GPU Video Enhancement on Windows: Test First, Then Batch
· ClapClip Team · 7 min read
Plan from a short sample, not from a specification sheet. ClapClip runs its models on your GPU through DirectML, and the only reliable way to know what a clip will cost is to process thirty seconds of it and multiply. Choosing an output resolution before that test is guessing.
Key takeaways
- The local path uses DirectML on NVIDIA, AMD or Intel graphics. The file stays on your PC; the browser path is the one that uploads.
- Processing time is set by the mode, input resolution and length, output resolution, available VRAM and other GPU load. There is no universal speed multiplier.
- 4 GB of VRAM is guidance for 1080p work, and 8 GB or more for 4K output or a batch. That is product guidance, not a certified minimum.
- Batch mode applies one settings pass to every file in the queue. Split mixed sources instead of hoping one setting suits all of them.
When is the local GPU path the right choice?
Use it when:
- The footage is private, client-owned, or large enough that uploading it is unattractive.
- You have several files to process and want one queue with one set of settings.
- You need video at all. The browser tool is a still-image tool: JPG, PNG or WebP, one image, 2× or 4×, PNG out.
- You want to preview without committing to a full render.
Look elsewhere when:
- A single photo is all you need; the browser needs no install.
- You are on macOS or Linux. ClapClip is a Windows 10 and 11 application, with no Mac or Linux build.
- The machine has no GPU the driver exposes to DirectML. The app targets NVIDIA, AMD and Intel graphics through DirectML, and the Windows page covers the platform requirement.
Rule: if the job is one photo, open the browser; if it is a clip or a folder, open the app.
What actually determines the processing time?
There is no universal speed multiplier, and any article quoting one is describing a specific machine. For a given clip, the time is set by:
- Model and mode. General, Anime and Portrait do different amounts of work per frame.
- Input resolution and length. More frames and more pixels per frame mean more work.
- Output resolution. 4K output is four times the pixels of 1080p for the same duration.
- Available video memory. Too little headroom slows the job or makes it fail.
- Other GPU load. A game, a browser with hardware acceleration, or a second render competes for the same card.
That is why the workflow below measures first and scales second. A sample from your own clip is the only number that applies to your machine.
Rule: if a test is too slow, cut the output resolution first, then the clip length; those two reduce work per frame and total frames respectively.
How much memory does a job need?
ClapClip's GPU enhancement page suggests 4 GB of VRAM or more for 1080p work, and 8 GB or more when the output is 4K or a batch is running. That is practical guidance from the product side, not a certified minimum, because memory use varies with the model, the frame size and what else the card is doing. Integrated graphics work but are significantly slower than a discrete card.
If a job fails partway through, the first things to try are a lower output resolution and a shorter clip, in that order. Both reduce peak memory before you spend time on anything else, and you should also keep free disk space for the original, the working files and the export.
Rule: if a job fails partway, lower the output resolution before you shorten the clip; memory per frame is the usual wall.
Which pass fixes which fault?
Match the pass to what is wrong. Upscaling a noisy clip, or sharpening a low-resolution one, mostly enlarges the problem.
| What is wrong | First thing to evaluate | What to watch for |
|---|---|---|
| Low resolution | A modest increase in output size | Crisp but invented texture |
| Low-light grain | Denoising | Faces turning waxy |
| Mild softness | Gentle sharpening | Halos, exaggerated compression blocks |
| Old tape or film | A short restoration test | Amplified tape artifacts |
Grain removal has its own page on video noise, and focus and motion problems are covered in the blurry video guide. Neither is fixed by upscaling, which is why the 4K guide puts cleanup before resolution.
Rule: if the fault is grain, denoise before you upscale; if it is softness, sharpen before you upscale; if it is resolution, upscale only after both.
Which settings exist in the app?
The app has a small set of controls, and each one changes a different part of the job. The table below lists what each one does.
| Setting | Options | What it changes |
|---|---|---|
| Input | MP4, MOV, AVI, MKV, WEBM | Which containers the app can open |
| Mode | General, Anime, Portrait | How the model treats the frame |
| Output | The resolution you choose | Frame size of the export |
| Audio | Original track kept | The audio is not processed |
| Batch | One settings pass, one output folder | Repeats the same job across files |
There is no frame-rate conversion in the enhancement path, and no stabilization. If a clip needs either, do it in an editor.
Rule: if you need a different frame rate or a steadier frame, fix that before enhancement; the app will not do it for you.
What do the common failures mean?
| What you see | Cause | What to do |
|---|---|---|
| "Choose a valid JPG, PNG or WebP image." | A clip was dropped on the browser tool | Open the video in the Windows app |
| "Enhancement failed." | The job used up its retries, usually on a long or large file | Split the file, or lower the output size |
| The app is much slower than expected | Another GPU-heavy app is sharing the card, or the laptop is on a power-saving profile | Close what you are not using and plug in |
| The export is bigger and softer than expected | Output resolution was raised but the bitrate was left at the old setting | Raise the bitrate to match the new frame size |
Rule: if the app is slow only on battery, fix the power profile before you blame the model.
How do you test a clip before committing?
- Pick a sample that can fail. Ten to thirty seconds, including motion, a face, fine texture and a dark scene if the source has them. A static bright shot hides every weakness.
- Write down the test conditions: input dimensions, duration, mode, output dimensions and elapsed time.
- Process the sample once. Compare it with the original at the same displayed size, then at 100%.
- Play it at normal speed. Check for flicker, halos, hair or foliage that crawls, and audio sync.
- Estimate the full clip from the sample. Multiply by the ratio of full duration to sample duration, and expect the real figure to be somewhat higher if the full clip has harder scenes.
- Process the rest only after the sample passes. For similar files, batch enhancement applies the same settings and writes to one folder.
- Separate mixed sources. Noisy night footage and clean animation are not the same job, and batch mode does not vary settings per file.
Rule: if the sample has no motion, no face and no dark scene, it is not a test; pick a section that can actually fail.
What's the bottom line?
The local path trades a queued cloud service for your own hardware: no upload, no waiting behind other people, and a failure mode that is yours to manage through memory, thermals and disk. That is a good trade for a folder of private clips and a poor one if you only ever enhance a single screenshot. Keep the original in both cases, because an enhanced video is a new file with details the model supplied, not a better copy of the archival master, and download ClapClip for Windows only after the thirty-second test passes.