A YouTube video upscaler for the file you upload

Softness in an upload has one primary cause: the file that was sent. If the camera or the recording app captured 480p or 720p, that is the detail the file contains, and no playback setting, viewer preference or platform option adds the missing pixels back. This is arithmetic rather than opinion: Wikipedia's 480p article gives standard-definition and 480p pictures at 640 × 480, 720 × 480 or 854 × 480, and a 720p picture at about 0.92 megapixels, against about 2.07 megapixels at 1080p.

Softness can also arrive later, as compression. A clip that was sent through a messaging app, re-downloaded from another service or exported at a low bitrate carries blocking and mosquito noise from that encode. Wikipedia's compression artifact article describes how block-based transform coding leaves discontinuities at block boundaries that are most apparent in flat areas, and how video artifacts persist across decompressed frames and travel with the optical flow — which is why such material can look almost clean in a still and crawl in playback.

So there are two questions to answer before spending anything. Is the file the best export you can produce, at the highest resolution you still hold? And is the softness baked in by a capture or by a later re-encode? If the answer to the first is no, export again first: replacing a 480p export with the original 1080p one beats any model here. If you genuinely hold only a soft file, upscaling is the available lever, and the rest of this page is about doing it without overpaying.

Choose the output tier from what the source can support and what the destination expects, not from the largest number available. At the Standard model, 720p and 1080p both bill 9 credits a second, so 1080p is the free choice; 2K bills 17 and 4K bills 33, about 3.7 times the 1080p rate. A 480p source holds about 0.31 megapixels, so a 4K output is roughly twenty-four times the recorded pixel count, while 1080p is roughly six times.

That does not make 4K pointless. It makes it a compatibility decision: a 4K delivery specification, a project that already lives on a 4K timeline, or a destination that handles 4K frames better than 1080p. What it is not is a way to recover detail, because the same reconstruction happens at 1080p for a third of the rate. If you want more from a difficult clip, the cheaper lever is usually a heavier model at 1080p rather than the cheapest model at 4K.

One practical rule for uploads: process at the tier you intend to deliver and stop there. Re-upscaling an already-upscaled file multiplies cost and re-encodes the picture again, and the second pass is working from reconstruction rather than from the original. Keep the untouched source, and keep the first completed output as the deliverable.

Pick five to ten seconds that represent the hardest part of the upload: an opening shot in low light, a fast pan, small text on screen, a face in motion. Upload it, choose the model and the output tier, and read the credit quote before submitting. Because billing starts at a five-second minimum, this test costs a handful of credits and settles both questions at once: which model, and which tier.

Fix the pass and fail rules before looking. Reject a setting when small text changes shape or becomes unreadable, when a face stops being the same face, when flat backgrounds crawl, when edges acquire haloes, or when the result looks sharpened rather than resolved. Accept it when the clip plays smoothly at the tier you will upload, when edges are stable through motion, and when a 100 percent crop shows plausible texture rather than a repeating pattern.

Then process the rest in 10-minute segments, keeping notes per segment: source export, scene, model, tier, credits and verdict. Only failed segments need a re-run, which is where the heavier models pay for themselves, and keeping the untouched export means the whole job can be redone later if a better master appears.

All figures come from the published credit table. At the Standard model 1080p bills 9 credits a second, which is 540 credits a minute and 32,400 an hour of source footage; 2K bills 17 a second and 4K bills 33. FlashVSR bills 30 credits a second at 1080p, and SeedVR2, Pro and Ultimate bill 49 credits a second at 1080p.

Against the yearly plans, Creator is $238.80 for 28,000 credits, which is about 10 hours of 1080p Standard work; Starter is $118.80 for 9,600 credits and Studio is $598.80 for 80,000 credits. If you only want to test whether a particular clip is worth processing, the smallest public general pack is $29.99 for 1,500 credits.

Read those figures per minute of finished video rather than per file, because a channel's output is a stream of short clips. Ten minutes of finished upload at 1080p Standard is 5,400 credits; the same at Standard 4K is 19,800. That is the arithmetic that makes a 1080p default sensible for routine uploads, with 4K reserved for the pieces where a delivery spec or a sponsor brief actually asks for it.

The pipeline changes the picture and nothing else, so everything around the picture is your job. The request carries the file or URL, the model, the target resolution tier and the billed duration: there is no bitrate target, no codec choice, no frame-rate setting, no aspect-ratio option and no audio handling. If the clip needs loudness normalisation, a caption track or a specific export preset, that is the editor's work before and after the upscale.

Remux the audio before publishing. Because the request carries no audio settings, the original soundtrack has to be re-attached to the upscaled picture; doing that at export time keeps the deliverable a single file and avoids publishing a silent clip. Also settle the frame shape deliberately: the tool does not crop, pad or letterbox, so a 4:3 source stays 4:3 unless you decided otherwise in the editor.

Finally, do not re-encode the output harder than you need to. If the delivery is a single platform upload, a normal export from your editor is fine; what you are avoiding is a chain of re-exports that each subtract quality from a picture that has already been reconstructed once. Keep the first completed output as the master and export delivery copies from it.

Do two passes. First, compare the same frame at 100 percent against the untouched source: this is where invented texture, reshaped letters and haloed edges show up, and where you can tell whether a model has resolved a detail or manufactured it. Then watch the whole segment at normal size and speed, because an upload is consumed as motion and a still-frame win that flickers in playback is not a win.

Check the three faults that matter most at upload scale: faces, on-screen text and flat backgrounds. Faces because a viewer notices identity immediately; text because small lettering is where reconstruction most often changes shape; flat backgrounds because gradients in skies and dark interiors either band, crawl or go plastic depending on the model. If any of the three changes between adjacent frames, change the setting rather than the scene.

Write the result down in one line per segment — source, timestamp, model, tier, credits, verdict, one reason — and keep the completed output beside the untouched source. When the next clip arrives, that record turns a forty-minute experiment into a two-minute decision.