AI Restoration at Archive Scale: GPU Throughput Planning

Overview

Indian broadcasters and studios hold enormous back catalogues that were never mastered above standard definition, and streaming platforms want them at 4K with clean audio. AI upscaling, denoising, scratch removal and frame interpolation now do this at quality that passes commercial review, which turns restoration from an artisanal per-title project into an industrial throughput problem. The unit of planning is the frame: a two-hour feature at 24 fps is about 172,800 frames, and a thousand-title library runs to hundreds of millions. This article sets out how to size for that.

From catalogue to GPU hours

The estimate has four inputs: number of titles, average runtime, frame rate, and processing rate per GPU for your chosen model chain. Multiply the first three for total frames, divide by the fourth for GPU-hours, then divide by available GPUs for wall-clock. The discipline is to measure the processing rate on your actual content rather than assuming a published figure, because grain, resolution and damage level change it substantially.

Two adjustments follow. Restoration is rarely one pass: a typical chain runs denoise, then scratch and dust removal, then upscale, sometimes then interpolation, and each pass reads and writes the whole frame sequence. And a meaningful fraction of titles fail first-pass review and need reprocessing with different parameters, so plan a rework allowance of 15 to 30 percent rather than assuming a clean run.

Where the cost sits by stage

The highest-value discipline is deciding which stages a given title actually needs. Applying the full chain uniformly across a catalogue wastes most of the compute; a triage pass that classifies titles by source condition and target tier lets you route cheaply through what does not need heavy treatment. That triage is itself a light model and pays for itself quickly on a large library.

Storage is usually the real bottleneck

Frame-sequence restoration reads and writes uncompressed or lightly compressed images at every stage. A single 4K frame in a 16-bit intermediate format is tens of megabytes; a feature’s worth is measured in terabytes; and with multiple stages you hold several generations simultaneously. Studios that size GPUs carefully and storage casually discover their accelerators idle waiting on I/O.

Three mitigations. Keep intermediates on fast NVMe local to the processing node and only write finals to shared storage. Chain stages in memory where the toolchain allows, avoiding a round trip per pass. And enforce a cleanup policy that deletes intermediates on delivery, because a restoration programme that retains every generation will consume capacity faster than any budget accommodates — particularly at current flash pricing, discussed in the 2026 memory and NAND squeeze.

Quality control does not scale by itself

The uncomfortable truth of large restoration programmes is that the GPU stage can be scaled by adding hardware and the review stage cannot. Someone must watch the output, and AI restoration failures are specifically the kind that automated metrics miss: a face subtly reshaped by upscaling, text made legible but wrong, a grain structure replaced by plastic smoothness that only reads as wrong in motion.

Practical approaches reduce but do not remove the burden. Automated checks catch gross failures — frame count mismatches, colour shifts, dropped frames — before human review. Sampling protocols, reviewing high-risk sections such as faces, titles and rapid motion rather than everything, cut review time substantially. And per-title parameter presets learned from earlier reviews reduce the rework rate over a programme’s life. Budget review capacity explicitly in the plan; it is the constraint on delivery date.

Own or rent the capacity

Restoration is batch work with no latency requirement, which makes it unusually well suited to whichever capacity is cheapest per GPU-hour. For a one-off campaign — a rights holder digitising a library once — rented capacity avoids buying hardware that idles afterwards, and India’s subsidised national compute at rates near a dollar per GPU-hour is competitive for this profile.

For a sustained programme, where a broadcaster restores continuously as part of operations, owned hardware wins on unit cost over a depreciation cycle, and it keeps masters inside the facility. That second point matters: archive masters are valuable rights assets, and many licensing agreements restrict where material may be processed. Confirm the contractual position before moving a library to any external service. The TCO comparison method is in on-prem vs cloud GPU TCO.

Planning an Indian restoration programme

Sequence it commercially. Restore the titles with demonstrable streaming demand first, since they fund the rest, and use that first cohort to calibrate processing rates, rework rates and review effort before committing to a full-catalogue schedule. Establish the target tier per title early — not everything needs 4K, and a good 2K master from a damaged source is often the better outcome.

Two further notes. Source condition dominates outcome: a well-preserved negative scanned properly produces a far better result than any amount of AI applied to a degraded telecine, so scanning quality deserves investment before processing capacity. And retain the unrestored scan permanently alongside the master, because restoration techniques improve and a future pass will want the original. The studio infrastructure context is in GPU for rendering and generative AI in India.

Frequently asked questions