Nearest-Exact, Bilinear, Area, Bicubic, or Lanczos? Upscale Methods Compared

Key takeaways

• For upscaling photos and video frames, Lanczos and bicubic keep the most detail. In our 4× test they scored about 5 dB above nearest-exact and 2.5 dB above bilinear.
• Lanczos and bicubic sharpen edges by overshooting them. That's the faint halo you see around high-contrast lines.
• In ComfyUI, "area" gives the same result as nearest-exact when you upscale by a whole number. In FFmpeg, "area" upscales like bilinear.
• For downscaling in ComfyUI, pick lanczos or area. Bilinear and bicubic there skip anti-aliasing and leave strong moiré.
• None of these methods adds detail. For that you need an AI upscaler, which is a different kind of tool.

For upscaling, choose Lanczos or bicubic; they keep the most edge detail, at the cost of slight halos. Use nearest-exact only for pixel art or masks. Bilinear is softer but halo-free. Downscaling is different: in ComfyUI, bilinear and bicubic create moiré, so use lanczos or area.

These names show up in ComfyUI's Upscale Image node, in FFmpeg's scale filter, and in most image editors. In this guide, you'll learn what each method does, what our measurements showed, which one to pick for each job, and where classic resampling stops and AI upscaling starts.

Contents

1. What is an upscale method?
2. The five methods in plain terms
3. Our test: one photo, five methods, two tools
4. Halos: the price of a sharper edge
5. The downscaling trap in ComfyUI
6. Which method should you choose?
7. FFmpeg: the flags that match
8. When a resampling method isn't enough
9. How to compare both in KwaFlux
10. Troubleshooting
11. FAQ

What is an upscale method?

An upscale method, or resampling filter, is the rule a program uses to calculate new pixels from the existing ones when it changes an image's size. It decides how many neighboring pixels to look at and how much each one counts. It can't invent detail that wasn't captured; it only decides how smooth or crisp the in-between pixels look.

ComfyUI's Upscale Image node offers five: nearest-exact, bilinear, area, bicubic, and lanczos. The latent version of the node swaps lanczos for bislerp, a method for latents rather than finished images.

The five methods in plain terms

Sidenote: In ComfyUI, four methods go through PyTorch's interpolate function with its default settings. The lanczos option takes a different path: it converts each image to 8-bit and resizes it with the Pillow library on the CPU. On long frame batches that path can be slower, and it rounds the picture to 8-bit.

Our test: one photo, five methods, two tools

We took a detailed 1920 × 1080 night-city photo, shrank it to a quarter of its width and height (480 × 270), and scaled it back up with each method. Then we compared every result with the original using PSNR, a pixel-accuracy score where higher is closer.

Three other sample photos gave the same order, except that on the softest one bicubic edged out Lanczos at 4× by 0.07 dB. Two results stand out:

• "area" isn't the same method in both tools when you upscale. In ComfyUI it produced exactly the nearest-exact picture. In FFmpeg it produced exactly the bilinear picture. Area averaging only matters when several source pixels fall into one output pixel, which happens when you shrink.
• Lanczos and bicubic are close. The 0.2 to 0.4 dB gap at 4× is hard to see. The bigger jump is from bilinear to either of them.

Halos: the price of a sharper edge

Bicubic and Lanczos look sharper because their weighting curves dip below zero. Next to a hard edge, that pushes the dark side darker and the bright side brighter than either original value. Our test used a step from gray level 40 to 215, scaled 4×:

On a photo you rarely notice this. On text, line art, UI screenshots, and bright lights against a dark sky, it shows up as a thin light or dark outline.

In ComfyUI, bicubic can also push values past pure white or pure black. On our night photo the 4× result ran from -4.6% to 113% of the normal range before it was clipped. Clipping hides it, but very bright areas lose a little gradation.

The downscaling trap in ComfyUI

People use the same Upscale Image node to shrink images, for example to fit a model's input size. That's where the method matters most.

When you shrink an image, fine patterns that are too small for the new size have to be blurred away first. Otherwise they fold into false patterns, called aliasing or moiré. PyTorch's interpolate only does that pre-blur when you ask for it, and ComfyUI doesn't ask. So in ComfyUI, bilinear and bicubic downscaling look at too few pixels.

We shrank a test pattern of fine rings from 1024 to 256 pixels and measured how much false pattern was left, where 0 is ideal and about 90 means full strength:

FFmpeg and Pillow pre-blur automatically, so all their smooth methods do well. In ComfyUI, use lanczos or area to shrink.

Which method should you choose?

Pro tip: If you upscale and then run a sharpening step or an AI model, the method matters less, because the next step reshapes the edges anyway. Pick bicubic as a safe default and spend your time on the step after it.

FFmpeg: the flags that match

FFmpeg's scale filter uses bicubic unless you set flags. In our test, leaving it out produced a file identical to flags=bicubic.

# Upscale to 1080p with Lanczos
ffmpeg -i input.mp4 -vf "scale=1920:1080:flags=lanczos" -c:v libx264 -crf 18 -c:a copy output.mp4

# Keep the aspect ratio: set the height, let FFmpeg pick an even width
ffmpeg -i input.mp4 -vf "scale=-2:2160:flags=lanczos" -c:v libx264 -crf 18 -c:a copy output_2160p.mp4

# Pixel art: nearest neighbor, exact 4x
ffmpeg -i sprite.png -vf "scale=iw*4:ih*4:flags=neighbor" sprite_4x.png

FFmpeg also offers spline, which scored between bicubic and Lanczos in our test (35.39 dB) with slightly smaller halos than Lanczos. For more scaling detail, see how to upscale 720p to 1080p.

When a resampling method isn't enough

Every method above calculates new pixels from the old ones with a fixed formula. None of them can tell a face from a wall or recover texture the camera didn't record. If a 4× result looks soft no matter which method you choose, that's the limit of resampling, not a wrong setting.

AI upscalers work differently. A trained model predicts plausible detail from what it learned about real images, and it can also remove compression blocks and noise along the way. The trade-off is that the added detail is a prediction. AI upscaling vs bicubic and Lanczos compares the two approaches side by side, and 480p to 4K: what AI can recover shows where even AI runs out.

How to compare both in KwaFlux

KwaFlux's comparison view shows a resampled version next to the AI result, so you can judge the difference on your own footage:

1. Import a clip or image and open Quality Repair.
2. Set Output Resolution to the size you need, for example 2× or 4K.
3. Render a 1-, 3-, or 5-second preview. No account is needed.
4. Open Slider compare. The "before" side is your source scaled to the output size; the "after" side is labeled with the AI upscaled resolution.

For photos, the same steps work from the image enhancer. One license covers 23 video modules, 15 photo modules, 4 audio tools, and subtitle and file tools, with no credits or per-minute fees; the Perpetual License lists at $199.99, paid once.

Troubleshooting

Final thoughts

Pick Lanczos or bicubic to enlarge, nearest-exact for pixels that must stay hard, and in ComfyUI pick lanczos or area to shrink. The method changes how the edges look, but it can't create detail. When you need detail that the source doesn't hold, that's a job for AI upscaling.

See AI upscaling next to a classic resize

KwaFlux runs the AI video upscaler and image enhancer on your own GPU. AI processing runs locally on your GPU, and 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: 1, 3, or 5 seconds for video, and the whole picture for images. Signing in with an active paid plan is required only to export, convert, repair a file, or download a file from a link.

Download KwaFlux · See plans