Magic Kernel Sharp (Classic, 2013, and 2021) resampling filter for upscaling filter

[ntsc-rs-fan]

Expected behavior of the wanted feature

Magic Kernel Sharp is a resampling algorithm by John Costella. There are Classic, 2013 and 2021 versions so add all three. The website is https://johncostella.com/magic/ and it describes its use. I'm wanting it implemented as I like resampling filters (ones with math) and MPV and MPV Android have never had this one before. Read the website for a more detailed explanation.

https://github.com/Viddeleer/Image-Processing has a C++ implementation of the upscaling filter (resampling methods) and Magic Kernel Sharp would be useful for both the Android and Windows versions of MPV. Currently Magic Kernel Sharp is in no MPV version, even on Android. ImageMagick has added Magic Kernel Sharp due to being a resampling algorithm, and I got the developer of https://github.com/T8RIN/ImageToolbox to add Magic Kernel Sharp too.

Alternative behavior of the wanted feature

No response

Log File

No response

Sample Files

No response

[kasper93]

It provides better results when upscaling due to its algorithm.

It does not. But feel free to prove me otherwise.

[ntsc-rs-fan]

Sorry. Oh I was requesting it because it was never in MPV and I wanted it there. It has other resampling algorithms but not this one.

[HedioKojima]

https://github.com/hooke007/mpv_PlayKit/tree/main/portable_config/shaders/MagicKernel

[kasper93]

https://code.videolan.org/videolan/libplacebo/-/merge_requests/873
#18290

Memes. I didn't know anyone would want that, but if you do it's easy to add. Nothing special though... Use ewa_lanczossharp if you need to have "sharp" in your kernel name.

[ntsc-rs-fan]

Thanks. I wanted Magic Kernel Sharp by John Costella. I wish it was in the main MPV software

[arch1t3cht]

How much does John Costella pay you to mention his name every time you mention the Magic Kernel Sharp by John Costella (trademark, copyright, all rights reserved)?

[ntsc-rs-fan]

No, he doesn't pay me to use his name but I just clarified who created Magic Kernel Sharp here, to not confuse it with other resampling methods. Thanks for adding Magic Kernel Sharp.

…

On Sat, Jul 25, 2026, 10:26 AM arch1t3cht ***@***.***> wrote: *arch1t3cht* left a comment (mpv-player/mpv#18288) <#18288 (comment)> Does John Costella pay you to mention his name every time you mention the *Magic Kernel Sharp by John Costella* (trademark, copyright, all rights reserved)? — Reply to this email directly, view it on GitHub <#18288?email_source=notifications&email_token=B7HXOS3CD3OTTS3IRO6ZVZ35GS7S3A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMBXHA4DINBQHE4KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-5078844098>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/B7HXOS3NO4IJX4OZHMXP6E35GS7S3AVCNFSNUABDKJSXA33TNF2G64TZHM3DEMBRGA4TEO2JONZXKZJ3GQ4TMOJYHEZDAMJXUF3AE> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/B7HXOS2YJ6XC4WMLD5JQ4CL5GS7S3A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMBXHA4DINBQHE4KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KUZTPN52GK4S7NFXXG> and Android <https://github.com/notifications/mobile/android/B7HXOS4UOZYLVBMP35VKA635GS7S3A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMBXHA4DINBQHE4KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>

[kasper93]

How much does John Costella pay you to mention his name every time you mention the Magic Kernel Sharp by John Costella (trademark, copyright, all rights reserved)?

No, but there is precedence for this. https://code.videolan.org/videolan/libplacebo/-/work_items/329 and changes that followed after this report.

[kasper93]

Note that, I did it because it was trivial to plug the kernels, such that people can grab the build and test it for themselves. However, I will not merge this unless there is compelling reason to.

[kasper93]

note that those PSNR values doesn't mean much, check images with your own eyes.

I did some checking, how it works. I got help from friendly robot, to format and generate/organize the images and generate some metrics. Here goes rough summary of the mks kernels compared to some of the already available options in mpv.

• Upscaling: mks2013 has the best kernel PSNR on all three test images under mpv defaults + AR 0.6 - by at most 0.4 dB, at identical GPU cost to lanczos (section 6). The sharp-kernel field spans ~0.6 dB.
• Downscaling: mks2013 aliases more than lanczos (4.90 vs 3.74 RMS) - replica ghosts suppressed, near-Nyquist leakage worse. The un-sharpened magic kernel (quadratic) is the best separable antialiaser measured (1.94, near-EWA): the Sharp step is what costs the antialiasing.
• Ringing: mks2013 has the deepest and widest first ring (14.3 %) but no secondary halo - single-ring structure; with AR 0.6, lanczos rings less (4.6 % vs 5.7 %). Non-negative kernels (quadratic, hermite, bilinear) are ring-free.
• Chroma: mks2013 below lanczos and spline36.
• mks2021: at or below mks2013 on every test, at 1.5x the GPU cost.
• ArtCNN reference: whole-dB swings either way by content type - dwarfs all kernel differences.

All images are vo=gpu-next renders: --no-config --deband=no --dither-depth=no, OSD off, captured via screenshot-to-file ... window. Antiring runs use --scale-antiring=0.6. "sig" = mpv default (sigmoid-upscaling=yes); downscales use --correct-downscaling=yes.

1. Restore tests (degrade -> upscale back -> compare vs reference) - PSNR / SSIM

clown x4 - general-cathexis.com/interpolation canonical pair: input ClownLanc3_0.5X_Box0.25X.png (192x256), reference ClownLanc3_0.5X.png (768x1024):

scaler	PSNR sig	SSIM sig	PSNR sig+AR	SSIM sig+AR
ArtCNN_C4F32_DS	23.359	0.6182	-	-
mks2013	22.929	0.5967	22.957	0.5985
ewa_lanczos4sharpest	22.904	0.5942	22.907	0.5938
lanczos	22.864	0.5896	22.858	0.5909
mks2021	22.863	0.5916	22.850	0.5921
spline36	22.843	0.5887	22.832	0.5895
catmull_rom	22.780	0.5873	22.769	0.5874
ewa_lanczossharp	22.718	0.5771	22.740	0.5805
ewa_lanczos	22.679	0.5740	22.701	0.5775
hermite	22.531	0.5774	22.530	0.5774
bilinear	22.344	0.5633	22.344	0.5633
quadratic	22.015	0.5400	22.015	0.5400

lighthouse crop 512x512 (box /2 -> x2):

scaler	PSNR sig	SSIM sig	PSNR sig+AR	SSIM sig+AR
ewa_lanczos4sharpest	34.388	0.9298	34.448	0.9306
mks2013	34.215	0.9318	34.447	0.9323
lanczos	34.271	0.9291	34.249	0.9285
ewa_lanczossharp	34.111	0.9241	34.139	0.9244
spline36	34.139	0.9276	34.103	0.9268
mks2021	34.138	0.9286	34.097	0.9277
ewa_lanczos	34.019	0.9224	34.051	0.9228
ArtCNN_C4F32_DS	33.811	0.9175	-	-
catmull_rom	33.698	0.9225	33.665	0.9219
hermite	32.366	0.9059	32.363	0.9058
bilinear	32.168	0.8976	32.168	0.8976
quadratic	31.522	0.8828	31.522	0.8828

kumiko 512x512 crop (artoriuz's anime test image, box /2 -> x2):

scaler	PSNR sig	SSIM sig	PSNR sig+AR	SSIM sig+AR
ArtCNN_C4F32_DS	34.649	0.9835	-	-
mks2013	32.271	0.9674	32.613	0.9732
ewa_lanczos4sharpest	32.590	0.9742	32.558	0.9733
lanczos	32.162	0.9679	32.239	0.9715
ewa_lanczossharp	31.868	0.9662	32.070	0.9706
spline36	32.018	0.9682	32.059	0.9707
mks2021	32.011	0.9689	32.039	0.9709
ewa_lanczos	31.765	0.9655	31.959	0.9699
catmull_rom	31.554	0.9674	31.553	0.9684
hermite	30.257	0.9604	30.255	0.9604
bilinear	29.776	0.9564	29.776	0.9564
quadratic	29.309	0.9506	29.309	0.9506

Under sig+AR, mks2013 has the highest kernel PSNR on all three images (22.957 / 34.447 / 32.613) - best outright on clown x4 and kumiko, and within 0.001 dB of ewa_lanczos4sharpest on lighthouse. Without AR the picture is mixed: lanczos leads mks2013 on lighthouse PSNR (34.271 vs 34.215) and ewa_lanczos4sharpest leads on kumiko; mks2013 holds the best kernel SSIM on clown x4 and lighthouse, mks2021 on kumiko. Antiring helps mks2013 (+0.03 to +0.34 dB - it clips the kernel's overshoot errors) and leaves lanczos essentially unchanged. mks2021 trails mks2013 on PSNR in every configuration. The non-interpolating quadratic lands below bilinear on every restore test - the un-sharpened magic kernel is a smoother, not an interpolator. ArtCNN measures +2.5 dB over lanczos on kumiko (its training domain), +0.5 dB on clown x4, and -0.4 dB on the lighthouse photo - larger swings in both directions than the entire sharp-kernel spread.

2. Ringing (step edge x8)

with AR 0.6:

scaler	overshoot	+ AR 0.6	ring width (px @ x8)	2nd halo	taps/axis
bilinear	0.0 %	0.0 %	0	0.0 %	2
quadratic	0.0 %	0.0 %	0	0.0 %	3
hermite	0.0 %	0.0 %	0	0.0 %	2
ewa_lanczos4sharpest	2.9 %	5.7 %	7	1.1 %	polar r4.2
catmull_rom	7.4 %	2.9 %	8	0.0 %	4
mks2021	10.3 %	4.0 %	8	1.7 %	9
spline36	10.9 %	4.6 %	8	1.7 %	6
lanczos	12.0 %	4.6 %	8	2.9 %	6
ewa_lanczos	12.0 %	4.6 %	8	2.9 %	polar r3.2
ewa_lanczossharp	12.0 %	5.1 %	8	2.9 %	polar r3.2
mks2013	14.3 %	5.7 %	11	0.0 %	5

Ring width: extent of the first ring (contiguous deviation at least 0.5 code values) on the dark side of the edge. 2nd halo: amplitude of the brighter-than-plateau band beyond the first ring.

Non-negative kernels - quadratic (the plain magic kernel), hermite, bilinear - do not ring at all; ringing enters MKS entirely through the Sharp step. Among the sharp kernels the structures differ qualitatively: mks2013 has the deepest and widest first ring - its single negative lobe reaches -9/64, about -14%, and spans 1.5 sample units (vs 1.0 for lanczos), hence 11 px vs 8 px at x8 - but no secondary halo whatsoever: the kernel never returns positive after its negative lobe, so the step response is structurally limited to one ring. lanczos, spline36 and the EWA kernels add a 1.7-2.9 % second bright halo from their positive third lobe; mks2021 re-acquires a faint 1.7 % second halo through the echo of its sharpening taps at distance 2; catmull_rom (support 2, no third lobe) is also single-ring, at half the depth. So the trade is one deep, wide ring (mks2013) versus a shallower ring plus visible secondary rings (lanczos-class). With AR 0.6 all sharp kernels collapse into the 3-6 % band, with lanczos (4.6 %) below mks2013 (5.7 %). ewa_lanczos4sharpest measures worse with --scale-antiring=0.6 than without - the option overrides its built-in AR 0.8.

3. Downscale aliasing (zone plate /2.56, dscale=<kernel>)

Outside the center disc the ideal result is flat gray; RMS there = total aliasing (lower is better). No AR, separable antiring is not available.

scaler	RMS (gamma)	RMS (linear)
ewa_lanczos	1.60	6.43
ewa_lanczossharp	1.74	6.57
quadratic	1.94	4.63
ewa_lanczos4sharpest	2.67	7.57
spline36	3.73	7.53
lanczos	3.74	7.74
mks2021	4.49	7.78
catmull_rom	4.80	7.52
bilinear	4.88	5.55
mks2013	4.90	8.79
hermite	6.65	7.93

Two distinct artifact classes are visible. The replica ghosts (off-center bull's-eyes) are visibly suppressed by all magic-family kernels - a consequence of the third-order zeros their $\mathrm{sinc}^3$ spectrum places at multiples of the sampling frequency. The decomposition is the striking part: the un-sharpened quadratic measures 1.94 - best separable kernel, within reach of the polar EWA family, and best of the entire set in linear light - while adding the Sharp step (mks2013) boosts near-Nyquist response so strongly that total leakage rises to 4.90, more than lanczos (3.74) and every other sharp kernel; in linear light it is the worst of the whole set. The $\mathrm{sinc}^3$ antialiasing claim is thus real, but it belongs to the B-spline; the sharpening spends it. hermite - mpv's --dscale default, chosen for being ring-free and cheap - has the weakest stopband in gamma light (6.65), hidden on real content by its blur. Single frames understate all of this; in motion the leakage is shimmer.

4. Real content

Clown x3, wheel crop - gamma:

sigmoid + AR 0.6:

Lighthouse /2.5 linear light, fence:

At realistic ratios the differences are small: mks2013 renders slightly softer than lanczos with a tighter edge halo.

5. Chroma upscaling (cscale, 4:2:0 -> 4:4:4)

artoriuz's aoko art, 512x512 crop -> lossless FFV1 4:2:0, fully tagged (BT.709 matrix, limited range, sRGB transfer, left-sited chroma, subsampled left-sited by swscale to match) -> rendered 1:1 so only --cscale engages -> PSNR vs the RGB original. SSIM omitted: it is luma-dominated and lands at 0.998 for every kernel.

scaler	PSNR
ewa_lanczos4sharpest	37.138
lanczos	36.783
spline36	36.688
mks2021	36.634
ewa_lanczossharp	36.585
mks2013	36.521
ewa_lanczos	36.515
catmull_rom	36.337
hermite	35.410
bilinear	35.398
quadratic	34.593

For chroma, mks2013 measures below lanczos and spline36 - its sharpening does not pay off on the smoother chroma planes - while mks2021 sits just under spline36. ewa_lanczos4sharpest leads.

6. GPU cost

Measured with vo-passes (GPU timer averages) during 60 fps playback of a 480x270 4:4:4 test pattern upscaled x4 to a 1920x1080 window, sigmoid + AR 0.6. Absolute numbers are specific to the test machine's GPU, look at the ratios between numbers.

scaler	scaler passes (us)	whole frame (us)
bilinear	0 (direct sampling)	30
hermite	44	59
quadratic	52	67
catmull_rom	65	80
mks2013	106	121
spline36	106	121
lanczos	106	121
mks2021	161	176
ewa_lanczossharp	308	323
ewa_lanczos	317	333
ewa_lanczos4sharpest	416	432

mks2013 renders at exactly the same cost as lanczos and spline36 despite its nominally lower tap count. libplacebo's separable sampler pads the filter row to an even width, so radius 2.5 rounds up to the same 6-tap rows as radius 3. mks2021 costs ~1.5x lanczos; the polar EWA kernels cost ~3-4x any separable kernel.

Commentary

Quotes are copied verbatim from johncostella.com/magic.

• Rediscovery, not discovery: m(x) is the quadratic B-spline. MKS is a B-spline + discrete sharpener, the classic generalized-interpolation idea.
• "Magic Kernel Sharp is superior to any of the Lanczos kernels" - not reproduced: within 0.4 dB of lanczos either way when upscaling, more total aliasing downscaling, worse chroma.
• "This avoids introducing 'ringing'" - true for non-negative m(x): measured, quadratic rings 0.0 %. The Sharp step re-adds it (14.3 %, highest in the set). What survives is one ring with no secondary halo - and with default-profile antiring, lanczos rings less anyway.
• "These high-order zeros greatly suppress any potential aliasing artifacts...at low frequencies" - reproduced, and the un-sharpened kernel proves how well: quadratic measures 1.94, near-EWA. But the sharpener amplifies near-Nyquist leakage, which dominates: mks2013 lands at 4.90, above lanczos. The antialiasing belongs to the B-spline; the Sharp step spends it.
• "Facebook has used it to resize all its images since 2013" - a CPU story: 5 taps matter in server-side thumbnailing of stills. On GPU the row pads to lanczos width - identical cost (section 6).
• Perspective: the sharp-kernel field spans ~0.6 dB; the first-order differences live elsewhere - EWA (half the aliasing, 3-4x cost) and trained doublers (a full dB either way depending on content).

Bottom line: a competent kernel family whose verifiable contributions are replica-ghost suppression, a single-ring step response, not really the claimed superiority over Lanczos.