it.
That can negate the benefit of the upscale and waste the credits you spent creating it.
This is also why there is usually little value in upscaling work-in-progress versions. Iterate at normal resolution and upscale the version that actually becomes your final cut.
What video upscaling does—and what it does not do
The video upscaler increases resolution and reconstructs detail.
It is not a replacement for fixing problems earlier in the creative pipeline.
If your source already contains issues such as motion artifacts, banding, compression damage, or other generated artifacts, increasing its resolution will not necessarily repair them.
Solve those problems upstream.
Then use the upscaler to prepare the finished asset for delivery.
Video limits, processing, and cost
A single video upscale supports clips up to 120 seconds.
The technical guidance lists typical execution at roughly 47 seconds, with 4K generally taking longer than 1080p.
Cost is calculated per second.
A 1080p upscale is approximately 10 credits per second, while a 4K upscale is approximately 17 credits per second.
Because cost increases with duration, trim your video to the actual delivery cut before running the upscale.
There is little reason to spend credits processing footage that will not appear in the final asset.
Working with videos longer than two minutes
Videos longer than 120 seconds cannot be sent through the upscaler as one job.
The technical guidance recommends handling these as a segmented workflow:
split → upscale each segment → re-stitch
Where possible, make the splits on hard edits. This can reduce the chance of visible seam artifacts between processed sections.
You can split a video via right-click → more actions → split.
You can also trim any video through the same menu, upscale and then select the videos you want to cut back together in the order you want them concatenated and just tell the agent to stitch these videos together in this order
The guidance also recommends keeping the original audio on the master and laying it back over the reconstructed sequence afterward.
This is a special case compared with the normal upscale last rule. For standard clips, upscaling remains the terminal step. For videos above the 120-second limit, segmenting the video requires a separate long-form workflow.
Getting higher-resolution images
There are three main paths: generating at the final resolution, generatively upscaling an existing image, or using Split Grid for grid outputs.
Generate at your final resolution
If you are creating the image yourself, this is the preferred approach.
Ask for the resolution you need when the image is generated:
Create a product shot of a matte-black water bottle on marble, 4K.
Image models can generate at resolutions such as 1K, 2K, or 4K, depending on the model.
If you already know the asset needs to be 4K, creating it at 4K from the beginning gives you a cleaner workflow than trying to reconstruct a lower-resolution version later.
Whenever possible, perform the rest of the image-editing workflow at that final resolution as well.
Upscale an existing image generatively
If you already have a low-resolution image, you can select it and ask the Agent to recreate it at a higher resolution with an image model such as Nano Banana Pro.
For example:
Upscale this image to 4K with Nano Banana Pro.
Keep composition, subject, colors, and lighting identical.
Increase resolution and micro-detail only — no creative changes.
This is different from the dedicated video upscaler.
The model is regenerating and reconstructing details at a larger resolution rather than simply enlarging the existing pixels.
That means the result can look sharper and more detailed, but some details may not be perfectly identical to the source.
Tell the image model what to preserve
When generatively increasing image resolution, preservation instructions matter.
Do not only tell the Agent what should change. Tell it what should stay the same.
Depending on the image, that might include:
- Composition
- Subject
- Colors
- Lighting
- Face
- Pose
- Wardrobe
- Background
- Framing
For example:
Upscale to 2K. Preserve the character's face and pose exactly. Sharpen textures without changing wardrobe, background, or framing.
Being explicit reduces unnecessary reinterpretation.
Know when generative image upscaling is risky
Because the image is being reconstructed, small details can drift.
Use extra caution with:
- Logos
- Small or precise typography
- Legally or brand-sensitive text
- Fine graphic details
- Identity-critical faces
For exact logos or typography, the technical guidance recommends generating the asset at the required resolution from the source rather than relying on generative upscaling.
For face-sensitive work, consider using GPT Image 2 where maintaining identity is more important.
The underlying principle is simple: the more exact the asset needs to remain, the less you should depend on generative reconstruction.
Avoid stacking multiple image upscales
Do not progressively send the same image through:
1K → 2K → 4K
Each pass gives the model another opportunity to reconstruct and reinterpret the source.
Small differences can compound.
If you know the final target is 4K, aim for 4K in a single clean pass.
The same principle applies to repeated image editing. Running the asset through multiple generative passes can gradually reduce fidelity to the original.
Upscaling images from a grid
If an image contains a grid of several outputs, use a split grid method.
Split grid separates the individual cells and can upscale each one up to 4K.
For example:
Split this 2x2 grid into four separate images at 4K.
This is especially useful when you used a grid to explore several visual directions and now want to turn individual cells into standalone assets.
The tool can also handle the gutters and borders between cells while separating them.
Extracting higher-resolution frames from video
A frame extracted from a standard video is limited by the resolution of that source video.
If your final goal is a high-resolution frame grab, the technical guidance recommends upscaling the video first, then extracting the frame.
That keeps the operation in the deterministic video pipeline rather than extracting a low-resolution frame and then generatively rebuilding it as an image.
Choosing the right workflow
For a final video going to the web or social, upscale to 1080p as the final step.
For a video intended for a large screen or premium delivery, upscale to 4K as the final step.
For a video longer than 120 seconds, use the split, upscale, and re-stitch workflow.
For a new high-resolution image, generate at 2K or 4K from the start.
For an existing low-resolution image, use a generative upscale with a preservation-focused prompt.
If a logo, typography, or another detail must remain exact, generate the asset at the final resolution rather than relying on generative reconstruction.
For identity-critical faces, use extra caution with generative upscaling.
If you need an HDR or VFX-grade master, use an HDR or EXR-capable workflow instead of the standard video upscaler.
For high-resolution individual images from a grid, use split grid.
A helpful reminder
Higher resolution should be intentional, not automatic.
For video, finish the creative work before spending credits on a final-resolution version.
For images, generate at your target resolution whenever possible. If you need to reconstruct an existing image at a larger size, clearly state what must remain unchanged and avoid repeatedly passing the asset through generative models.
In both cases, one well-planned final pass is better than repeatedly increasing resolution throughout the workflow.
Key takeaway
Think of resolution as a finishing decision.
For video, use the dedicated upscaler at the end of production and choose 1080p or 4K based on the destination.
For images, prioritize generating at the final resolution, and use generative upscaling only when you understand that the model may reconstruct some of the original details.
Choose the workflow based on what needs to stay exact, where the asset is going, and what kind of final master you actually need.