Why archival video restoration comes up against manual labor
Archival video restoration almost always comes up against the same thing: the main time sink is not digitizing the archive but manual frame retouching. Automation confidently handles clean shots and falls apart where film is scratched, overexposed, or spliced from pieces.
An operator manually traces every scratch and repaints damaged areas — minutes of finished material turn into weeks of department work.
Film defects do not follow a single rule. Grain in dark scenes reads as noise, while in bright scenes it reads as texture, and the same filter setting either eats detail or leaves dirt.
Scratches "float" with the camera in motion, and stand as a dot in static shots; dust and hairline strokes are present in one frame and disappear in the next. It is physically impossible to tune thresholds for the entire body of material at once.
A separate headache is splices and shrinkage. Over decades, film shrinks unevenly, the perforation pitch drifts, the frame jitters and "breathes" vertically, and jumps and duplicates appear at the joints.
Frame-by-frame stabilization using reference points partially helps, but at transitions it requires manual marking — otherwise interpolation from neighboring frames drags artifacts along and blurs the background.
The cost of the process is not only deadlines. Manual retouching is subjective: two operators will produce different results on the same material, and a "washed-out" face in a frame cannot be brought back.
The larger the archive, the more people must be kept in the department, and restoration stops scaling. Below we look at how this is addressed and what routine work can actually be handed over to the model.
What the system delivers: processing speed, defect rate, downtime
Before automation, the department's output was measured by eye: how many films were reviewed per shift and how many were cleaned in time.
In the system, every batch follows the same route, so five numbers emerge at the output that show the real restoration performance — without subjective assessments and a "feeling of progress."
First — archive processing speed. Material moves in parallel queues, so an hour of source footage no longer stretches into a shift: the department's throughput is measured in days, not in individual films, and the release of the collection can be planned by calendar.
Second — the share of repeat passes. If the result is not accepted, the frame goes back for processing; in a properly configured system, such returns are only a few percent, and the rest is resolved on the first pass.
Every extra pass means more department time again, so reducing it has an effect across the entire queue.
Third — the share of frames requiring manual finishing. The operator gets involved only where automation has honestly flagged a questionable area: severe damage, missing frames, track desynchronization. The flow of manual work collapses to isolated spots, and it stops being a daily routine.
Fourth — consistency of results between batches. Old and new acquisitions are processed by the same rules, so adjacent releases do not diverge in grain, contrast, and color — editing comes together without additional color correction.
Fifth — freed-up time for archivists: it goes to selection, description, and verification against source documents, that is, to what truly requires a human.
Film and digital master restoration: comparison of modes
The choice of mode is dictated not by the volume of material but by the nature of the source: a film scan with unstable grain and splices behaves differently from a smooth digital master from an editing system.
Frame-by-frame processing maintains temporal coherence between neighboring frames and delivers maximum sharpness recovery, but requires intermediate sequences and space for them.
Stream processing reads material sequentially — it is predictable in memory and faster over long segments where defects are uniform.
In practice, a mixed scheme more often wins: the rough pass runs in stream mode, problematic shots are handled frame by frame, and finishing and assembly are brought together in DaVinci Resolve, where grain, color, and sound are combined. Below is how the modes differ in applicability and what limits each one.
| Option | When it fits | Limitations |
|---|---|---|
| Full frame-by-frame pass | Archive with unstable grain, flaking emulsion, frequent splices; maximum sharpness recovery required | Slow per minute of material, requires intermediate sequences and disk headroom |
| Stream processing | Smooth digital master, uniform noise, large volumes | Holds temporal coherence less well: artifacts between frames are possible during fast motion |
| Selective finishing | Material is generally usable, and the list of problematic shots is already marked up | Markup is done manually; a local defect outside the list remains unnoticed |
| Mixed scheme | Large projects with heterogeneous source material | Requires stitching results at shot boundaries and monitoring junctions between modes |
The common limitation is reversibility. We keep a map of applied operations and intermediate versions, so rolling back an individual shot does not entail reprocessing the entire archive.
How we run a project: from intake of materials to archive delivery
A restoration project rarely fits into a single pass: the collection arrives in batches, film and digital copies behave differently, and the requirements for the result depend on the archive's purpose. Therefore, each stage ends with an artifact that can be checked before moving to the next.
- Inventory and intake of materials. We reconcile the inventory list with the actual contents, record media types, formats, film condition, splices, and lost frames. The output is a collection map and a list of problematic units.
- Test fragment. We take a few minutes from the most challenging part in terms of condition and run the full cycle: it is more reliable to catch an error here than in the middle of multi-batch processing.
- Calibration of processing for the specific collection. We select noise reduction, stabilization, interpolation, and color correction parameters so that "plastic" faces do not appear and the grain of the shooting era is not erased. We assign the set of settings to the collection.
- Batch processing. We process material in parts with logging: the log shows which setting was applied to which batch, and any part can be reassembled from scratch.
- Quality control. We compare critical scenes frame by frame with the original and selectively check the rest of the material, calculate the share of frames with residual defects, and correct them selectively.
- Archive handover and training. We provide a lossless master copy, versions for screening and publication, metadata, and storage regulations; we show your team how to repeat the settings on new acquisitions.
What remains with the client: schemes, rules, regulations, access permissions
A system cannot be considered handed over if, after launch, only our team understands it.
Therefore, along with the working service, the client receives a set of settings and documents that let the archive continue independently: add new materials, adjust processing parameters, and grant and revoke access without contacting us.
- Processing scheme — the end-to-end route of material from media intake to delivery of the finished copy: where automatic analysis runs, where the operator gets involved, and at which points intermediate versions are created.
- Processing profiles — ready-made parameter sets for material types: black-and-white film, color negative, magnetic tape, damaged media. A new batch is created by selecting a profile, not by setting it up from scratch.
- Access rules and roles — permissions mapped to real archive positions: who starts restoration, who approves the result, who sees only source files without editing rights.
- Change log — records every settings change, manual correction, and processing status. It shows who intervened in the result and when.
- Storage regulations — how long source files, intermediate versions, and master copies live, and what goes into long-term storage and when.
- Instructions for archivists — step-by-step guides in Russian for daily operations: intake, batch launch, error review, export.
- Disaster recovery plan — procedures and processing priorities if part of the equipment is unavailable.
Describe your task — we will come back with an estimate and a plan for what from this list your archive specifically needs.
Case study: restoring newsreels without losing source materials
In the collection of the Archive of Film Documents — 240 newsreel issues on film: shrinkage, splices with varying brightness, vertical scratches, and frame jitter.
Each issue underwent manual finishing: an operator cleaned defects frame by frame, up to 30% of frames required correction, and a repeat pass after assembly was needed for almost every issue.
The film also went through the equipment several times — and this is the main risk, because mechanical intervention is irreversible.
What changed
Scanning was moved into a separate stage, and a master copy was immediately fixed at maximum resolution, which is not overwritten afterward.
Image work is handled by models: stabilization using reference frames, brightness and grain alignment along splice boundaries, restoration of damaged areas using "defective — reference fragment" pairs.
All edits live in separate layers and are logged, so any version can be reassembled without going back to the film.
What the result was
The share of frames going to manual finishing dropped from 30% to 4%, and repeat passes — from 9 out of 10 issues to 2 out of 10. Time per issue was reduced by about three times.
The integrity of source materials is confirmed by checksums before and after processing, and the archive collection itself received two results: versions for screening and immutable master copies for long-term storage.
How does frame-by-frame processing differ from stream processing?
Stream processing maintains a window of neighboring frames: the algorithm evaluates noise, jitter, and motion across the scene as a whole and carries the decision over to the current frame.
Frame-by-frame processing works with a single image independently of its neighbors — restoration is based on the statistics of that frame itself. The difference is fundamental: the former relies on time, the latter on the frame's space.
Hence the difference in artifacts. Stream mode suppresses grain and low-frequency noise smoothly, but the error is smeared across the window: trails appear during fast motion, and at splices the algorithm "pulls" a frame toward the previous one.
Frame-by-frame mode produces a sharp result on static shots and graphics, but loses temporal coherence — brightness and grain begin to "boil," and flicker is visible on uniform backgrounds.
The second axis is controllability and reproducibility. Frame-by-frame mode parallelizes across nodes almost linearly and, with a fixed model, produces a repeatable result: an episode can be reprocessed without recalculating the scene.
Stream mode requires alignment across the window and is sensitive to frame order — it handles gaps and duplicates from digitization poorly.
Therefore, the choice of mode is set by the source type: film scans with grain and stable geometry benefit from stream processing, while television recordings with interlacing and cassette digitizations with unstable geometry require a frame-by-frame pass with separate temporal stabilization.
In practice, modes are combined: stream processing removes noise, jitter, and flicker; restoration and upscaling run frame by frame; and the final pass aligns brightness across the scene.
The selection criteria are the nature of the defect, the amount of motion, the smoothness of the digitization, and the requirement for repeatability of the result.
Technical doubts worth resolving before the start
Before starting, what is tested is not the model but its behavior on your material. Below are five questions we address on the test fragment.
The film is severely damaged — what can actually be restored?
Some defects are removed by frame-by-frame registration and stabilization, others by masks. Completely lost frames are not silently redrawn: they are marked, and their share is visible in the report.
The archive is already digitized with errors — work with it or rescan?
We work with what exists: scanning defects and film defects are corrected by different methods. If the distortion is irreversible — channel loss, frame cropping — we will say so directly after the test fragment.
How does the system distinguish grain from a defect?
Grain is stochastic, while a defect stays in one place between frames — the separation is based on this. The threshold is adjusted manually: what we remove completely and what we leave as film texture.
How are batches joined at boundaries?
Material is cut with overlap, parameters are calculated against a common reference — there is no jump in density or color at the joint. We check batch boundaries separately by comparing frames before and after the transition.
What happens to the source materials?
The master copy is not overwritten: edits go into a separate layer, and masks are stored as decisions. The result is reproducible — it can be reassembled without rescanning.
Let's discuss your task: send material samples
So that the scope estimate is based on the material rather than a retelling, send the samples themselves: one or two fragments for each type of media and defect in the collection.
Take pieces with the most problematic areas — graininess, scratches, dropped frames, jitter, audio desynchronization, fading. 20–60 seconds per fragment is enough, preferably in the original form, without intermediate re-compression.
What to include in the package along with the samples:
- one sample per media type: film, magnetic tape, digital archive, digitization;
- fragments in the original container — otherwise the defect cannot be distinguished from a conversion artifact;
- collection inventory: number of storage units, hours or footage, availability of a catalog and its format;
- a list of defects critical specifically to you — we use it to set priorities;
- delivery requirements: broadcasting, archival storage, or editing, and the master copy format;
- access conditions for restricted materials.
Based on this data, we calculate the scope estimate — by defect and media types, not by total running time — and assemble a work plan:
- how many storage units and processing hours, and where the bottlenecks are;
- the sequence of stages and the order of collection processing;
- pilot section: one unit or 10–15 minutes, where we verify quality before scaling;
- acceptance criteria: by what signs we consider a frame restored and the sound mixed.
This is the draft proposal for your collection: you can see what we do, in what order, and what we use to verify quality.
