# What is the best AI upscaler for old films in 2026?

ai-videoupscale.com · August 25, 2026

> The short answer: for most people restoring old films at home, the best AI upscaler for old films is Topaz Video AI, because it combines dedicated...

The short answer: for most people restoring old films at home, the best AI upscaler for old films is Topaz Video AI, because it combines dedicated film-restoration models (Artemis and Proteus), frame interpolation, deinterlacing, and batch processing in one package. If you want maximum control and don't mind a steeper learning curve, the open-source combination of Real-ESRGAN or Anime4K-style models inside Video2X or Flowframes, paired with AviSynth/VapourSynth filters like QTGMC for deinterlacing, produces results that rival commercial tools at zero cost — provided you have a reasonably modern GPU. For casual users who want a one-click experience, VideoProc Converter AI offers an accessible all-in-one toolkit that Fstoppers tested favorably for photo and video restoration work.

There is no single 'best' tool because old films are not one problem but several stacked problems: low resolution, film grain, interlacing artifacts from TV broadcasts, color fading, scratches and dust, and sometimes heavy compression from DVD-era encodes. The right choice depends on which of those problems dominate your source material. A 1940s black-and-white studio print needs different treatment than a 1990s VHS home video or a 480i broadcast recording of a cartoon.

**Also worth reading:** [What is the best AI video upscaler to 4K in 2026, and how do I upscale my videos to 4K properly?](https://ai-videoupscale.com/knowledge/what_is_the_best_ai_video_upscaler_to_4k_in_2026_and_how_do_i_upscale_my_videos_to_4k_properly.php) · [SeedVR2 vs FlashVSR comparison: which AI video upscaler delivers better 4K results in 2026?](https://ai-videoupscale.com/knowledge/seedvr2_vs_flashvsr_comparison_which_ai_video_upscaler_delivers_better_4k_results_in_2026.php) · [What is the best free AI video upscaler in 2026 for converting footage to 4K without hidden costs or watermarks?](https://ai-videoupscale.com/knowledge/what_is_the_best_free_ai_video_upscaler_in_2026_for_converting_footage_to_4k_without_hidden_costs_or_watermarks.php)

## Why Old Films Need AI Upscaling in the First Place

Most films made before the mid-2000s were shot on 35mm film, which actually contains far more detail than standard-definition video transfers capture. The problem is that home releases were mastered at 480p (DVD) or 576p (PAL), and broadcast versions were interlaced at 480i or 576i. When you play that content on a modern 4K television, every pixel gets stretched four times horizontally and vertically, producing soft edges, jagged lines, and visible blockiness.

Traditional upscaling algorithms — bicubic, Lanczos, spline — simply interpolate between existing pixels. They make the image bigger but not sharper, because they cannot invent detail that was never encoded. AI upscalers trained on millions of image pairs learn what real textures look like: skin pores, fabric weave, brick, foliage, film grain structure. When they encounter a soft edge, they predict plausible fine detail rather than just blurring the transition. This is why a good AI upscale of a DVD can genuinely look closer to a Blu-ray remaster than to a naive stretch.

That said, the technology has limits that enthusiasts and professionals have documented extensively. PetaPixel's coverage of AI-enhanced old films notes that not everyone is happy with the results, and No Film School has asked pointedly whether AI in film restorations has gone too far. The concern is real: models can hallucinate faces, smooth away intentional grain, and alter the texture that gives old films their character. The infamous case of the Super Mario Bros. Super Show! broadcasts on MeTV — covered by both Nintendo Life and Engadget in late 2023 — showed what happens when AI processing goes wrong on cel-animated cartoons: warped faces, melting backgrounds, and smeared line art. Cartoons are the hardest category for AI upscaling precisely because hand-drawn art violates the photographic assumptions baked into most models.

## The Contenders Compared

Here is how the leading options stack up as of August 2026:

| Feature | Topaz Video AI | Open-source (Real-ESRGAN + VapourSynth) | VideoProc Converter AI |
| --- | --- | --- | --- |
| Price | ~$299 one-time (frequent sales near $199) | Free | ~$45–79/year subscription |
| Ease of use | High — GUI, presets | Low — command-line and scripting required | Very high — guided workflow |
| Dedicated film models | Yes (Artemis, Proteus, Iris) | Community models, varies by project | General enhancement models |
| Deinterlacing | Built-in (good quality) | QTGMC via VapourSynth (reference quality) | Basic built-in |
| Frame interpolation to 60fps | Yes (Chronos/Apollo) | RIFE/Flowframes (free) | Limited |
| Batch processing | Excellent | Scriptable, unlimited | Good |
| GPU requirement | NVIDIA RTX recommended, 8GB+ VRAM ideal | Any CUDA-capable GPU; CPU possible but slow | Moderate; runs on modest hardware |
| Control over grain preservation | Model parameters only | Full manual control per filter chain | Preset-based |
| Output ceiling | Up to 8K/16K | Whatever your pipeline supports | Up to 4K |

Topaz Video AI remains the reference point for consumer-grade restoration because it packages everything: denoising, deinterlacing, upscaling, and interpolation with per-model parameter tuning. Its Proteus model lets you adjust detail recovery, dehalo, deblock, and grain synthesis independently, which matters enormously when working with damaged sources. The company ships frequent model updates, and version releases through 2025–2026 improved handling of archival footage noticeably.
The open-source route deserves serious consideration despite its difficulty. A Pocket-lint writer documented reviving an entire DVD collection 'with nearly perfect picture for free' using open-source tools, proving the ceiling is high. A typical pipeline looks like this: QTGMC in VapourSynth for deinterlacing (still considered the gold standard after more than a decade), then Real-ESRGAN or a community-trained model for the resolution bump, then optional RIFE interpolation if you want smoother motion. MakeUseOf has also highlighted that some GPUs ship with built-in upscaling features — NVIDIA's RTX Video Super Resolution, added to drivers in early 2023 — which works in real time in browsers and VLC without any extra software, though it offers less control than offline processing.

VideoProc Converter AI occupies the middle ground. Fstoppers tested it as an all-in-one media toolkit covering photos and video, and found it competent for casual restoration. It will not match Topaz on difficult archival material, but for family VHS tapes and DVDs where 'noticeably better' is the goal rather than 'archival grade,' it delivers strong value at a fraction of the price.

## How AI Upscaling Actually Works on Film Material

Understanding the mechanics helps you set realistic expectations. Modern upscalers are neural networks, typically convolutional or transformer-based, trained on pairs of images: a degraded low-resolution version and its high-resolution original. During training, the network learns the statistical mapping between blur, noise, and compression artifacts and the underlying sharp image. At inference time, it applies that learned mapping to your footage frame by frame.

Film adds complications that generic training data does not fully cover. First, grain: 35mm film has organic, per-frame random grain structure. A naive AI pass treats grain as noise and removes it, leaving a plasticky, waxy surface often described as the 'AI look.' Better workflows either preserve grain through the upscale or remove it, sharpen, and then re-synthesize grain so consecutive frames do not shimmer. Second, temporal consistency: each frame is processed somewhat independently, so details can flicker between frames — a known artifact class that newer temporal-aware models reduce but have not eliminated. Third, cadence: telecined film stored as 29.97fps interlaced video requires inverse telecine (IVTC) before upscaling, or you bake combing artifacts into every frame permanently.

This is why professional restorations still start from the original film elements scanned at 2K or 4K whenever possible. AI upscaling of a DVD is a reconstruction, not a recovery. It produces something that looks better on a 4K screen, but it is inventing detail the studio never released. Collectors should treat AI-upscaled discs as viewing copies, not replacements for future official remasters.

## Practical Steps: Restoring an Old Film Step by Step

Start with the best source you can obtain. If a film exists on Blu-ray, upscale from that, not the DVD — garbage in, garbage out applies brutally to AI models. For DVD sources, rip losslessly with MakeMKV or HandBrake in lossless mode; never feed a heavily compressed streaming rip into an upscaler, since the model will amplify compression blocks instead of removing them.

Next, handle interlacing and cadence before anything else. If your source is 480i broadcast material, run it through QTGMC (VapourSynth) or use Topaz's deinterlacing model first. If it is telecined film on DVD, apply inverse telecine to recover clean progressive 23.976fps frames. Skipping this step is the single most common mistake and ruins otherwise good results.

Then choose your upscale target conservatively. Going from 480p to 1080p usually yields convincing results; pushing the same source to 4K multiplies hallucination risk because the model must invent four times more detail. Many experienced restorers upscale SD sources to 1440p or 1080p and let the TV's scaler handle the final stretch, which often looks more natural than a full 4K AI pass.

Finally, evaluate motion, not just stills. Pause a frame and it may look stunning; play ten seconds and flickering textures or warping faces reveal themselves. Check scenes with fine patterns (brick, fences, credits text) and close-up faces, which are where artifacts concentrate. Export a short test clip first — 30 seconds of the hardest scene in the film — before committing hours of GPU time to a full feature. A 90-minute film upscaled on a mid-range RTX card typically takes 4 to 12 hours depending on model and output resolution, so test clips save real time.

## Where AI Upscaling Goes Wrong: Documented Failures

The Mario cartoon incident is worth studying because it encapsulates every failure mode. MeTV's broadcast of the 1989 Super Mario Bros. Super Show! appeared to have been processed with AI upscaling and frame interpolation, and the result was widely mocked: characters' faces drifted and morphed, outlines wobbled, and background art smeared. Cel animation has flat colors, inconsistent line weights, and off-model frames by design — exactly the irregularities AI models are trained to 'correct.' The lesson generalizes: hand-drawn animation, stop-motion, and heavily stylized footage need animation-specific models (Anime4K-family, or Topaz's Iris tuned for animation) and conservative settings, or no AI at all.

Over-smoothing is the second chronic failure. Removing all grain makes 1940s Hollywood films look like modern digital video, destroying period atmosphere. PetaPixel's reporting on AI-enhanced old films captures the purist backlash: many cinephiles argue that grain, gate weave, and even print damage are part of the artifact's history. If you care about this, look for tools with explicit grain preservation or re-graining controls, and keep enhancement strength below maximum — around 50–70% recovery often preserves texture while cleaning genuine damage.

Face hallucination is third. Models occasionally invent teeth, eyes, or skin texture that differs from the actor's actual appearance, which crosses an ethical line for documentary and historical footage. Jess Weatherbed's coverage of Adobe's Project Res Updater preview in October 2023 highlighted how seriously major vendors treat fidelity claims, but independent verification still matters. Always compare your output against a high-quality still photograph of the same scene or actor when restoring historically important material.

## Cost Breakdown and Hardware Requirements

Budget shapes your options more than anything else. Topaz Video AI costs roughly $299 at list price, though sales bring it closer to $199 several times a year; it includes a year of model updates, after which upgrades cost extra. VideoProc Converter AI runs about $45–79 per year depending on promotions. The open-source stack costs nothing but demands a GPU to be practical: expect 6GB of VRAM minimum for 1080p output and 10–12GB for comfortable 4K work. An RTX 3060 (12GB) handles most jobs well; an RTX 4070 or better cuts render times roughly in half.

Free alternatives exist at every tier. NVIDIA RTX Video Super Resolution costs nothing beyond owning an RTX 20-series-or-newer card and works live in Chrome, Edge, and VLC — MakeUseOf called out that 'the feature is already installed' for many users who never enabled it. MPV players with shader packs (FSRCNNX, Anime4K) provide excellent real-time upscaling during playback without ever rendering files. North Penn Now's comparison of six free AI video enhancers found meaningful quality gaps among them, reinforcing that 'free' spans a wide range from basic to near-commercial quality.

One hidden cost is storage. Upscaling a 90-minute film from a 4GB DVD rip to 4K ProRes or high-bitrate H.265 can produce 20–80GB files. Plan for fast NVMe scratch space and archive space accordingly, and encode final deliverables to H.265/HEVC or AV1 at sensible bitrates (15–25 Mbps for 4K SDR) to keep libraries manageable.

## Common Mistakes to Avoid

The most frequent error is over-processing: stacking denoise, sharpen, upscale, and interpolation at maximum settings until the film looks like a video game cutscene. Restraint wins. Run each stage at moderate strength and inspect results at 100% zoom on a calibrated display.

Second is ignoring audio and sync. Long GPU renders occasionally introduce frame drops or audio drift; always verify sync at the start, middle, and end of your export. Third is interpolating to 60fps blindly. The 'soap opera effect' is widely disliked for narrative film because 24fps motion cadence is part of cinematic grammar; high-frame-rate smoothing suits sports and documentaries far better than dramas. If you interpolate, consider 48fps as a compromise, or keep 24fps and let motion blur carry the realism.

Fourth is trusting presets. Default settings are tuned for average smartphone footage, not 1970s film prints. Spend an evening learning what the detail-recovery and artifact-suppression sliders actually do on your specific source; two hours of experimentation improves a whole library. Fifth is deleting originals. Keep your untouched rips forever — next year's models will be better, and re-running an improved pipeline on the clean source beats trying to undo a bad upscale, which is essentially impossible.

## When to Act and What the Future Holds

If you have a collection of DVDs or VHS tapes deteriorating in storage, there is a genuine urgency argument: magnetic tape degrades measurably every year, with VHS oxide shedding becoming noticeable after 20–30 years. Digitize tapes now with a decent capture device (a used Blackmagic Intensity or a quality USB capture unit), archive the raw captures, and upscale later when you have time — the capture is the irreplaceable step, not the enhancement.

The tool landscape keeps improving. Adobe's research into AI upscaling previewed in October 2023 has matured into shipping features across Creative Cloud, and open-source temporal-consistency research published through 2025–2026 continues to shrink the gap between free and paid pipelines. Whatever you choose today, build your workflow around preserving originals and documenting your settings, so any future improvement can be applied retroactively. That habit — more than any single tool purchase — is what separates a lasting restoration project from a one-off experiment.", "faq": [ { "q": "Can AI upscaling really make a DVD look like 4K?", "a": "It can make a DVD look dramatically sharper on a 4K screen, approaching Blu-ray quality in favorable cases, but it reconstructs rather than recovers detail. Results depend heavily on source condition, and fine textures are predicted, not restored. Treat it as enhanced viewing, not a substitute for an official remaster from original film elements." }, { "q": "Is AI upscaling bad for old movies?", "a": "It can be, if applied aggressively. Over-smoothing removes film grain, models can hallucinate facial details, and the MeTV Mario cartoon broadcast showed how badly generic AI processing damages hand-drawn animation. Conservative settings, grain preservation, and animation-specific models avoid most problems." }, { "q": "Do I need an expensive GPU to upscale old films?", "a": "A CUDA-capable GPU with at least 6GB VRAM makes the process practical; 10–12GB is comfortable for 4K output. CPU-only processing works but can take days per film. NVIDIA RTX owners also get free real-time upscaling via RTX Video Super Resolution built into drivers since early 2023." }, { "q": "Should I upscale old films to 60fps?", "a": "Usually not for narrative cinema. Interpolating 24fps film to 60fps creates the soap opera effect that many viewers dislike, and it contributed to the widely criticized Mario cartoon broadcast. Keep 24fps for dramas; consider higher frame rates only for sports, concerts, or personal home videos." }, { "q": "What should I do before AI upscaling a VHS tape?", "a": "Digitize it immediately with a quality capture device, since magnetic tape degrades every year and the capture itself is irreplaceable. Archive the raw capture untouched, deinterlace properly (QTGMC is the reference standard), then upscale from the clean progressive file. Never upscale directly from an interlaced source." } ], "quick_facts": [ { "label": "Category", "value": "AI video upscaling / film restoration software" }, { "label": "Timeline", "value": "A 90-minute film takes roughly 4–12 hours to upscale on a mid-range RTX GPU" }, { "label": "Cost", "value": "$0 (open-source) to ~$299 one-time (Topaz); subscriptions from ~$45/year" }, { "label": "Best for", "value": "Home archivists digitizing DVDs, VHS tapes, and broadcast recordings" }, { "label": "Hardware", "value": "CUDA GPU with 6–12GB VRAM recommended; RTX cards include free driver-level upscaling" }, { "label": "Key warning", "value": "Avoid aggressive AI processing on hand-drawn animation and preserve film grain" } ], "sources": [ "https://hackernoon.com/how-i-built-a-pipeline-to-restore-old-bw-photos-to-4k-color-using-open-source-ai", "https://www.pocket-lint.com/how-i-revived-my-old-dvd-collection-with-nearly-perfect-picture-for-free", "https://www.nintendolife.com/news/random-ai-has-ruined-a-classic-mario-cartoon-on-metv", "https://www.engadget.com/the-super-mario-bros-cartoon-is-back-but-looks-really-weird-and-ai-might-be-to-blame", "https://petapixel.com/old-movies-are-being-enhanced-with-ai-tools-and-not-everyone-is-happy", "https://nofilmschool.com/has-ai-in-film-restorations-gone-too-far", "https://fstoppers.com/can-ai-really-restore-old-photos-testing-videoproc-converter-ai", "https://www.makeuseof.com/gpu-upscale-old-videos-feature-installed", "https://northpennnow.com/top-6-free-ai-video-enhancers-compared", "https://www.the-sun.com/weatherbed-adobe-previews-ai-upscaling" ], "follow_up_keyword": "deinterlace VHS before AI upscaling"

Canonical: https://ai-videoupscale.com/knowledge/what_is_the_best_ai_upscaler_for_old_films_in_2026.php
Markdown: https://ai-videoupscale.com/knowledge/what_is_the_best_ai_upscaler_for_old_films_in_2026.php/index.md