AI video restoration for film, camcorder and archive footage

If the material comes from film, Video8, MiniDV, a digital archive or another physical carrier, the capture step defines the ceiling for everything that follows. Use a well-maintained playback device appropriate to the format and, where possible, time-base correction for unstable analog timing. Capture once at a high-quality, edit-friendly setting. Replaying fragile media repeatedly to try different AI settings creates physical risk without adding information to an already weak digital copy. Tape-specific handling — deinterlacing a 480i or 576i capture, keeping the 4:3 frame, and what a cassette costs per minute in credits — is covered in the VHS to 4K guide.

Label the untouched capture with the carrier, date, device, and capture settings. Store at least two verified copies in different locations before restoration. Do not overwrite it with a cropped, denoised, deinterlaced, or upscaled file. A messaging-app transfer or low-bitrate DVD export is not an archival master merely because it is convenient. If several generations exist, compare them; an older-looking copy may still contain more usable information than a later compressed export.

Many old-video defects are not resolution problems. Analog captures may show head switching noise, horizontal timing bends, dropouts, chroma misalignment, or interlaced combing. Film transfers may have incorrect speed, duplicate frames, gate weave, dust, or exposure flicker. Correct playback speed and frame cadence first. Deinterlace interlaced material with a method that respects field order before feeding progressive frames to an enhancement model. Repair missing or broken frames only with a documented method.

Crop unstable borders conservatively and keep a version that shows the full captured frame. Avoid removing all grain or tape texture. Those patterns can carry visual continuity, and excessive denoising can turn faces into flat shapes before the model sees them. Do not sharpen strongly. Halos around hair, shoulders, and titles may be amplified into false contours. Export a short prepared test from the restored timeline without an additional low-bitrate delivery encode.

A useful restoration test contains the information you cannot afford to change: a person turning toward camera, a handwritten sign, a uniform, a vehicle plate, fabric patterns, or a familiar room. It should also contain typical defects and movement. Upload five to ten seconds, choose a modest output such as 720p or 1080p, and note the displayed credits. The purpose is to find a reason to reject the method before it is applied to the whole archive.

Compare the source and result frame-aligned. Ask someone familiar with the people or place to review without presenting the AI output as authoritative. Look for changing eye shape, teeth, jewelry, lettering, and background objects. Watch grain and texture during motion. If the model invents a plausible but unsupported feature, do not hide that problem with color, music, or a smaller preview. Try a less aggressive target or preserve the scene without AI enhancement.

An AI-enhanced version is a derivative. Keep it beside, not in place of, the untouched capture and the minimally corrected restoration master. Name files so the distinction is obvious. A family compilation, documentary edit, or social post may reasonably use the enhanced copy, but the source of record should remain available. If the result is published, a simple disclosure that AI-assisted restoration was used prevents viewers from mistaking estimated detail for original evidence.

Retain audio separately when it has historical value. The current upscaler request does not expose audio controls; remuxing, restoration, wow and flutter correction, noise removal, synchronization, and missing speech repair are separate decisions. Confirm that the final delivery plays correctly and that synchronization is unchanged. Store checksums or use a backup system that periodically verifies files; an untested drive on a shelf is not a preservation strategy.

Credit cost grows with billed duration, resolution, and selected model. The Standard workflow starts at 45 credits for a five-second 720p job, while 2K and 4K cost more. Old sources do not necessarily benefit from the largest target. A stable 1080p derivative can preserve the character of a tape or film transfer better than an oversharpened 4K output and is easier to review, store, and deliver.

Divide long archives into logical reels or scenes and create a processing log. Retry only failed or rejected segments instead of resubmitting the whole recording. Allow for queue and provider variation; an estimate is not a preservation deadline. Download and verify each approved result before the hosted link expires or the project moves on. The goal is a reproducible chain with controlled changes, not the fastest route to a dramatic before-and-after thumbnail.

This tool accepts one source of at most 10 minutes, and that cap is real: a two-hour family film has to be submitted as 12 separate parts. What that costs is continuity, not resolution. Parts processed with different models, resolutions, or settings look like different films once they are joined, and a cut made in the middle of a shot is the one an audience notices. Split at a scene or shot boundary instead, so every cut you create is a cut the film was already going to have.

Keep one model and one resolution for every part of the same film, and write down which ones you used. Keep the untouched capture and every processed part on disk until you have watched the whole reassembled film: a hosted result link is a live window, not an archive. The current upscaler request does not expose audio controls, so re-attach the original audio track when you rejoin the parts in your editor, in the same order you split them. If you later change model or resolution, redo every part rather than only the ones you dislike, and re-run the parts that failed instead of resubmitting the whole recording.

For one whole family film there are three honest routes, and they differ in what you pay and who does the work. Our hosted route prices in credits per hour of footage; running the open-weights model on your own GPU removes the credit cost and moves the waiting and the troubleshooting to your desk; a professional transfer house also handles the physical tape and an operator's judgment. The table below compares them on cost, queue time, control, and what happens when a run fails, using only figures published on this site.