Free AI Video Upscaler Test: SeedVR2 at 1080p, 2K and 4K

SeedVR2 is a free, open-source AI video upscaler. We measured 1080p, 2K and 4K with spectra, thousands of edges and a shrink-back test. 2K adds real detail. 4K mostly adds pixels.

Published: 2026-11-02
Updated: 2026-11-02

SeedVR2 is an open-source AI video upscaler under Apache 2.0. The model is free, with no watermark and no usage limit. You need a GPU.

Our first reaction to the output was two words: "Not bad."

How not bad? Results first, then we measure.

Upscale to	Size	Time per second of video	System RAM
720p	1248×720	11 s	32 GiB
1080p (1K)	1890×1080	26 s	62 GiB
1440p (2K)	2520×1440	48 s	57 GiB (7 s run)
4K	3780×2160	110 s	59 GiB (3.5 s run)

Our pick: a 768 source, 7B INT8, upscaled to 1080p, with lab color correction. The rest of this article is why.

768 upscaled to 1080p. From left: original, 3B INT8, 7B INT8. 492×600 crops at 1:1.

How long does it take to upscale to 1080p, 2K and 4K?

Double the pixels, double the time. From 1080p up, 80% of that time goes to the VAE and only 7 to 8% to the model's own sampling, which is why switching precision later barely moves the total.

Stacked bars of time per second of video from 720p to 4K, split into VAE decode, VAE encode, sampling and other: 11 seconds at 720p, 26 at 1080p, 48 at 1440p and 110 at 4K, with system RAM of 32, 62, 57 and 59 GiB — RTX 5090, 3B INT8. 720p comes from the 480 source, the others from 768.

We didn't run 14 seconds at 2K or 4K, only the opening seconds. Scaled up, memory would hit 115 and 235 GiB. Our machine stops at 93 GB, so you'd split the clip, upscale the parts and join them.

At 4K, one second of video takes nearly two minutes. Is the detail you get for that time real?

Does AI upscaling add real detail, or just pixels?

There's a standard way to check: the frequency spectrum.

A 768 source can only hold so much detail. At 1080p that limit sits at 0.356 cycles per pixel, the yellow box below. Plain upscaling (lanczos) stretches the pixels and stays inside the box. Anything outside it wasn't in the source.

2D spectra for 768 and 480 upscaled to 1080p, three panels each: lanczos original, 3B and 7B. Lanczos stays almost entirely inside the yellow box, while 3B and 7B spread gray energy outside it. Outside the box: 0.006%, 0.23% and 0.14% for 768, then 0.013%, 0.48% and 0.38% for 480 — 2D spectrum, with low frequencies at the center and fine detail toward the edge. The yellow box is the source limit. 9 frames averaged.

Share of all energy outside the box: lanczos 0.006%, 3B 0.23%, 7B 0.14%. The 480 source has its box at 0.222, and the shares are 0.013%, 0.48% and 0.38%.

Those are small numbers, because nearly all of a picture's energy sits at low frequencies. The gap isn't small: 21 to 37 times.

The curve makes it plainer. Past the dashed line the source falls away. 3B and 7B carry on.

Radial power curves for 768 and 480 upscaled to 1080p. Dashed lines mark the source limits at 0.356 and 0.222. Past them the lanczos original drops fast, while 3B and 7B decline slowly, with 3B slightly above 7B — The same spectra averaged in rings. The dashed line is the source limit.

Beyond the limit, that's 14 to 16 dB higher on the 768 source and 21 to 22 dB on the 480. The softer the source, the more the model adds.

So yes, it's real. It also didn't come from the source. The model made it. Making 1080p detail doesn't mean it can make 4K detail.

Are 2K and 4K worth upscaling to?

Put 1080p, 2K and 4K on one axis: the same first 3.5 seconds, the same area of the picture, frequency in cycles per picture height.

Then pick a ruler. Effective lines are twice the highest frequency still above −57 dB. That's the level the 768 source has at its own limit, so the source measures exactly 768 lines.

1080p measures 1080: it runs out of pixels before it runs out of detail. 2K measures 1424 out of 1440. 4K measures 1329.

A 2160-line picture with 1329 lines of detail. Fewer than 2K.

Power curves for the 768 original, 1080p, 1440p and 4K on one frequency axis, with a table. Effective lines are 768, 1080, 1424 and 1329. Energy above the source limit of 384 is 0.19, 5.20, 5.64 and 5.69 — 3B, 8 frames from the first 3.5 seconds, the same picture area. The −57 dB threshold is the source's level at its own limit.

Loosen the threshold to −60 dB and 4K reaches 1573, still nowhere near 2160. The totals agree. Energy above the source limit is 5.20 at 1080p, 5.64 at 2K and 5.69 at 4K, so going from 2K to 4K adds 1%. The model isn't making more detail at 4K. It's spreading about the same amount higher up.

2K is different. Above 540 cycles, where 1080p runs out of room, 2K has nearly three times the energy. That part is new.

Your eyes say the same. 2K adds a little hair. In this frame the 4K ear edge also picks up a red fringe.

Close-ups of the ear in the 768 original and its 1080p, 1440p and 4K outputs. 1440p shows slightly more hair than 1080p, and the 4K ear edge shows a red color fringe — Frame 60, 3B INT8. The same area at the same display size. This is one area.

4K takes four times as long as 1080p, while 2K takes under twice as long. Stop at 1080p if it's enough, and go to 2K if you want more. We wouldn't run 4K.

More detail should mean a sharper picture. We got this one wrong the first time.

Does upscaling sharpen edges? Is there a halo?

We started by measuring one edge, on an eyelid. It came out 3× steeper.

Then we measured 9,258. Nine frames, strong edges with flat sides in the source, one at a time.

Median 10–90% rise: source 6.6 pixels, 3B 5.9, 7B 6.2. That's 11% and 6% narrower.

The 3× belonged to that one edge.

Median edge curves with interquartile bands for 768 and 480 upscaled to 1080p. The three lines almost overlap. For 768 the rise is 6.6, 5.9 and 6.2 pixels and P90 overshoot is 11.4%, 22.6% and 16.1%. For 480 the rise is 7.9, 7.2 and 7.2 pixels and P90 overshoot is 20.6%, 34.4% and 29.3% — 19,170 edges across both clips, each aligned on its midpoint and normalized to its own plateaus. Bands are the interquartile range.

The halo lives in the tail. The median edge has almost no overshoot. The worst tenth does: 11% in the 768 source, 23% with 3B, 16% with 7B, and on 480 it's 21%, 34% and 29%. The source already had some, and on 768, 3B doubles it.

The official model card warns about this: on lightly degraded AI video the model tends to generate too much detail and can oversharpen. AI video is exactly what we tested.

So edges barely move, yet energy outside the box is up dozens of times. Most of it has to be texture. And texture is where a model can get things wrong.

Can you trust the added detail?

The check is simple. Shrink the output back to source size and compare. If the model only added detail and left the source alone, the two should nearly match.

They don't nearly match. On 768, 7B scores 34.1 dB PSNR and 0.954 SSIM, 3B scores 32.0 dB and 0.923, and plain upscaling shrunk back scores 58 dB. The model changes things inside the box too.

What it changes is fine detail. Blur the difference and the score climbs to 43 to 46 dB. The difference map agrees: edges and texture are dark, flat areas are white. Shapes, clothes and positions stay put. The texture is new.

Frame 240 of the 768 original beside the difference with 7B shrunk back to source size, magnified 8 times and concentrated on edges and texture. A table lists PSNR and SSIM: 768 7B 34.1 dB and 0.954, 3B 32.0 dB and 0.923; 480 7B 36.0 dB and 0.959, 3B 34.2 dB and 0.933; lanczos round trip 58.0 dB — Averaged over 112 frames. The right panel is magnified 8×. Darker means a bigger difference.

The same ruler answers the 4K question from the other side. On 3B's first 3.5 seconds, 1080p scores 31.7 dB, 2K 33.6, 4K 35.7. 4K's generated detail sits at higher frequencies, so it vanishes when you shrink it, and what's left looks most like the source.

Two sources, three numbers, and 7B is closer every time. Zoom in and you can see why.

3B or 7B?

3B adds orange-brown scratches to the robe that the source doesn't have. 7B keeps the fabric close to the source.

The original, 3B INT8 and 7B INT8 on the toad's eye and robe. 3B adds small orange-brown scratches and woven texture to the robe, while 7B stays close to the original fabric — 768 to 1080p, frame 240. 716×400 crops at 1:1.

If you want crisp, take 3B and accept that it makes things up. We use 7B.

If the texture is generated, is it spread evenly or piled into some brightness range?

Do shadows and highlights survive?

Every band gains detail, reaching 1.4× to 2.9× the source. In the shadows, 3B reaches 1.7 to 1.9× and 7B 1.4 to 1.6×.

Bar charts of high-frequency detail in five luma bands for 768 and 480 upscaled to 1080p. 3B and 7B beat the original in every band, by 1.4 to 2.9 times. 7B beats 3B in the highlights, and 3B is usually higher elsewhere — 9 frames. Detail is the RMS of each pixel minus a Gaussian blur of 1.5 pixels, sorted by the original's local luma.

The 2.9× is in the highlights, and there 7B beats 3B. That's one shot, not the whole clip. The window in the second shot is blurred in the source, and the model sharpens it. In shots one and three, 3B still adds more. If you want that background soft, check this yourself.

The tone range itself holds. Mean luma drops about 1 level, the 99th percentile moves 1 level at most, and clipping tops out at 0.017%.

Luma waveforms and histograms for the original, 3B and 7B. The three waveforms look nearly the same and the histograms almost overlap. Means are 79.3, 78.1 and 78.1, p99 is 217 for all three, and clipping is 0%, 0.008% and 0.017% — 768 to 1080p. Waveforms are frame 240. The histogram and figures average 8 frames across the clip.

Single frames pass. But nobody watches video one frame at a time.

Does upscaled video flicker?

The model works chunk by chunk, about 2 seconds each in our setup. Chunks leave seams.

Lay one column of pixels out over time. The lines don't break, but the texture changes at every seam. The source has nothing like it.

Magnified time slices of the original, 3B with no overlap and 3B with overlap 2. Horizontal fine lines continue unbroken, and in 3B the texture visibly changes at the black seam ticks, which does not happen in the original — 480 to 1080p, first shot. One column of 100 pixels over time. Black ticks mark chunk seams.

In numbers: still areas of the source change 0.13 levels per frame, 3B changes 0.37 and 7B changes 0.30. That's 2 to 3 times as much, and still tiny. In the three frames from a seam it doubles again, to 0.56 for 3B and 0.50 for 7B.

Line chart of still-area change per frame for 768 to 1080p. 3B and 7B sit above the original and rise at each seam tick. Two charts below align and average the seams: the three frames from a seam reach 0.56 for 3B and 0.50 for 7B, about double the frames before, and overlap 2 lowers the 480 clip from 0.55 to 0.46 — Still areas are pixels where the original changes under 0.6 levels per frame. Gaps are shot cuts.

Overlap 2 helps a little: from 0.55 to 0.46 on the 480 clip, for about 8 more seconds across the clip. We leave it on.

Brightness is stable. What about color?

Does color shift?

With no correction, saturation goes up 13%. lab stays within 1.5%. Neither gets worse over the clip.

Vectorscopes of the original, lab correction and no correction. No correction spreads wider, with 12.9% more saturation, while lab stays close to the original at 1.5% less — 480 to 1080p, 3B INT8, frame 240. The closer the shape is to the original, the closer the hue and saturation.

lab adds about 14 seconds on a 14-second clip. The ComfyUI template ships with none, so change it.

Two choices left: precision and source.

Which precision should you use?

INT8. For both models it sits top left: fast, and about 45 dB PSNR against FP16. NVFP4 is the fastest and only reaches 37 to 40 dB.

Precision only touches sampling, not the VAE. From 768 to 1080p, fastest and slowest differ by 19 seconds with 3B and 49 with 7B.

Scatter plot of speed against quality for four precisions of 3B and 7B. INT8 sits at the top left, NVFP4 is fastest and lowest, and dashed lines mark FP16 sampling time — Sampling speed across, closeness to FP16 up.

Should you start from 480 or 768?

Time is about the same. The 480 output has more energy outside the box, and it also grows wrinkles behind the ear that the source doesn't have. The 768 output is much cleaner.

Close-ups of the ear in the 480 and 768 originals and their 1080p outputs. The 480 output adds fine wrinkles behind the ear, while the 768 output keeps a clean ear shape — Frame 60 at 3×, 3B INT8. These are separate takes, so compare each only with its own original.

The less the source gives it, the more the model guesses.

The settings we use

"Not bad" has numbers now.

Setting	Choice
Model	7B (3B for more sharpness, and more made-up detail)
Precision	INT8
Target	1080p (2K for more detail)
Color correction	`lab`
Chunk overlap	2
VAE tile	768 is 11% faster than 512. 1024 ran out of VRAM on a 32 GB card

How we tested

Two 14-second clips generated with Pika H3 Flex, 832×480 and 1344×768, at 24 fps and 336 frames. They are separate takes of one scene, so each is compared only with its own original. Our H3 step comparison used the same scene in a separate test.
RTX 5090 32GB, ComfyUI 0.35.0 built-in nodes, models repackaged by Comfy-Org, a fixed seed and H.264 at CRF 16, for 29 outputs.
2K ran only the first 7 seconds and 4K the first 3.5, with shorter chunks than 1080p, and only with 3B. Memory for 14 seconds at those sizes is an estimate.
The 1080p runs measured 62 to 84 GiB of system RAM, and the table shows the lowest. Peak VRAM was about 19 to 25 GiB with INT8. That is not a minimum, and we did not test smaller cards.
We did not test 7B with tile 768, or specifically test face close-ups, subtitles or large motion.