AI Upscaling vs Bicubic and Lanczos: What Actually Changes

AI upscaling and traditional upscaling both turn small frames into big ones. The difference is where the new pixels come from. Bicubic and Lanczos calculate them from neighboring pixels, so the picture gets bigger but not more detailed. An AI upscaler predicts detail from patterns it learned in training.

As a result, edges, hair, and textures can end up sharper than the source, and occasionally invented. Which one fits depends on how big the jump is and what state the source is in.

• Traditional upscaling (nearest neighbor, bilinear, bicubic, Lanczos) calculates each new pixel from the pixels around it. The picture gets bigger, but no new information appears.
• AI upscaling uses a trained model to predict plausible fine detail, and it usually reduces noise and compression damage at the same time.
• Lanczos looks crisper than bicubic, with more risk of faint halos along hard edges. Nearest neighbor suits pixel art and little else.
• AI earns its extra processing time on jumps of 2x or more, on compressed or noisy sources, and on line art. For small resizes of clean footage, a good traditional scaler is often enough.
• Neither kind can recover detail the camera didn't capture. Judge results at 100% zoom on the screen the video is meant for.

Upscaling increases the pixel dimensions of an image or video, for example from 1280 × 720 to 1920 × 1080. Every upscale has to fill in pixels that didn't exist. A 1080p frame has 2,073,600 pixels and a 720p frame has 921,600, so about 56% of the pixels in a 720p-to-1080p upscale are new.

Your TV, browser, and editing app upscale constantly, usually with a traditional scaler, whenever a video is smaller than the area it's shown in. HandBrake's documentation points out the cost: videos below 720p are typically blown up to fill the display, which magnifies even minor imperfections.

Each traditional method is an interpolation rule: a formula for how much each nearby source pixel contributes to a new one. FFmpeg's scaler documentation lists the common ones.

• Nearest neighbor (point sampling). Copies the closest source pixel. The result is blocky, but hard pixel edges stay hard, which is exactly what pixel art and retro game captures need when scaled by a whole number (2x, 3x, 4x).
• Bilinear. Blends the nearest pixels in a straight line. Smooth and soft.
• Bicubic. Fits a smooth cubic curve through nearby pixels. Sharper than bilinear and the usual general-purpose choice. It's FFmpeg's default, configured as a Mitchell-Netravali-style curve.
• Lanczos. A windowed sinc filter that weighs a wider neighborhood (three taps by default in FFmpeg). It keeps edges crisper than bicubic, and along high-contrast edges it can add faint light or dark outlines, often called ringing.
• Others. Area averaging is mainly for downscaling (FFmpeg notes it's equivalent to bilinear when upscaling); Gaussian and spline sit between the options above.

What they share matters more than how they differ: every output pixel is a weighted average of input pixels. That's why a traditionally upscaled 720p video still looks like 720p on a bigger canvas. Edges get smoother or crisper, but no new texture appears, and any noise, compression blocks, or blur in the source grow with everything else.

An AI upscaler, often called a super-resolution model, is a neural network trained on pairs of low- and high-resolution images. In training it learns which high-resolution patterns usually correspond to a low-resolution patch: what brickwork, eyelashes, or foliage tend to look like with more pixels. At run time it predicts those patterns for your frames.

That has three consequences:

1. It adds plausible detail. Edges come out clean and textures look denser than any interpolation can make them.
2. It usually cleans up as it scales. Models trained on degraded inputs also learn to reduce noise, compression blocks, and softness. KwaFlux's Smart Repair does this on purpose: it reads each clip's quality detection result and combines denoising, sharpening, color and brightness correction, upscaling, and face and fine-detail restoration as needed.
3. The detail is a prediction. When the source doesn't hold enough information, the model draws something that looks right rather than what was there. Small faces, distant text, and fine repeating patterns are where this shows first.

Video adds a fourth: consistency across frames. If texture is predicted frame by frame, it can shimmer as the clip plays. Many video models use neighboring frames to keep detail steady, and it's still the first thing to look for in a preview.

Sidenote: "AI upscaling" and "AI enhancement" overlap. Many upscalers repair and resize in one pass, which is why the same clip can look cleaner at its original size too. If you only want a bigger frame with nothing else changed, a traditional scaler is the honest tool.

• Small resizes of clean footage. Going from 1080p to 1440p with a sharp, well-exposed camera file rarely needs more than bicubic or Lanczos.
• The screen will scale anyway. If the viewer's TV or the player is going to scale the video regardless, pre-scaling with the same kind of math adds file size and little else.
• Edit timelines and proxies. When a clip just has to match a project resolution, size matters more than detail.
• Pixel art and retro captures. Nearest neighbor at 2x, 3x, or 4x keeps every pixel square.

• Jumps of 2x or more: 720p to 1440p, 1080p to 4K, SD to HD. The bigger the jump, the more of the frame a traditional scaler has to smear across. See how to upscale 720p to 1080p for a worked example.
• Damaged sources. Compression blocks, sensor noise, and soft focus get bigger under interpolation. AI upscaling usually reduces them on the way up.
• Anime and line art. Clean edges and flat color are what the eye checks first. A model tuned for animation keeps lines thin and fills flat areas cleanly.
• Footage you'll keep. For family video and archives, the extra processing time is spent once. For realistic expectations from very small sources, read 480p to 4K: what AI can recover.

Reading about scalers only gets you so far. Pick one shot with fine detail (hair, foliage, fabric, small text) and compare at the size you'll deliver.

In KwaFlux

1. Open Quality Repair. In the sidebar, go to Quality Enhancement and open the Smart Repair card, or a Scene Repair card that matches your footage, such as Anime.
2. Import the clip and set the output size. In Export Settings, set Resolution to your target, for example 1080p or 4K.
3. Render a preview. Choose Preview 1s, 3s, or 5s on the detailed shot.
4. Compare against a same-size baseline. In the before/after view, the Before (same size) baseline is the source upscaled to the output resolution with conventional scaling. Putting the AI result next to it shows what the model adds beyond plain resizing, at identical dimensions.
5. Read the scores, then your eyes. With Compare quality after preview turned on in Settings › Preview, KwaFlux scores the change: Clarity, Cleanliness, Aesthetic, and Technical, plus Resolution gain and Similarity to source (VMAF). The app's own hints say how to read them: VMAF is labeled "Higher means fewer changes, not better quality," and the panel notes that these are objective algorithmic metrics and your own eyes make the final call. For how the scoring works, see how we score video quality.

With FFmpeg (free)

Make a bicubic and a Lanczos version, then stack them side by side:

ffmpeg -i input.mp4 -vf "scale=1920:-2:flags=bicubic" -c:v libx264 -crf 16 -an bicubic.mp4
ffmpeg -i input.mp4 -vf "scale=1920:-2:flags=lanczos" -c:v libx264 -crf 16 -an lanczos.mp4
ffmpeg -i bicubic.mp4 -i lanczos.mp4 -filter_complex hstack -c:v libx264 -crf 16 compare.mp4

-2 keeps the aspect ratio and rounds the height to an even number. A low CRF (16) keeps the encoder from hiding the differences you're trying to see. Open compare.mp4, pause on the detailed shot, and look at 100% zoom: left is bicubic, right is Lanczos.

Sidenote: flags=lanczos is the long-standing way to pick an algorithm in FFmpeg's scale filter. The current scaler documentation marks algorithm selection through flags as deprecated in favor of a separate scaler option, so run ffmpeg -h filter=scale to see what your build accepts.

Pro tip: Compare at the size people will watch. A difference that's obvious at 100% zoom on a desktop monitor may vanish on a phone, and a halo you can't see on a laptop may be distracting on a 65-inch TV.

• Recover detail that wasn't captured. Unreadable text stays unreadable. An AI model may draw letter-like shapes, which is worse than leaving it soft.
• Turn heavy focus blur into a sharp shot. Mild softness improves; a badly missed focus doesn't come back.
• Make motion smoother. Upscaling adds pixels, not frames. For smoother motion, that's frame interpolation.
• Undo artifacts baked in by earlier processing. Over-sharpening halos and strong noise reduction from a phone or a previous export are part of the image now. AI can soften some of them, not reverse them.

Traditional scalers decide how to stretch the pixels you have. AI upscalers decide what the missing pixels probably looked like. Use the first when the source is clean and the jump is small, the second when the jump is large or the source is damaged, and compare them side by side before you commit to an export.

KwaFlux runs its video upscaler on your own GPU. AI processing runs locally on your GPU — media files are not uploaded for processing. It runs on Windows 10 (64-bit, version 1809 or later), Windows 11, and macOS 12 or newer on Apple Silicon. Previews need no account: render a 1-, 3-, or 5-second preview on any clip. Signing in with an active paid plan is required only to export, convert, repair a file, or download a file from a link.