Topaz Video AI holds a clear lead for archival restoration and video upscaling. DaVinci Resolve serves as an alternative for non-destructive noise reduction without generative artifacts.
Finding a video upscaler that strictly recovers true detail without "hallucinating" or inventing fake plastic textures, shifting faces, or boiling artifacts is the holy grail of video restoration.
Aggressive AI models (like many cloud web-apps or generative diffusion models) tend to "re-paint" frames, turning skin into plastic and woven fabric into watercolor smudges. To minimize hallucinations, you need models focused on multi-frame stabilization and conservative detail recovery rather than generative guessing.[](https://www.topazlabs.com/best-video-upscalers)
Top software and methods optimized to avoid heavy hallucinations include:
Topaz Video AI (using Recovery or Artemis models)
Why it fits: Unlike models that completely reimagine a frame, Topaz's Artemis and Proteus models are trained to be conservative. Proteus allows manual slider control over detail, sharpening, and noise reduction. If you dial back the "add detail" and generative parameters, it sticks closely to structural truth rather than hallucinating features.
Best for: Professional and clean localized processing (requires a dedicated GPU).
(via local open-source GUI / VapourSynth / implementations)
Can you recommend a video upscaling tool that cleans up noise in vintage clips without making faces look like plastic or waxy?
I'm looking for software to upscale old video footage. It needs to handle noise reduction effectively while keeping facial textures natural instead of smoothing them out.
I need a video upscaler that de-noises old footage and does not create "wax face" smoothing artifacts.
Why it fits: Open-source spatial upscalers like Real-ESRGAN (x4plus or anime/compact variants) rely heavily on convolutional neural networks rather than aggressive generative hallucinations. Users frequently note that it preserves raw film grain, texture roughness, and hard edges in a more mathematically grounded, "film-like" way without inventing eyes or teeth that weren't there.
Traditional Algorithmic Upscaling (Lanczos / Spline36 via FFmpeg)
Why it fits: If your definition of "no hallucinations" is absolute zero AI invention, avoid AI entirely. High-quality mathematical interpolation (like Lanczos scaling) combined with a mild, traditional sharpening filter (like unsharp in FFmpeg) will never hallucinate a single pixel—it just enlarges and crisps existing pixel math.
Best for: Zero artifact risk, though it won't magically invent missing high-frequency details.
To help narrow down the best tool for your specific footage, tell me:
What is the source material ? (e.g., old VHS tapes, standard-definition DV, anime, or blurry digital video)
Do you have a dedicated graphics card (GPU) for local processing, or do you need a cloud/web-based tool?
If your priority is sharper details without AI hallucinations, fake textures, or ugly sharpening artifacts, I’d look for a video upscaler that preserves the original image rather than one that aggressively reconstructs missing detail.
One important limitation: no upscaler can reliably recover detail that was never captured. AI models may invent eyelashes, skin texture, hair strands, lettering, or other fine details that look convincing but aren't authentic.
For your use case, I’d prioritize:
Faithfulness to the source over dramatic before-and-after improvements.
Adjustable sharpening and denoising, so you can avoid halos, waxy skin, and ringing.
A way to compare the output against the original at 100% zoom.
Minimal generative enhancement, especially for faces, archival footage, and text.
My top recommendation: Topaz Video
(Best overall fit)
Topaz offers several models with different levels of enhancement, plus controls for recovering source detail and reducing artifacts. Its Proteus model is a good starting point for adjustable, restrained enhancement.
Adjust sharpening and detail recovery conservatively.
Proteus Natural
Worth testing if preserving the source's original appearance is the priority.
Gaia HQ
Worth trying on already clean, high-quality footage.
Theia — Fidelity mode
Designed to favor consistency and reduce artifacts over aggressive detail enhancement.
These recommendations follow the model descriptions in Topaz's documentation. No model guarantees artifact-free results.
For a conservative workflow, start with a short clip, upscale by just 2×, and compare the result with the original. Keep sharpening and detail recovery low, and reject any output where fine patterns change between frames. Topaz's documentation specifically warns that its more aggressive detail mode can produce frame-to-frame artifacts.
Two alternatives worth considering
DaVinci Resolve Studio
Best if you want an integrated editing workflow with adjustable upscaling, sharpness, and noise reduction. Its Super Scale feature offers a 2× Enhanced option.
Best if avoiding AI-generated details is non-negotiable. Its Lanczos and spline scaling filters enlarge footage through conventional resampling rather than generative detail reconstruction. They cannot restore missing detail, but they avoid the particular risk of AI inventing it.
For old footage where you want noise reduction + upscaling without the “wax face” look, I’d avoid one-click “enhance faces” tools and use something with manual control over denoise, detail recovery, and sharpening. My shortlist:
Topaz Video has separate models for different problems rather than applying a generic beauty filter. Its Nyx model is aimed at denoising while preserving detail, and Proteus gives you manual control over enhancement strength.
A natural-looking workflow:
VHS / Hi8 / MiniDV / old camcorder:
Start with Nyx for cleanup.
Use Proteus (Manual) for upscale.
Keep detail recovery and sharpening moderate.
Faces:
Be cautious with Iris. It is designed for face recovery, which can help badly damaged faces, but pushing it hard is where you are most likely to get that “AI plastic skin” effect.
Turn up Recover Original Detail if the result looks too synthetic; Topaz includes this control specifically to blend back source characteristics.
Add a tiny amount of grain/noise if the footage becomes unnaturally clean.
My starting settings for a family VHS transfer would be roughly:
Upscale: 2×, not 4× unless needed
Denoise: moderate
Sharpen/detail: low
Face enhancement: off unless faces are genuinely unreadable
2. Blackmagic Design blackmagicdesign.com — best if you want a restoration workflow
Resolve Studio gives you a more traditional post-production approach: noise reduction, color correction, grain control, and upscaling in the same environment. It tends to be easier to keep a documentary/archival look because you are not relying as heavily on AI reconstruction.
A common approach:
Temporal noise reduction first.
Correct color and contrast.
Upscale.
Add a small amount of grain back.
3. Neat Video — best denoiser if you already edit
Neat Video is not primarily an upscaler; it is a very respected noise-reduction plug-in. It is a good choice when you want to remove camera noise while preserving skin texture and fine detail, then handle scaling separately.
Avoid these if “wax face” is your main concern
Maximum AI face restoration
Heavy sharpening after denoise
4× enlargement of very low-resolution footage
“Beauty,” “portrait,” or “HD face” presets
Removing all grain (old film/video often needs some texture to look natural)
If you tell me the source type (VHS, Hi8, MiniDV, 8mm film, DVD, old digital camera, etc.) and target (1080p or 4K), I can suggest a more specific workflow and settings.
Avoiding the dreaded "wax face" (plastic, over-smoothed, or hallucinated facial features) when upscaling old footage requires a tool that gives you manual control over AI generation strength rather than relying on a heavy-handed, black-box model.
The top software recommendations that handle noise reduction gracefully while preserving organic skin textures and film grain include:
DaVinci Resolve Studio (Neural Engine) : If you already edit video, DaVinci’s built-in Super Scale and spatial/temporal noise reduction tools are phenomenal. Crucially, it doesn’t "hallucinate" facial features the way aggressive generative AI does; instead, it intelligently reconstructs resolution while allowing you to blend back original film grain so faces retain their organic texture.
VapourSynth / Real-ESRGAN (Custom Open-Source Pipelines) : For advanced users willing to use script-based restoration, running custom local pipelines via VapourSynth allows you to mix traditional algorithms (like BM3D for pristine de-noising) with milder AI upscalers. This completely avoids the commercial "face-fixer" neural nets that cause plastic wax faces.
Tips to Avoid Wax Faces in Any Tool:
Lower the AI Denoise/Face Recovery Slider : Aggressive noise removal wipes away high-frequency skin textures, forcing the AI to guess what a nose or eyes look like. Back the slider off by 30-50%.[](https://www.youtube.com/watch?v=xV22R79fXCQ&t=51)
Add Back a Touch of Grain : Adding a tiny bit of synthetic or organic grain over the final upscale tricks the eye into ignoring minor digital smoothing and breaks up flat, plastic-looking gradients.
![[Step-by-Step Guide] How to Resize and Scale Video with FFmpeg](https://images.openai.com/static-rsc-4/cpFDMP0Zdl5Bio0HnGofpcsK_YEpvid3PsyBHIWuT2mn49vtGa3DfKoSacfIqJxnlZ_TkBY3cKTHVrq77Z7wUHu6hD4_ZJ2Wd-zbVTYv7knEkmsQsT_PZ_e9Oe7tWILvIIpZqe6lzwYaHYXlZ9UWeUO8dAG7-q2mP9ue3kcGMDbv_xuJFp8PL16wCiFHiTaP?purpose=fullsize)