AI Upscaling vs Bicubic and Lanczos Interpolation, Explained

When you enlarge an image, every new pixel needs a value. Traditional resampling calculates that value from nearby pixels using a fixed formula. AI super-resolution predicts it using a model that has learned what real images look like. Both approaches have a place, and understanding how they work makes it obvious when to use which.

The problem every upscaler solves

Take a 100×100 image and enlarge it 4× to 400×400. You now need 160,000 pixels but only have 10,000 of them. The other 150,000 must be filled in. Interpolation fills them by blending known neighbours. An AI model fills them by recognising patterns such as edges, fur or text and generating detail consistent with them.

Traditional interpolation methods

Nearest neighbour

Each new pixel copies the value of the closest original pixel. Nothing is blended, so the result is a grid of larger, hard-edged squares.

• Good for: pixel art and screenshots of UI, where you want blocks to stay blocks. Use whole-number factors (2×, 3×, 4×) or the blocks become uneven.
• Bad for: photos, where it produces obvious jagged staircasing.

Bilinear

Each new pixel is a weighted average of the 2×2 nearest original pixels, blended linearly in both directions.

• Smoother than nearest neighbour, but noticeably blurry at large factors.
• Common in real-time graphics because it is cheap.

Bicubic

Uses a 4×4 neighbourhood (16 pixels) and a cubic curve to weight them. Because the curve can slightly overshoot, bicubic preserves more apparent edge contrast than bilinear.

• The long-time default in many image editors.
• Produces soft results at 2× and quite blurry ones at 4×.
• Can create faint light or dark fringes along high-contrast edges.

Lanczos

Uses a windowed sinc function, the mathematically ideal reconstruction filter cut down to a practical size. The common Lanczos-3 variant considers a 6×6 neighbourhood (36 pixels).

• Generally the sharpest classical method and a strong choice for downscaling.
• Prone to ringing: thin ripples parallel to strong edges.
• Still cannot create detail. A soft source stays soft, just larger.

Comparison at a glance

Method	Neighbourhood	Speed	Sharpness	Typical artifact
Nearest neighbour	1 pixel	Fastest	Blocky	Staircase edges
Bilinear	2×2	Very fast	Soft	Blur
Bicubic	4×4	Fast	Moderate	Slight halos
Lanczos-3	6×6	Fast	Highest classical	Ringing
AI super-resolution	Learned, large	Slower	High, with new detail	Invented texture

If you want to try the classical methods yourself, ImageMagick makes it a one-liner:

# Lanczos 2x enlargement
magick input.jpg -filter Lanczos -resize 200% out-lanczos.png

# Nearest neighbour for pixel art (keeps hard edges)
magick sprite.png -filter Point -resize 400% sprite-4x.png

The fundamental limit of interpolation

All of these methods are interpolation: they only ever produce values in between (or slightly beyond) the values that are already there. That is why they never add detail. An edge that spans three blurry pixels in the original will span twelve blurry pixels after a 4× enlargement. Sharpening afterwards increases edge contrast but cannot reconstruct the hair strands or fabric weave that were lost.

How AI super-resolution works

AI upscalers are neural networks trained on huge numbers of image pairs. Training data is typically made by taking high-resolution images and degrading them: downscaling, blurring, adding noise and JPEG compression. The network learns to map the degraded version back to the original.

A few milestones give a sense of how the field developed:

• SRCNN (2014) showed that a small convolutional network could beat bicubic interpolation on standard benchmarks.
• SRGAN (2017) and ESRGAN (2018) added adversarial training, where a second network judges whether output looks real. This pushed results toward sharp, natural-looking texture instead of the slightly smooth “average” outputs of earlier models.
• Real-ESRGAN (2021) trained on more realistic, mixed degradations, which made it far better on real-world photos that are compressed and noisy rather than just downscaled.
• Diffusion-based upscalers generate detail through an iterative denoising process. They can produce very convincing texture at large factors, and they can also drift further from the original content.

The key point: the model is not looking up the true detail. It produces detail that is statistically consistent with what it has seen. On a brick wall that means convincing bricks. On a face, it usually means a convincing face, but small features can change.

Strengths and weaknesses side by side

Interpolation

• Fully predictable and faithful: it never introduces content that was not implied by the source.
• Fast enough for real time and available everywhere.
• Soft and lifeless beyond about 1.5×.

AI super-resolution

• Crisp edges and plausible texture at 2× and 4×.
• Can remove noise and compression artifacts in the same pass.
• Slower and needs more compute.
• Can hallucinate: invented patterns, altered small text, subtly changed faces.
• Results depend on matching the model to the content (photo vs illustration).

When to use which

• Downscaling anything: Lanczos or bicubic. AI adds nothing when you are throwing pixels away.
• Small enlargements (up to ~1.5×) of clean images: Lanczos is quick and faithful.
• 2× to 4× of photos, product shots or artwork: AI super-resolution, with a model matched to the content.
• Pixel art: nearest neighbour at integer factors, or a dedicated pixel-art algorithm.
• Evidence, medical, scientific or legal imagery: interpolation only, because invented detail is unacceptable.
• Hitting an exact size: AI upscale to the next factor up, then Lanczos down to the exact dimensions.

Spotting a bad upscale

Whichever method you use, check the output at 100% zoom for:

• Halos and ringing along strong edges (over-sharpening or Lanczos ringing).
• Waxy skin or “oil painting” texture (over-denoising or an overly aggressive model).
• Repeated or nonsensical patterns in grass, hair or fabric (hallucination).
• Wrong letters in small text.

Summary

Interpolation estimates missing pixels from their neighbours, which keeps it honest but soft. AI super-resolution predicts missing detail from learned patterns, which makes it sharp but occasionally inventive. For most real-world enlargements of 2× and above, a good AI model is the better tool, provided you check the result.

That combination of AI upscaling for detail and classical resampling for exact sizing is the approach IncreaseAI is being built around. For a full step-by-step workflow, see how to increase image resolution without losing quality.