Faithful image upscaling · up to 16K
ImageUpscaling16K

NVIDIA RTX Video Super Resolution

rtx_vsr_pro

The image you approved, reconstructed through 8K or 16K. No prompt, no seed, no crop, and no generative repaint.

Runs on the Sogni Supernet. Nothing to install.

Progressive detailSame portrait · closer view
The original portrait at the eye and freckled cheek, the widest of the three framings
Original832×1,216
The 8K RTX VSR output at the same eye and cheek, framed closer
8K5,248×7,664
The 16K RTX VSR output at the closest framing, showing lash and skin-pore detail
16K10,504×15,352
Inspect matched 100% crops

About

RTX VSR is the faithful upscaler in Sogni's image toolkit. It enlarges one existing image while preserving its identity, composition, colors, and aspect ratio. It is promptless and deterministic, so use it after the creative work is finished.

Choose 2×, 3×, or 4× for a relative enlargement, 7,680px for the 8K profile, or 15,360px for the 16K maximum. The other edge is calculated automatically and both edges are aligned to eight pixels. Outputs above 7,680px use JPEG.

832×1216original Sogni still
5248×7664actual 8K result
10,504×15,352actual 16K result

Use it from an agent or API

Sogni Chat and the Creative Agent route pure resolution requests to upscale_image. Attach one source image and ask for 8K or 16K, or submit the exact tool arguments in a durable workflow.

sogni-agent -q --upscale ./input.jpg --target-longest-edge 15360 -o ./upscaled-16k.jpg
{
  "id": "upscale",
  "toolName": "upscale_image",
  "arguments": {
    "sourceImageIndex": -1,
    "targetLongestEdge": 15360
  }
}

Use targetLongestEdge: 7680 for 8K. The hosted tool also accepts scale: 2, 3, or 4. See the complete upscaling guide.

Native-pixel evidence

Inspect the actual pixels

Large images look similar when they are fitted to a screen. These matched 100% crops show what RTX VSR reconstructs at output resolution. Start with the source map so you know exactly where each crop came from, then drag each wipe between a high-quality resize and the actual RTX VSR output.

The complete 832 by 1216 source portrait with crop boxes over the lower-right sclera
Complete 832×1,216 source · boxes are drawn at their exact coordinates

Start with the whole image

The evidence comes from the lower-right sclera

The green 128×128 box feeds the 8K comparison. The amber 64×64 box is nested inside it for the 16K comparison. Both stay centered on the vessel-rich sclera beside the tear duct, where ordinary resizing loses the branching lines and wet-lid boundary.

8K · 128×128 source pixels 16K · 64×64 source pixels

8K profile

808×808 output pixels · ~128×128 source region · 6.3× inspection depth

The lower-right sclera from a 128 pixel source region stretched to 808 pixels with Lanczos resizing The corresponding native 808 pixel lower-right sclera crop from the 8K RTX VSR output
Stretched sourceRTX VSR
vessel networktear-duct edgelower-lid boundary
Open the 808×808 crop at 100% →

16K profile

808×808 output pixels · ~64×64 source region · 12.6× inspection depth

A nested 64 pixel region of the lower-right sclera stretched to 808 pixels with Lanczos resizing The corresponding native 808 pixel lower-right sclera crop from the 16K RTX VSR output
Stretched sourceRTX VSR
vessel branchesscleral texturewet-lid edge
Open the 808×808 crop at 100% →
The complete Hello World brush painting with a crop box over the bottom of the second l in Hello
Complete 1,024×1,024 source · exact 128×128 crop marked in green

Hard-edge test

Brush strokes at 8× scale

The marked source region is the bottom of the second “l” in Hello, where the brush lifts and the stroke breaks into separate bristle tracks. The left side stretches those 128×128 pixels to 1,024×1,024; the right side is the corresponding native crop from the 8,192×8,192 RTX VSR output.

tapered terminussplit bristlesdry-brush speckle
Open the 1,024×1,024 crop at 100% →
A 128 pixel crop of the bottom of the painted letter l enlarged eight times with Lanczos resizing The corresponding native crop of the painted letter bottom from the 8K RTX VSR output
Source stretched 8×RTX VSR 8K

Woven-texture test

Cloth weave and stitching

Fabric is where an ordinary enlargement gives itself away. A weave is finer than the pixels describing it, so interpolation has nothing to work from and hands back a smooth gradient. Both halves below are 448×448 pixels cut from the same coordinates: high-quality bicubic resampling on the left, the RTX VSR output at the identical size on the right.

The weave and the stitch line are not sharper in the reconstruction — they are present. Note what did not move: the fold falls the same way, the seam sits in the same place, and the color is unchanged.

weavestitch linefold shadow
Open the 448×448 crop at 100% →
A shoulder of woven cloth enlarged by high-quality bicubic resampling, the weave and stitching smeared into soft tone The same shoulder at the same size from the RTX VSR output, with the weave and the stitch line resolved
Bicubic resizeRTX VSR
16K in the physical world

A 161-megapixel production master

16K is wasted as a fitted browser thumbnail. Its value is the pixel budget left for a large print, a tight client crop, or both. This portrait's actual 16K-profile output is 10,504×15,352—roughly 159 times as many pixels as the 832×1,216 source.

The files themselves

Download the full-size samples

Every crop on this page comes from one render — an 832×1,216 portrait generated on Sogni with Chroma v48 Detail — and the two RTX VSR reconstructions of it. All three files are published in full, so you can judge the pixels in your own viewer rather than ours.

When to use it

Finished image in, faithful larger image out

Composition locked

Framing, identity, edges, and clothing return as they went in. No new objects, no reinterpretation.

Cleaner, not repainted

Built for compressed video, so blocking and ringing are reduced on the way up rather than magnified, and edges sharpen naturally—without inventing a pore.

1→2

Stacks after a creative pass

Want detail the frame never had? Run a creative or detail upscaler first, then RTX VSR to reach print size without a model having further opinions.

16K

Large-format output

Move from screen-sized art to print and display dimensions. See the actual sizes →

Every Sogni surface

Use Sogni Create, Chat, the public Creative Agent, or hosted workflows.

Supernet powered

NVIDIA RTX workers supply the hardware. No local GPU or setup required.

Pricing

Covered by Unlimited, or pay as you go

RTX VSR is a standard Supernet model, not one of the metered frontier partner models. A Sogni Unlimited subscription covers your upscales under fair use, and pay-as-you-go Premium Spark covers them without a subscription.

Premium Spark

from $1.99 250 Spark · one-time pack, no subscription

Pay as you go on the fastest priority queue. Larger packs lower the rate from $0.0080 down to $0.0053 per Spark.

Buy Spark Packs

Free

$0 Free Spark every month · filtered Supernet jobs

Every account gets monthly Spark for Supernet work, so you can run an upscale before committing to a plan.

Get Sogni free
FAQ

NVIDIA RTX VSR on Sogni

What does 16K mean?

16K is the profile name. Sogni fits the source inside a 15,360px longest-edge box, preserves its aspect ratio, and aligns both output edges to eight pixels.

How large can this 16K portrait print?

Its actual 10,504×15,352 file maps to 35.0×51.2 inches at 300 PPI, 43.8×64.0 inches at 240 PPI, or 70.0×102.3 inches at 150 PPI before bleed and trim. Choose the density with the print provider based on medium and viewing distance.

What does an RTX VSR upscale cost?

It is covered by a Sogni Unlimited subscription ($20/mo or $199/yr; Unlimited Pro is $50/mo or $498/yr) under fair use, with no credit counting. Without a subscription it runs on pay-as-you-go Premium Spark, and Spark Packs start at $1.99. The Spark estimate scales with the output size you request and is shown before the job runs. See the pricing breakdown ↑

Is this a generative image edit?

No. RTX VSR is a promptless deterministic reconstruction path. Use Creative Enhance, restoration, or image editing when you want the content repainted or changed.

Why does the exact result vary from 15,360px?

The source aspect ratio is preserved and both dimensions use an 8px alignment. A portrait or landscape result may therefore land just below the profile's longest-edge ceiling.

Where can I inspect the output at full detail?

Use the 100% crop links above to inspect native output pixels. The linked Sogni Engineering article adds more source and upscaled examples, interactive comparisons, and the test methodology.

Make the image bigger. Keep it yours.

Upscale in Sogni Create or ask for it naturally in Sogni Chat.