Video denoise without erasing real texture

Film grain is an optical effect rather than dirt on the file. Wikipedia's film grain article describes it as the random optical texture of processed photographic film, developing from small particles of metallic silver or dye clouds produced from silver halide grains that received enough photons, and notes that its magnitude depends on both the film stock and the definition at which the picture is viewed, becoming objectionably noticeable in an over-enlarged film photograph. That last clause matters here: enlarging a grainy frame makes the grain look worse even though nothing was added to the file.

Digital noise has different causes. Sensor noise rises with ISO and with heat, and appears as fine luminance speckle plus blotchy colour; it is usually strongest in shadows. Compression noise is a third thing again: Wikipedia's compression artifact article explains that block-based transform coding quantises each block separately, which produces visible blocking and block boundaries that are most obvious in flat areas, plus ringing around edges that turns into mosquito noise in motion because the artifacts persist across decompressed frames and travel with the optic flow.

The practical reason to separate them is that they respond differently. Grain sits across the whole frame at roughly one scale; sensor speckle is concentrated in dark areas; blocking follows the encoder's block grid; mosquito noise hugs edges. A reconstruction pass that models the scene can suppress a good deal of speckle and blocking, but the same pass that removes grain from a 1970s interview will also remove the grain that told you it was shot on film.

A hosted job here carries four things to the provider: the video file or a public URL, the chosen model, the target resolution tier and the billed duration. That is the whole request. There is no denoise-strength value, no noise-profile preset, no chroma/luma split, no audio option and no field-order or frame-rate setting. If you need a precisely tuned spatial or temporal denoiser with a strength parameter, that is a different class of tool and this page will not pretend otherwise.

What the models are is video restoration and enlargement pipelines: heavy reconstruction models that re-synthesise detail and, as part of that reconstruction, tend to reduce speckle, fine blocking and mosquito noise. SeedVR2 is the diffusion-based model and the heaviest of the five, FlashVSR is a faster reconstruction model, and Standard is the cheapest general-purpose option. None of them exposes a number that says how much noise to remove, so the control you have is model choice, output tier, and which clip you submit.

Two consequences follow. First, dose is chosen by the model rather than by you, so the honest workflow is trial and comparison, not dialling in a value. Second, anything the service does not accept has to be handled before upload: if a capture is interlaced, deinterlace it first; if the soundtrack matters, keep the original audio and re-attach it after the upscale, because the request carries no audio settings at all.

Pick five to ten seconds from the worst part of the recording: the darkest shot, the fastest movement, any on-screen text, and the specific texture you are worried about losing — fabric, foliage, skin, brick. Upload it, choose the model and the output tier, and read the credit quote shown before submission. Because the minimum billed duration is five seconds, a test costs a small fraction of the full clip, and it is the only way to find out what a model does to your particular noise.

Write your pass and fail rules before you look at the result, so the comparison stays honest. Reject a setting when skin turns waxy, when hair and foliage collapse into flat smears, when straight edges shimmer or crawl between frames, when fine texture becomes a repeating pattern, or when lettering changes shape. Accept it when grain reduces without the picture going plastic and when detail holds steady across a camera move rather than flickering.

Only then process the rest, in segments. One source is capped at 10 minutes, so a long recording becomes a series of submissions; keep a note of the model, tier and credits per segment, because the cheapest mistake to avoid is discovering a bad setting on the fortieth minute of a fifty-minute file. Keep the original untouched so you can always re-run a scene at a different model.

Every figure on this page comes from the published credit table rather than from an estimate. 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 the SeedVR2, Pro and Ultimate models all bill 49 credits a second at 1080p, with heavier rates above that tier.

Convert that into the plan you would actually buy. Creator is $238.80 for 28,000 credits, which covers about 10 hours of 1080p Standard work at 32,400 credits an hour, and 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 a sample, the smallest public general credit pack starts at $29.99 for 1,500 credits.

The useful comparison is between models, not between resolutions, for noise work. Dropping from 4K to 1080p at the Standard model cuts the bill to roughly a quarter while keeping the same reconstruction, and 720p bills identically to 1080p, so there is rarely a reason to choose it. Moving from Standard 1080p to SeedVR2 1080p multiplies the per-second cost by more than five, so reserve the heaviest model for the scenes where the cheap one visibly fails.

Review at full size on the display you will actually use, and compare the same frame in both files — the original on one half of the screen and the result on the other, cropped to the same 100 percent view. A small preview hides exactly the damage that matters: smeared texture, reshaped lettering and stabilised-looking edges that are being invented frame by frame.

Watch motion, not just stills. Compression artifacts and denoising both behave differently across time: blocking that is invisible in a still can crawl, and a reconstruction that looks crisp in one frame can flicker on hair or foliage in the next. If your player can step frame by frame, do so around a fast pan and around any face, then watch the same scene at normal speed.

Keep the notes short and comparable: source, scene timestamp, model, tier, credits, verdict and the single reason for that verdict. Notes like that are what let you change one variable at a time instead of re-testing blindly, and they are also what stops a second pass from undoing the grain you decided to keep.

Grain carries information. On film it is a record of the stock and the exposure, and in a documentary or an archive transfer it is part of what the footage is. A model that smooths it away produces a cleaner frame and a less truthful one, so decide before you start whether the grain is a defect to remove or a property to preserve, and write that decision down with the test notes.

There are also cases where the honest answer is to change the source instead of the settings. If the noise arrived in a re-encode — a messenger copy, a screen recording of a player, a download from a platform that compressed it again — then recovering the original file removes more noise than any model will. If the noise came from a very dark exposure, re-shooting or re-capturing with more light beats reconstruction every time.

And if the soundtrack needs repair, this is not the tool for it. Audio noise reduction is a separate job with separate tools; this pipeline takes the picture, and the audio has to be preserved and re-attached afterwards if you need it. Being clear about that boundary saves credits and avoids a deliverable that nobody can use.