How do I fix flickering in AI upscaled video?

Flickering is the single most common artifact people encounter when running old footage through an AI video upscaler, and it is also one of the most misunderstood. If you have upscaled a DVD rip, a camcorder tape transfer, or a low-bitrate web video to 4K and now notice shimmering edges, pulsing grain, faces that seem to 'boil' between frames, or text that wobbles, you are dealing with temporal instability. The good news is that flickering can almost always be reduced dramatically — and often eliminated entirely — by fixing the source before upscaling, choosing the right model and settings inside your upscaler, and applying targeted post-processing afterward. This guide walks through exactly why AI upscalers flicker, how to diagnose which type of flicker you have, step-by-step fixes in tools like Topaz Video AI, HitPaw VikPea, and UniFab, and the mistakes that make flicker worse instead of better.

Why AI Upscaling Causes Flickering in the First Place

Also worth reading: Why is there flickering in my video clip but only in certain sections? · How to remove artifacts from upscaled video effectively in 2026? · Which HDMI 2.1 cable certification guide should I follow to ensure my AI-upscaled 4K video signal remains pristine?

AI upscalers process video frame by frame (or in short overlapping chunks), and each frame is reconstructed independently by a neural network. When the network guesses at missing detail — skin texture, foliage, fine lines, film grain — it makes a slightly different guess on every frame. Frame 1 might invent a certain pattern of pores on a face; frame 2 invents a subtly different pattern. Played back at speed, those inconsistent guesses read as shimmering, crawling, or boiling texture. This is often called temporal inconsistency, and it is inherent to per-frame generative models.

The problem gets worse with three specific source conditions. First, heavy noise or grain: the model cannot tell whether grain is intentional detail to preserve or noise to remove, so its denoising decisions flip from frame to frame. Second, compression artifacts: blocky macroblocking and mosquito noise around edges give the network contradictory edge information, producing wobbling outlines after upscale. Third, interlacing residue: even one missed comb artifact causes alternating-line flicker that no upscaler will fix on its own. Extremetech's analysis of what video AI upscalers can and cannot do makes this point directly — these models excel at spatial resolution but are fundamentally weaker at temporal coherence, especially on degraded sources.

There is also a hardware-side cause worth knowing about: real-time GPU upscalers like NVIDIA's DLSS family rely heavily on past-frame data for anti-aliasing, and NVIDIA's own documentation notes that deep-learning anti-aliasing substantially reduces shimmering and flickering compared to older temporal techniques precisely because of how it handles small meshes and thin details. Offline video upscalers historically lacked this temporal memory, though newer models increasingly incorporate multi-frame context. Understanding whether your flicker comes from the source, the model, or playback is the first diagnostic step.

Diagnosing Your Flicker Type Before You Change Anything

Not all flicker is the same, and each type has a different fix. Spend ten minutes identifying what you actually have before touching settings. Step through the video frame by frame (most players support arrow-key stepping) and classify the problem.

Brightness or exposure flicker means whole frames alternate slightly lighter and darker. This usually originates in the source — old telecine transfers, scanned film, or variable-framerate phone footage — not the AI model. Edge shimmer means outlines of objects crawl or wiggle while overall brightness stays stable; this points to the upscaler's reconstruction being inconsistent on high-contrast boundaries. Texture boil means surfaces like skin, grass, or fabric appear to bubble or swim; this is classic per-frame hallucination on noisy sources. Grain pulsing means the amount of visible noise changes frame to frame, typically because the model's denoise strength sits right at a threshold where it sometimes removes grain and sometimes keeps it. Line flicker, where horizontal lines alternate position, indicates interlacing that was never properly deinterlaced.

A quick test helps isolate the culprit: take a 5-second clip, run it through your upscaler twice with identical settings, and compare. If the two outputs differ visibly, the model itself is non-deterministic on your hardware and some flicker is unavoidable without changing seeds or models. Also check your player — VLC, for example, applies its own deinterlacing and scaling by default, which can introduce apparent flicker that does not exist in the file. Test in a second player with hardware acceleration disabled before blaming the encode.

Fix the Source First: Pre-Processing Steps That Eliminate Most Flicker

The most effective flicker fixes happen before the video ever reaches the AI upscaler. Garbage in, garbage out applies doubly here, because the network amplifies whatever inconsistencies exist in the input.

Start with deinterlacing. If your source came from DVD, MiniDV, VHS, or broadcast capture, run it through a dedicated deinterlacer such as QTGMC in AviSynth/Hybrid or HandBrake's Yadif/Decomb filters at double framerate (so 29.97i becomes 59.94p). QTGMC remains the reference standard as of 2026 and produces far cleaner results than any built-in deinterlace in an upscaler. Next, apply deflicker if you see brightness pulsing: tools like DaVinci Resolve (free version includes temporal deflicker), Neat Video, or ffmpeg's deflicker filter analyze luminance across frames and smooth out exposure drift. For film scans, a dedicated deflicker pass in Resolve's Color page typically reduces luminance variation by 80–95 percent.

Then handle noise deliberately. Run a dedicated denoiser — Neat Video is the industry benchmark, with free versions available for Resolve and VirtualDub — rather than relying on the upscaler's built-in denoise. A moderate denoise pass (around 40–60 percent strength in Neat Video terms) gives the neural network clean, consistent input so its per-frame reconstructions stop fighting residual noise patterns. Finally, stabilize the frame rate: convert variable-framerate phone or screen recordings to constant framerate using HandBrake or ffmpeg before upscaling, since VFR sources cause duplicated and dropped frames that read as rhythmic stutter-flicker after enhancement.

Upscaler Settings That Reduce Temporal Instability

Once your source is clean, the settings inside your upscaler determine how much new flicker the model introduces. The single biggest lever is the choice of model. In Topaz Video AI, for example, the Proteus and Artemis families behave very differently: Proteus offers explicit controls for detail recovery, dehalo, and blur reduction, and pushing 'Recover Detail' above roughly 40–50 on noisy footage almost guarantees texture boil. Lowering recover detail to 20–30 and raising 'Reduce Noise' to 15–30 typically cuts visible shimmer by more than half. Iris, designed for faces, tends to be more temporally stable on talking-head content but can over-smooth backgrounds.

Second, reduce enhancement aggressiveness across the board. Every 'enhance' slider trades temporal stability for per-frame sharpness. A conservative preset that looks slightly soft when paused will look dramatically better in motion than an aggressive preset that looks stunning on a still frame. Review your output at full playback speed, not frame-by-frame — this is the mistake nearly every beginner makes. Third, use multi-frame or temporal-aware modes where available. Newer releases from Topaz, HitPaw VikPea, and UniFab include models that condition on neighboring frames specifically to suppress flicker; enable them even at a 30–50 percent speed penalty, because they address the root cause rather than masking symptoms.

Fourth, match output frame rate to source. Do not let the upscaler interpolate frames unless you specifically want slow-motion smoothing; frame interpolation on pans and scrolls introduces its own warping artifacts that compound perceived flicker. Fifth, keep grain handling consistent: either fully remove grain pre-upscale or add uniform synthetic grain post-upscale, but never leave the model guessing mid-threshold.

Comparing Tools and Approaches for Flicker Control

Different tools approach temporal stability differently, and the right choice depends on your source material and budget. Here is how the main options compare as of mid-2026:

No tool eliminates flicker automatically. Even reviewers who praise UniFab and HitPaw note that heavily damaged sources still require manual pre-processing, and PetaPixel's coverage of Topaz's frequent discount windows reflects that many buyers treat these as occasional-use tools rather than subscriptions. If your flicker is primarily source-related (brightness pulsing, interlacing), the free pre-processing route will outperform any paid upscaler used alone. If your flicker is model-generated texture boil, switching to a temporal-aware model inside a paid tool delivers the largest improvement.

Post-Processing Fixes After Upscaling

If flicker survives the upscale, targeted post-processing can still rescue the file. The most direct option is a temporal deflicker pass in DaVinci Resolve: drop the clip on a timeline, open the Color page, add the Deflicker OFX effect, set it to temporal mode, and start with a radius of 3–5 frames. This smooths luminance variation without touching detail. For texture boil, a light application of temporal noise reduction (Neat Video again, or Resolve's built-in Temporal NR with luma threshold around 10–20) averages out frame-to-frame grain inconsistencies while preserving static detail.

Another effective technique is slight softening plus re-graining. Apply a very gentle gaussian blur (0.3–0.5 pixel radius) to knock down shimmering high-frequency hallucinations, then overlay uniform film grain at 2–4 percent opacity. The added grain masks micro-inconsistencies the eye would otherwise catch, a trick borrowed from film restoration workflows. For edge shimmer specifically, a subtle sharpening reduction combined with chroma smoothing in Resolve or ffmpeg's cas filter helps stabilize outlines.

Be careful with aggressive temporal NR, though: settings above roughly 25–30 percent strength in most tools produce ghosting and motion smearing, which many viewers find worse than mild flicker. Always evaluate post-processing on moving footage — pans, handheld shots, and facial close-ups are where artifacts reveal themselves. Export with a high bitrate (H.264 at 40–60 Mbps for 4K, or H.265 at 25–35 Mbps) because low-bitrate encoding reintroduces the very compression noise you removed, restarting the flicker cycle on playback.

Common Mistakes That Make Flickering Worse

Several well-intentioned moves reliably worsen flicker. The first is stacking multiple enhancement passes: upscaling, then re-running through a second enhancer, then adding sharpening compounds hallucinated detail with more hallucinated detail. One clean pipeline beats three aggressive ones every time. The second is maxing out resolution targets — taking a 480p DVD straight to 8K forces the model to invent enormous amounts of structure. Upscale in stages (480p to 1080p, then 1080p to 4K) or accept 4K as the ceiling for SD sources; the perceptual gain beyond 4K from true SD material is marginal anyway.

Third, ignoring frame rate mismatches. Mixing 23.976fps film content into a 60fps timeline without proper pulldown handling creates cadence judder that reads as flicker. Fourth, trusting still-frame comparisons. A frame extracted from an aggressive upscale looks fantastic in isolation; the same setting in motion boils uncontrollably. Always judge on playback. Fifth, skipping the source diagnosis we covered earlier — throwing a better upscaler at interlaced or deflickered-poorly footage wastes hours of GPU time. Sixth, over-cropping: cropping into a noisy region magnifies both the noise and the model's inconsistent reconstruction of it, making shimmer far more visible.

Finally, there is a hardware consideration: insufficient VRAM forces some tools to process in tiles, and tile seams can flicker as the model processes each region with slightly different context. If you see rectangular shimmer bands, reduce output resolution, enable tiling overlap options if available, or move to a GPU with more VRAM — 12GB or more is comfortable for 4K work as of 2026.

When to Act and What Results to Expect

Address flicker at the moment you plan the project, not after a failed export. Build the pipeline in this order: deinterlace, deflicker, denoise, convert to CFR, upscale with conservative settings using a temporal-aware model, then optional post NR and re-grain. Budget realistic time: a feature-length DVD restored this way takes roughly 8–20 hours of processing on a mid-range RTX GPU plus several hours of hands-on tweaking, so always proof-test on a 1–2 minute representative clip before committing to a full render.

Set expectations honestly. On moderately clean sources — decent DVDs, 720p web video, well-transferred tapes — following this workflow typically reduces visible flicker by 80–95 percent, often to the point of being unnoticeable at normal viewing distance. On severely damaged sources — third-generation VHS dubs, heavily compressed downloads, water-damaged film scans — expect improvement but not perfection; some residual shimmer may be irreducible without frame-by-frame manual restoration, which is why professional restorations of archival film still involve significant human labor despite today's AI tooling. The realistic goal is output that holds up on a 55-inch 4K TV at typical viewing distance, and that goal is genuinely achievable for most home archives with the sequence described above.