Depixelate video: what a model can honestly rebuild

Pixelation in the strict sense is a display effect: Wikipedia's pixelation article defines it as showing a bitmap or part of a bitmap so large that the individual single-coloured display elements become visible, and adds that depixelization attempts to reconstruct the appearance of the original image. The word attempts is doing real work there, because the pixels that were discarded at some earlier encode are not stored anywhere in the file you have now.

The blockiness people usually mean is a compression artifact. Wikipedia's compression artifact article explains that block-based transform coding quantises each block separately; below a certain bitrate the higher-frequency coefficients are quantised to zero, so each block reconstructs as a blurry low-resolution patch, and neighbouring blocks differ enough to leave visible discontinuities at block boundaries that show up most clearly in flat areas.

Motion adds a second artifact on top. The same article describes mosquito noise — a shimmering blur of dots around edges, caused by block-based DCT compression — and notes that in motion-predicted video the artifacts persist across several generations of decompressed frames and move with the optical flow. That is why a still frame can look almost clean while the same shot crawls in playback, and why a re-encoded copy of a re-encoded copy is the hardest and least rewarding case.

Low native resolution is the first cause and the one a model can help with most directly: a 480p or 360p source simply has few real pixels, and enlarging it makes that visible. Low-bitrate blocking is the second: the frame size may be perfectly adequate while the encoder had too few bits, which is common in long recordings, in screen recordings of a video player and in anything re-uploaded to a chat app. Deliberate mosaic blur, used to obscure a face or a plate, is the third, and it is not a compression artifact at all. Second-generation compression is the fourth, and it is the accumulation of the previous two.

Before spending credits, decide which one you are looking at. If a face has been deliberately mosaiced, a reconstruction model will invent a plausible face where the original was never recorded, and the page you are reading will not pretend that is recovery. If the file is a chat-app copy of a platform download, finding the original upload is worth more than any setting here. If the flat grey wall is banded with blocks while the foreground looks reasonable, that is classic low-bitrate blocking and it is a fair target.

One practical test separates resolution loss from bitrate loss: watch the same scene on a large screen and on a small one. Resolution loss looks soft at both sizes, because there is genuinely less detail; blocking is most visible at large sizes and in flat areas, and it crawls when the camera moves. Write down which one dominates before you upload anything.

A hosted job here sends the video file or a public URL, the model, the target resolution tier and the billed duration, and nothing else. There is no deblocking strength, no sharpening amount, no noise preset, no field-order option and no audio option. So the controls you actually have are which model you pick, which output tier you pick, and which five seconds you submit first.

The models are reconstruction and enlargement pipelines, not arithmetic deblockers. A deblocking filter smooths block boundaries by a rule; a reconstruction model re-synthesises the frame from a learned prior, which is why it can make a blocky wall look continuous and also why it can invent texture that was not in the original. Wikipedia's compression artifact article is candid about this class of fix, noting that no post-processing technique has been shown to improve image quality in all cases, which is why this page asks you to test rather than trust.

Two handover rules follow. Deinterlace before upload if your source is interlaced, because the request exposes no field-order control and an enhancement pass can otherwise sharpen comb teeth. Keep the original audio if the deliverable needs it, and re-attach it after the upscale, because no audio settings are sent.

Choose five to ten seconds that contain the worst case: a flat wall with visible blocks, a dark gradient with banding steps, a fast pan across texture, and any face you care about. Upload it, choose the model and the output tier, and read the credit quote before you submit. Because the minimum billed duration is five seconds, the test costs a small fraction of the full file and it is the only honest way to learn what a model does to your particular compression history.

Decide your pass and fail rules before viewing. Reject a setting when lettering changes shape, when a face becomes a different face, when flat areas acquire a crawl or a repeating pattern, when sharp edges grow haloes, or when detail that was stable across frames starts breathing. Accept it when block boundaries stop being visible, when the wall reads as a wall, and when faces and text hold their identity across a camera move.

Then process the rest in segments. One submission is capped at 10 minutes, so a long recording becomes several jobs; record the model, tier and credits per segment so a bad setting is not discovered late. Keep the untouched original so any scene can be re-run at a heavier model, and remember that a lighter model at a higher resolution is not the same trade as a heavier model at the same resolution.

All figures come from the published credit table. At the Standard model, 720p and 1080p both bill 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 all 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 and proportionally less at the heavier models; Starter is $118.80 for 9,600 credits and Studio is $598.80 for 80,000 credits. If you only need to check one scene, the smallest public general credit pack starts at $29.99 for 1,500 credits.

Two habits keep the bill honest. First, test blocky material at 1080p rather than 4K: the block grid is far larger than a 1080p pixel, so the extra output resolution mostly adds cost. Second, spend the heavier model where blocking is worst rather than everywhere, because the per-second difference between Standard and the heaviest models is more than five times.

Compare the same frame at 100 percent, original beside result, on the display you will deliver on. Blocky material is unusually easy to flatter: a soft result can hide blocking simply by blurring it, and a sharp result can hide it by adding texture that belongs to the model rather than to the scene. The question to ask at every crop is not whether the blocking is gone but whether what replaced it is plausible.

Always watch motion. Block boundaries that are stable in a still can crawl once the camera moves, and reconstruction that looks crisp in one frame can boil on texture in the next. Step frame by frame through the fastest pan, then watch the same clip at normal speed; if a face or a sign changes shape between frames, the setting is wrong for that scene regardless of how good the still looks.

Keep a short record: source, timestamp, model, tier, credits, verdict, and one reason. Notes like that let you change a single variable at a time, and they also protect you from a second pass that removes the blocking you fixed along with the detail you wanted.

Mosaic blur applied on purpose is a different case from compression, and it is worth stating plainly: when a face or a licence plate has been replaced with large blocks, the underlying pixels are not in the file. A reconstruction model asked to enlarge those blocks will produce a plausible-looking patch that is a synthesis, not a recovery, and presenting that as the original would be false. That boundary is not a technicality; it is the difference between a clean-up tool and a fabrication tool.

The same caution applies to anything whose value depends on being exact: a document on screen, a serial number, a date stamp, a medical image, a legal exhibit. Upscaling can make such a frame easier to look at and easier to encode for delivery, but it cannot be used as evidence about what a low-resolution patch originally contained, because every reconstructed pixel is a model output rather than a measurement.

Where the content is ordinary and the blocking came from bitrate, the calculation changes: a wall, a sky, foliage or fabric can be reconstructed convincingly, and the main risk is over-smoothing rather than invention. That is why the workflow on this page starts with the worst five seconds: it is the fastest way to see which of those two regimes your footage is in.