What is the best archival film AI upscaling workflow to 4K in 2026?

Why Archival Film Demands a Different Upscaling Workflow Than Modern Footage

Upscaling archival film in 2026 is not the same operation as upscaling a smartphone video or a digital camera clip. Cellulose nitrate and acetate stocks carry physical artifacts that digital sensors never produce: gate hairs, emulsion scratches, vertical tramlines, mold blooms, dye fading, and density shifts that vary scene by scene. A model trained primarily on clean modern footage will often hallucinate textures inside those artifacts, turning a faint scratch into a dark spider-web structure across a face. That is why the archival pipeline has to separate physical restoration from AI reconstruction rather than running a single one-click model over the whole image.

Also worth reading: How Does a K Video Restoration Workflow Use AI Upscaling to Reach 4K? · How Should Archival Footage Quality Control Work Before AI Upscaling to 4K? · What Is the Best Kling 3.0 4K Workflow for Upscaling and Polishing AI Videos?

The second complication is frame-rate. Silent-era and early sound-era footage was frequently hand-cranked at 14 to 18 fps, and even 24 fps material from the 1930s was often shown on projectors that ran hot or cold. Naively interpolating to 60 fps with a generic optical-flow model can amplify judder because the original motion cadence is non-uniform. A correct workflow either preserves the original cadence or performs a cadence-aware interpolation pass that explicitly models variable exposure times. Industry tooling from Telestream, Adobe, and Boris FX now ships with film-specific modes because of exactly these issues, not as marketing decoration.

A third issue is resolution and grain. A 35mm archival scan from a 4K wet-gate scanner carries real sub-pixel detail that a 4K consumer model may discard. Conversely, a 2K or HD telecine carries limited detail and the AI has to generate information that was never on the negative. Conflating these two cases is a common mistake. The source determines whether the workflow is primarily a detail-preservation job or a detail-generation job, and the parameters should be different in each case.

Finally, archival work is rarely a single-pass activity. Restoration teams typically scan, repair, color-correct, and then upscale, with each stage reviewed by humans. AI has not replaced those stages; it has compressed the gap between them. A workflow that treats AI upscaling as the final step rather than one stage in a chain tends to produce footage that looks glossy but loses period character, and that is exactly the criticism historians have raised about some YouTube upscaling projects.

The Six-Stage Archival Upscaling Pipeline Used in 2026

A practical archival workflow in mid-2026 usually consists of six stages. The first stage is acquisition: a 4K or higher wet-gate scan of the original negative or print, captured at the highest available bit depth (typically 12- or 16-bit linear) and a resolution that exceeds the final delivery target. Scanning at 4K-plus means the AI has authentic texture to work with, not just a guess. Sources such as the National Film Board and BBC restoration departments follow this principle even when the final delivery is 1080p, because the headroom is used for stabilization and reframing later.

Stage two is physical cleanup. This is where dust, hairs, and mild scratches are addressed with digital cleanup tools, sometimes paired with neural inpainting models. Tools such as Mocha Pro and After Effects content-aware fill have been used here for over a decade, but newer neural options added to DaVinci Resolve 19 and standalone products like Aiarty now automate much of the routine pass.

Stage three is stabilization and registration. Archival scans frequently exhibit shrinkage, weave, and breathing. Modern AI stabilizers correct these more accurately than legacy 2D trackers because they understand 3D camera motion. The output of this stage is a stable, registered image sequence that the upscaler can process frame by frame without artifacts shifting around.

Stage four is the AI upscale itself. In 2026, the most capable consumer-accessible options for archival work include Aiarty Video Enhancer (with its dedicated 4K upscaling profile), Topaz Video AI, and hardware-accelerated paths such as Telestream Vantage with NVIDIA acceleration. For pure quality-per-pixel on film grain, Aiarty's recent releases and Topaz remain the two names that come up most in independent comparisons, while Vantage is favored when the upscale has to integrate into a broadcast finishing pipeline. Output targets are typically UHD (3840×2160) for streaming and DCI 4K (4096×2160) for theatrical or archive masters.

Stage five is interpolation. Once footage is at 4K resolution, frame-rate conversion to 23.976, 24, 25, or 59.94 fps can be done with optical-flow or learned interpolation. For archival material, the conservative choice is to maintain the original cadence and only interpolate if delivery requires a different rate. PetaPixel's 2020 piece on a 109-year-old New York City film being upscaled to 4K 60fps used this exact logic, and the approach has only become more standard since then.

Stage six is finishing: color grading, audio restoration, and format wrapping for delivery. AI denoising should be applied carefully here. Over-denoising film grain removes the very texture that makes film look like film, so the threshold is typically the noise floor of the original stock, not zero.

Tool Comparison: Which Upscaler Fits Which Archival Job

Different tools serve different parts of the chain. The table below summarizes how the most commonly cited options in 2026 line up against archival needs. Pricing is approximate and based on publicly listed tiers as of late 2025 through early 2026.

For an independent archivist working on a small project, Aiarty's film-tuned presets deliver a strong balance of cost and quality. For a small post-production house that already uses Topaz for client deliverables, adding an archival pass is straightforward. For a national archive or a film lab with deadlines measured in terabytes per week, Telestream Vantage with NVIDIA acceleration remains the production-line choice, and the integration with QC and watch-folder workflows justifies its higher operational cost.

Practical Step-by-Step: Upscaling a 1920s Silent Film to 4K

A concrete example helps anchor the workflow. Suppose you have a roughly 15-minute silent-era short scanned at 2K from a 35mm print, with moderate dust, mild weave, and exposure flicker. The goal is a streaming-ready 4K UHD master at 23.976 fps with preserved grain. The first practical step is to inspect the scan at 1:1 pixel ratio and note the worst frames: gates, scratches, splice marks, and severe flicker. These determine how aggressive the cleanup stage can be.

The second step is to perform an initial cleanup pass in a tool like DaVinci Resolve or After Effects. Dust and hair are removed with manual or assisted selection; deep scratches are inpainted. This is the stage that sets a ceiling on the final quality, so it should not be rushed. A bad cleanup produces a bad upscale.

The third step is stabilization. Apply the AI stabilizer to a proxy first, check that the weave correction is not introducing edge warping on titles, and only then process at full resolution. For most archival shorts this adds 20 to 40 minutes per reel of footage.

The fourth step is the AI upscale. Export a DPX or high-bit-depth EXR sequence from the cleaned, stabilized master. Run the upscale through Aiarty's 4K preset or Topaz's Artemis model at the same bit depth. If the source is a 4K scan, an upscale to UHD is essentially detail-preserving and can run at default settings. If the source is 2K, expect to spend time tuning the strength of detail generation so the result does not look digital.

The fifth step is to bring the upscaled image sequence back into Resolve for color grading and grain management. Add a calibrated film grain plate at a low opacity if the original grain was overly suppressed. Without this, faces can take on a waxy appearance that audiences immediately read as AI.

The sixth step is delivery: encode to H.265 or H.266 (VVC) for streaming, and to ProRes or JPEG 2000 for archival storage. Expect a 15-minute 2K source to expand to roughly 60 to 90 GB of 4K mezzanine files per minute of footage, so storage planning matters as much as GPU time.

Common Mistakes That Ruin Archival Upscales

The most damaging mistake is over-denoising. AI denoisers in 2026 are remarkably capable of removing grain, but film grain carries real image information. Pushing grain removal to 80 or 90 percent leaves a smooth, plastic-looking surface that no amount of grain plate can fully recover. A safer threshold is to denoise only the noise above the original stock's grain density, which often means a setting closer to 20 to 40 percent than to 90.

The second mistake is running a single model across an entire feature without checking for scene changes. Lighting changes, costume changes, and stock changes require different parameters. A model trained on well-lit outdoor footage can hallucinate textures inside shadow detail from a dimly lit interior scene. Splitting the project at reel or scene boundaries and processing them as separate batches with per-batch settings takes more time but produces dramatically better results.

The third mistake is skipping physical cleanup. AI models can soften the appearance of a scratch, but they cannot reliably distinguish a vertical scratch from a vertical mast in the original scene. Removing the physical artifact before the upscale prevents the model from inventing geometry to fill it.

A fourth mistake is ignoring color space. Archival scans are often Log C, linear, or a custom scanner profile. Feeding a Rec.709-tagged file into a model that assumes a different color space can produce oversaturated highlights and crushed shadows. The upscale should happen on linear or Log data, with color conversion as the final grading pass.

A fifth mistake is treating AI upscaling as historically neutral. Several historians have publicly criticized YouTube upscaling projects for removing period-appropriate grain, smoothing skin tones, and adding detail that could mislead viewers about what the original footage actually showed. Documenting the processing chain in the deliverable metadata is not just good practice; it is increasingly an ethical requirement for institutions working with cultural heritage material.

Hardware, Cost, and Timeline in 2026

GPU choice has a measurable effect on throughput. In 2026, mid-range professional cards such as the NVIDIA RTX 5000-series Ada-Next generation and the AMD Radeon AI PRO R9700 deliver the best price-performance for archival jobs that run for hours. Consumer cards still handle short clips fine, but for multi-reel features the time difference is significant. A 90-minute 2K upscale to 4K on a top-tier consumer GPU can take 18 to 30 hours of processing, while a professional card with hardware-accelerated paths such as the Telestream-NVIDIA integration can cut that to under 6 hours.

Cost-wise, the realistic budget for an independent archivist working on a single short film ranges from roughly 100 to 500 USD in software, plus 50 to 200 USD per month in cloud GPU rental if local hardware is insufficient. A small restoration studio running regular projects typically spends 1,500 to 5,000 USD per year on tool subscriptions, with hardware amortization spread across multiple projects. National archives and broadcast facilities operate in a different cost bracket entirely, with annual tooling budgets in the tens of thousands USD and dedicated operators.

Timeline depends on source condition. A relatively clean 4K scan of a 1970s feature can be upscaled and graded in a week by one operator. A heavily damaged silent-era nitrate print might require three to six weeks even with AI assistance, because of the manual review at each stage. AI has compressed these timelines meaningfully, but it has not eliminated the human review that archival work demands.

When to Upscale and When to Leave Footage Alone

Not every archival clip benefits from upscaling. Footage that will only ever be viewed on a phone screen in a social-media post often does not need a 4K master. Footage whose value lies in its historical authenticity, such as raw war-zone recordings or newsreel footage that documents a specific event, can lose meaning when AI adds texture that was never captured. In those cases, a clean 1:1 transfer with light denoising is more appropriate than a 4K upscale.

The decision to upscale should be driven by delivery context and audience expectation. A documentary broadcast in 4K on a streaming service benefits from a true 4K master. A home-video release from a personal collection is often fine at 1080p. AI upscaling is a tool for making archival footage usable in modern viewing contexts, and like any restoration choice it should serve the material rather than demonstrate the technology.

For institutions concerned about authenticity, the strongest practice in 2026 is to retain the original scan alongside the AI-upscaled version, document the processing chain in metadata, and publish both. That preserves the source for future re-processing as models improve, while still letting present-day audiences experience the material at modern resolutions.