VIDEO UPSCALER
Same shot, more pixels
Raise footage you already have to 1080p, 2K or 4K without changing a frame of what is on screen. Upload an MP4, paste a URL, or pick a clip you generated here and upscale it in place.
Updated:
What video upscaling does
Video Upscaling increases the resolution of a clip you already have and leaves its content alone. The subject, the framing, the motion and the timing all stay exactly as they were, and only the pixel count changes.
That makes it the delivery step rather than a creative one. Most video models cap below 4K, so the practical route to a high resolution master is generating at the resolution the model does well and raising it afterwards, once the content is locked.
The upscaling form, control by control
Five steps, and only one of them is a decision. The rest is telling the form where the footage is.
Video Enhancer is the last entry in the generation type list.
Open the Video Enhancer
Upscaling lives under Video Enhancer, at the bottom of the generation type dropdown. Unlike the four types above it, this one works on footage that already exists rather than producing new content.
That distinction matters for how you use it. Nothing about the picture changes here except how many pixels it is made of.
Choose Video Upscaling
Inside the enhancer a second dropdown offers Video Upscaling and Motion Control. These are genuinely different jobs sharing a form, with different models and different required inputs.
Video Upscaling raises resolution and leaves the content alone. Motion Control re-animates the footage from a reference character image, which changes what is on screen. Picking the wrong one produces a result that is not close to what you wanted.
A second dropdown splits the enhancer into upscaling and motion control.
MP4 up to 1000 MB, or a direct URL, or one of your own previous results.
Supply the source video
The Video Source block takes an upload, a direct video URL, or one of your previous completed tasks. The drop zone accepts MP4 up to 1000 MB, with a minimum resolution of 300 by 300 pixels.
Previous Task is the fast route when the clip came from this workspace. A shot generated at 720p can be upscaled without downloading and re-uploading it, which keeps both versions in the same Artifact Gallery.
Select the model
Upscaling has a single model, the ByteDance Video Upscaler, labelled in the picker with its available outputs of 1080p, 2K and 4K. There is no comparison to make here.
The absence of choice is worth knowing before you go looking for it. Where the enhancer offers a model dropdown with two entries, you are in Motion Control rather than Video Upscaling.
One model for upscaling, labelled with its 1080p, 2K and 4K outputs.
Three target resolutions, defaulting to 1080p.
Choose the target resolution
Target Resolution offers 1080p, 2K and 4K, defaulting to 1080p. This is the only creative decision in the upscaling form, and the honest answer usually depends on the source rather than the destination.
Where the source is 720p or below, 1080p is the setting that earns its cost. Jumping straight to 4K from low resolution material rarely produces a visible improvement over 1080p, because the detail being added was never captured in the first place.
The upscaling model
One model handles every target resolution here, so unlike the generation forms there is no comparison to work through.
| Model | Best for | Notes |
|---|---|---|
| ByteDance Video Upscaler Default | Raising resolution without changing content | The only upscaling model. Outputs 1080p, 2K or 4K, and bills per second of video with the duration detected automatically from the file. |
The Video Enhancer form holds two task types with separate model lists. Selecting Motion Control instead of Video Upscaling replaces this single entry with the Kling 3.0 Pro motion model and adds a character image input, so confirm the task type before reading the model dropdown.
Settings reference
Every control in the upscaling form, what it changes and the values it accepts.
Task Type
Splits the Video Enhancer into its two jobs. Video Upscaling raises resolution, Motion Control re-animates footage from a reference character image.
Video Upscaling or Motion Control
Video Source
The clip to upscale. Accepts an upload, a direct video URL, or one of your previous completed tasks from this workspace.
MP4 up to 1000 MB, minimum 300 by 300 pixels
Model
One model only. The ByteDance Video Upscaler handles every target resolution offered here.
ByteDance Video Upscaler
Target Resolution
The output resolution. Where the source is 720p or below, 1080p usually delivers as much visible improvement as 4K at a fraction of the cost.
1080p, 2K or 4K
Duration detection
Runs automatically against the file you supply, because billing is per second of video. A status indicator confirms when the analysis has finished.
Automatic, billed between 5 and 600 seconds
Picking a target the source can support
The one real decision in this form is the target resolution, and the highest option is not automatically the right one. Upscaling generates detail that was never recorded, and how convincingly it can do that depends on how much information the source actually holds.
From a 720p source, 1080p is a modest step the model can make cleanly. 4K from the same source is a much larger extrapolation, and the result frequently looks no better on a normal screen while costing more and taking longer. The gap widens as the source gets softer.
Where 4K genuinely earns its place is a clean high resolution source going to a large display or into a finish that will be graded and re-encoded. Both cases start from material with enough real detail for the model to extend rather than invent.
Where upscaling belongs in a sequence
Upscaling is the last pass. Every earlier stage might still be regenerated, and an upscale of a shot you then replace is work that produced nothing. Locking the content first and raising resolution second is the order that avoids paying twice.
A typical run generates a still, animates it at a short duration and a low resolution until the movement reads correctly, regenerates that shot at the length the edit needs, and only then upscales the approved clip. Each stage is cheaper than the one after it, which is the point of running them in that order.
Where the pipeline is repeatable, the Nodes Graph Editor runs it in one pass. A Video Enhancer node accepts the output of a Text to Video or Image to Video node directly, and it detects the duration of the incoming clip so the credit estimate reflects the real length rather than a guess.
What upscaling cannot fix
Resolution is one dimension of quality and upscaling only addresses that one. Motion that stutters, a subject whose face changes between frames, hands that deform partway through a movement, and lighting that shifts without cause are all still present at 4K. They are simply larger.
That is worth stating plainly, because a soft clip and a broken clip can look similar at 720p. Judging a generation at its native resolution before upscaling it is the way to tell those apart, and a clip with temporal problems needs regenerating rather than enlarging.
Compression artefacts and noise are a separate problem again. Upscaling a heavily compressed source can enlarge the blocking along with everything else, so material that has already been through a lossy export is a weaker candidate than a clip taken straight from a generation.
Generating high, or generating low and upscaling
Several models in the video roster can produce 1080p directly, and one reaches 4K in image to video, which raises a fair question about whether upscaling is needed at all. The answer depends on what else you want from the shot.
Generating at high resolution in one pass is the simpler route when the model that offers that resolution is also the model whose motion you want. Where those diverge, and the model with the movement you need caps at 720p, generating on that model and upscaling afterwards gets you both. The motion comes from the model that does motion well, and the pixel count comes from a step that does not touch the motion at all.
Cost pushes the same way during iteration. Testing a shot at 720p and upscaling only the approved take is cheaper than generating every attempt at 1080p, because the attempts you discard were never paid for at full resolution.
Related tutorials
Coverage of upscaling behaviour, and of the models whose output usually arrives here.
Video upscaling questions
Should I always pick 4K?
No. Where the source is 720p or below, 4K rarely produces a visible improvement over 1080p, because the detail being generated was never captured in the source. It costs more and takes longer for a result most viewers cannot distinguish. Match the target to what the source can actually support.
What file formats and sizes does it accept?
The drop zone takes MP4 up to 1000 MB at a minimum of 300 by 300 pixels. You can also paste a direct video URL, or select one of your previous completed tasks without uploading anything.
Does upscaling change what is in the video?
No. Upscaling raises resolution and leaves the content as it was. If you want to change what is on screen, that is Motion Control, which is the other task type inside the same Video Enhancer form and uses a different model entirely.
How is upscaling billed?
Per second of video, at a rate that rises with the target resolution. The duration is detected automatically from the file you supply, and billing is clamped to a minimum of 5 seconds and a maximum of 600 seconds. Wait for the duration analysis to finish before reading the estimate, so the figure reflects the real length.
When in a workflow should I upscale?
Last. Upscaling a clip you then decide to regenerate is wasted work, so lock the content at a low resolution first and raise it only once the shot is final. This is the same argument that applies to image upscaling.
Can I upscale a clip I generated in this workspace?
Yes. Choose Previous Task in the Video Source block and pick the completed generation. Nothing is downloaded or re-uploaded, and both the original and the upscaled version stay in the same Artifact Gallery.
Can I chain upscaling into a longer pipeline?
Yes. In the Nodes Graph Editor a Video Enhancer node accepts the output of a Text to Video or Image to Video node directly, so generating a clip and upscaling it runs as one automated pass rather than two manual submissions.
Other ways to make a video
Video Upscaling is one of six routes into the video workspace.
Video & LipSync
- Video Generator
- Text to Video
- Image to Video
- Start-End Frame to Video
- Draw to Video
- Motion Control
- Video Enhancer
- Video Upscaler
- Video to Video LipSync
- Audio to Video LipSync
- Image to Video LipSync
- Video FaceSwap
- Seedance 2
- Vidu Q3 Pro
- Gemini Omni
- Google Veo 3.1
- Kling 3.0 Pro
- Luma Ray 3.2
- LTX 2.3
- Happy Horse 1.1
- WAN 2.7
- Kling 3.0 Motion
- ByteDance Upscaler
- InfiniteTalk
- InsightFace