AI Upscaling: 4K Images for AI Video

What 4K means for an AI image

Image models draw at a fixed size, and the size is a setting. 1K means about a million pixels — 1376 × 768 for a 16:9 frame. 4K means 3840 pixels or more on the long side — up to sixteen times the pixels. A model does not stretch a 1K picture to reach 4K: it draws the picture at that size, so every pore, hair and stitch gets pixels of its own.

Two models tested through an external API:

• Nano Banana 2 — the full model, not Lite — has a resolution switch: 0.5K, 1K, 2K and 4K. At 16:9 it returns 5504 × 3072.
• GPT Image 2.5 takes an exact width and height. 3840 × 2160 came back exactly; a request for 5504 × 3072 came back as 3840 × 2128 — 3840 on the long side is the ceiling.

Every image below starts from the same input — Mara’s character sheet, the face close-up next to the full body — and the same prompt: make the sheet again as a new studio shot, 16:9.

Nano Banana 2 · 1K vs 4K

With the same seed Nano Banana 2 drew nearly the same picture at both sizes — the same pose, the same light, the same freckles. So the crops below are the same parts of the same frame; only the number of pixels differs. The 1K crop is enlarged to the size of the 4K one, the way a video model or a big screen enlarges it.

GPT Image 2.5 · 1K vs 4K

This comparison is GPT Image 2.5 Sunburst at quality high — the first setup we tried. GPT Image has no seed, so the two sizes are two different takes of the same prompt: the crops show the same parts of Mara, not the same pixels. After all the tests below, our pick for GPT Image 2.5 is Flare at quality low: 2K for close-ups, 4K for wide shots — see why.

4K is not the same as quality

A size is a request, not a promise. Asked for 3840 × 2160, GPT Image 2.5 delivers every pixel — and fills them with texture it paints rather than draws. The same prompt at 2K comes out cleaner. A detail count cannot tell the two apart: it counts brush strokes and smears as detail. Your eye can. Click any tile to see it full size.

• At 4K the brows become brush strokes — combed parallel lines instead of hairs — at quality low and medium; high is cleaner but still painted in places.
• The fabric smears when you zoom in: the orange shoulders bleed into the green, the nylon gets blotchy. On a wide shot at screen size these patches are hard to see — that is why 4K still wins there (below).
• 2K low looks like a photograph — natural skin, hairs and weave — in 24 seconds, against up to 84 seconds at 4K.
• Nano Banana 2 is different: its 4K is the same picture as its 1K, drawn finer — real detail, not paint.

Close-ups at 2K, wide shots at 4K

The paint only matters where you can see it. In a wide shot the face is a few hundred pixels across, and there 4K is what keeps it a face: the same Flare run gives the full-figure face 538 pixels of width at 4K against 287 at 2K — freckles and the scar survive. Up close the same 4K paints.

Quality high: more time, no more quality

GPT Image 2.5 has a quality setting, and it is tempting to set it to high. On our sheets it did nothing you can see: the same skin, lashes and fabric as medium, with more waiting.

• No visible gain: at 1K the detail count is the same, at 4K high counts even a little less.
• More waiting: high took 10–14 seconds longer.
• Low is enough: our picks — Flare 2K and 4K — are quality low.
• One exception we saw: when GPT is used to upscale an existing image, high stayed closer to the original (drift 2.82 against 5.4 at low, one run each). For generating frames, keep it on low.

Our pick: GPT Image 2.5 Flare at quality low — 2K for close-ups (19 s), 4K for wide shots (35 s). Flare looks the same as Sunburst and is faster: 19 s against 24 at 2K, 35 against 84 at 4K. When a close-up has to be bigger than 2K, upscale it with Magnific — or use Nano Banana 2 at 4K.

An image-to-video model starts from your frame and has to keep the face for the whole shot. We gave MiniMax H3 Max Turbo the face close-up from the Nano Banana 2 sheet — once from the 1K picture, once from the 4K one — with the same prompt, the same seed and 1080P output. Nothing else changed.

The detail the image model drew stays in the video. Freckles, lashes and the scar are sharper from the first frame to the last. Measured on 26 frames (the mean absolute Laplacian of brightness at 1920 px — a plain count of fine edges), the video from the 4K frame has 1.15× the fine detail of the video from the 1K frame on average, and more in every one of them (at least 1.09×). The motion is the same — same seed, same prompt — the face is simply drawn finer.

One continuous shot, no cuts. Extreme close-up of her face; the camera holds still and then slowly pushes in a little closer. She breathes, her eyes move from the lens to the left side of the frame and back to the lens, she blinks once, her lips part slightly. The soft even studio light and the white background stay the same. No text.

Or upscale the 1K frame first

If the frame already exists at 1K, a generative upscaler can draw the missing detail before the video model sees it. We ran the same 1K face through the Magnific upscaler — Ultra Sublime 4×, Portraits, the most faithful 4× setting in the Magnific test below — and gave the result to H3 with the same prompt and seed.

• The upscale reaches native 4K. The video from the upscaled frame has 1.35× the fine detail of the 1K one (at least 1.22× in every frame) — against 1.15× for native 4K.
• It is a different kind of detail. Magnific adds skin texture — fine lines and pores — so it counts higher; native 4K keeps the freckles and lashes crisper. Look at the second slider: neither is a clear winner.
• Timing: 18 seconds for the still; the video was 6 s at 1080P.

Saving a 1K image bigger does not help

The opposite test: one 4K film still of Mara, and the very same still shrunk to 1376 × 768. Same prompt, same seed. Here the videos came out almost the same — 1.04× on average, within the noise. The video model shrinks a big frame to its own working size anyway. What it keeps is the detail that was drawn, not the file size: generate at 4K, or use an upscaler that draws new detail — not a plain resize.

Magnific: upscaler and skin enhancer

When the image already exists at 1K, a generative upscaler redraws it bigger and adds the detail the model never drew. We ran Mara’s character sheet — the face close-up next to the full body, 1200 × 896 — through the Magnific upscaler and skin enhancer with different settings.

The best upscale on the full figure

The same Ultra Sublime 2× Grainy Analog run upscaled the whole sheet: here is the full-figure half next to the original and to the same preset at 4×.

• 2× Grainy Analog works on the full figure too: the small face stays sharp and alive — no pitted skin like the Skin Enhancer gives small faces.
• 4× is not better. At the same viewing size it reads flatter.
• Faithful to the source — including whatever the source got wrong. Fix the character first, then upscale.

Upscale the close-up and the full figure separately

A character sheet puts a face close-up and a full figure into one image. Upscaled in one pass, both halves come out softer than the same close-up and full figure upscaled as images of their own. Below, Mara’s sheet through the best single-pass setting and through 4× Portraits, next to a separate 1K close-up (682 × 768) and a separate 1K full figure (688 × 768) through the same 4× Portraits.

• Separate images win on both halves. The close-up gets pores, lashes and freckles instead of flat, painted skin; the small face of the full figure gets sharp brows, eyes and hair.
• Timing: each half took 15–27 s.
• It repeats. We ran each separate image twice with the same settings: the same face both times, only the finest texture differs.
• The recipe: generate the close-up and the full figure as two images, upscale each with Ultra Sublime 4× · Portraits, and put the sheet together afterwards if you need one.

What the settings do

• Ultra Sublime 2× · Grainy Analog is the best upscale we tested — on the close-up and on the full figure: soft sharpening (5) and film grain (12) give skin, hair and fabric a photographic texture, the face, freckles and scar unchanged. 2400 × 1792.
• On this sheet, 4× downgrades the look. We ran the same 1200 × 896 sheet at 4× with Grainy Analog, Balanced and Portraits: faithful, 4800 × 3584 — and every one reads flatter and more “digital” than 2× Grainy Analog. On a small 1K image of its own (about 700 px) 4× Portraits is the best — see the test above.
• Upscale the close-up and the full figure separately. As images of their own, through 4× Portraits, both came out clearly sharper than the same halves of the sheet in one pass.
• When the focus is on skin, add the Skin Enhancer — Creative — on big faces only (see below).
• Creative Vivid and Wild change the person. Wild redraws the eyes and nose and wipes the freckles; Vivid turns freckles into moles. Fine for an illustration, wrong for a character that must stay the same.
• Creative Subtle, Ultra Photo and Grainy analog stay faithful at 2× when 2400 px is enough.
• Skin Enhancer is for close-ups only. It also enlarges about 4×, and on a big face Creative was the best of the four in this test. On the small face of a full figure every version went wrong — Creative invented pitted, older skin, Faithful changed the eye colour. For a sheet with a full figure, use the Upscaler’s Ultra Sublime instead.
• Upscale the sheet before it goes into a video model: the video model takes its detail from the start frame and cannot add what is not there.

The last step works on the video itself. Topaz Starlight is a generative video upscaler: it redraws every frame bigger and keeps them consistent in time. We ran two of our generations through it.

• Both add real detail — skin, hair, fabric and brick come out finer; the measure counts fine edges at the same 1920 px.
• The two runs are not a head-to-head: different clips, different sizes. Precise 2.6 went up 2.4× in size, Fast 3 stayed at 1080P.
• The measured upscale took 16 minutes. Upscale the final cut, not every take.

The recipe

1. Generate the start frame big. Nano Banana 2: resolution: "4K". GPT Image 2.5 Flare at quality low: 2K (2048 × 1152) for close-ups, 4K (3840 × 2160) for wide shots — up close its 4K paints textures instead of drawing them (see why).
2. Make the character sheet at 4K too. A sheet puts two views in one image, so each view gets half the width; at 4K each half is as wide as a whole 1080p frame or wider.
3. Give the video model the full 4K file. It will scale it down itself; the detail survives.
4. Already have a 1K image? Run it through a generative upscaler, which draws new detail, rather than resizing it. Pick the scale by the size: a small 1K close-up or full figure (about 700 px) → Magnific Ultra Sublime 4× · Portraits; a bigger frame or a sheet (about 1200 px) → Ultra Sublime 2× · Grainy Analog — 4× of a 1200 px image looks worse. In our video test 4× Portraits already reached native-4K detail.
5. Upscaling a character sheet? Upscale the close-up and the full figure as separate images — each through Ultra Sublime 4× · Portraits — and assemble the sheet afterwards. In one pass both halves come out softer (the test).
6. Focus on skin? Add the Magnific Skin Enhancer · Creative — on big faces only; on a full figure it ruins the face.
7. Upscale the final cut, not the takes. Topaz Starlight on the finished video adds detail again; the measured 15-second clip took 16 minutes.

More examples · film stills

The same upscaler on real film frames — small stills, as they come from a screenshot or a web page, run through Magnific Ultra Sublime at 4×. Click a frame for the details; drag across it to compare.

Seven more stills · Portraits or Balanced

Seven small frames, about 1046 × 580, each through Magnific Ultra Sublime 4× twice — preset Portraits and preset Balanced and under 35 seconds a run. The slider shows Portraits against the original; the row under it puts the two presets side by side.

• Portraits is sharper: pores, lashes, single hairs and metal come out, and it adds some texture to backgrounds.
• Balanced is softer and closer to the source: smoother skin, calmer backgrounds.
• Both keep the faces. No one turned into someone else in fourteen runs.

A smaller input · half the pixels

How small can the input be? We took the same still from The Silence of the Lambs, made it twice smaller — 523 × 291, a quarter of the pixels — and ran it through the same Ultra Sublime 4× · Portraits.

• The smaller the input, the more the upscale adds. From a quarter of the pixels it still drew the lashes, irises and lines — at the same viewing size as sharp as the upscale of the full frame.
• Smaller output. The result is 2092 px wide instead of 4176. Take the biggest source you have when the size matters.

The same small still · Magnific, GPT or Nano Banana

The half-size still again, now through our best image models as upscalers — GPT Image 2.5 Flare at 2K and 4K and Nano Banana 2 at 4K, each with the same strict prompt: the very same image at high resolution, nothing changed. Magnific needs no prompt.

• For a real face, Magnific. It is the only one that keeps the person exactly — the others draw a new photograph of someone very close.
• GPT Image 2.5 Flare 2K gives the sharpest photographic look, with a small redraw: colour and lines shift. Good when the frame is yours and a small change is fine.
• Flare 4K and Nano Banana 2 4K change the face — painted texture in one, invented veins, spots and age in the other.