DEV Community

Cover image for What a browser 'AI enhance' actually returns — measured, with hashes
LeGo build
LeGo build

Posted on Originally published at wangsheng1991.github.io

What a browser 'AI enhance' actually returns — measured, with hashes

#ai #imageprocessing #webdev #tooling

"Enhance" is one of those words that hides at least four different jobs: super-resolution, denoising, re-rendering in place, and re-composing the frame entirely. They produce very different files, and almost nobody shows you which one they ran.

So I measured one. Not by reading a marketing page — by pulling the public example from the command line, keeping both files, and printing their SHA-256. Here is everything that came out of it.

The run

The tool is a browser image enhancer; this example is a small Node script in its public repo that fetches one cached example pair the site already serves to guests. No account, no credits, no uploads, and nothing of yours is sent anywhere:

node bin/enhance-example.mjs --sample sample1 --out ./out
Enter fullscreen mode Exit fullscreen mode

My copy of that run, on 2026-10-30:

catalog: 5033 ms · cached=true
input.jpg     0.15 MiB   812 ms  sha256:19d965fed4699633…
output.webp   0.13 MiB  5377 ms  sha256:229b802a24474dd6…
Enter fullscreen mode Exit fullscreen mode

Three HTTP requests — catalog, input, output. No inference happened on my machine, and no credit was spent. The millisecond figures are network waits, not model time; treat them as nothing more than "these three requests are not slow".

What actually changed

Both files, measured rather than described:

input output
file input.jpg output.webp
bytes 154,319 134,222
dimensions 1200×768 1200×768
SHA-256 19d965fed4699633e9b6faae738fd678592ca5098a47f1535388ef133c2fc28a 229b802a24474dd6c8df6ddac02afb454b7563fe3dfcada76c25dfceac89e708

The same kitchen frame: the slider sweeps between the source JPEG and the re-rendered WebP

The pair this article is about, as a slider: source on one side, re-rendered output on the other, same 1200×768 frame.

Same dimensions in, same dimensions out. That is the whole headline, and it rules out the interpretation most people bring to the word: this is not upscaling. It is the same frame, redrawn.

The site's own FAQ is explicit about the ceiling: the longest output edge is capped at 1536 pixels, target factors are 2 / 4, and the requested scale is reduced to fit that cap. So even for your own uploads there is no path to "4K from a thumbnail" here.

Two caveats worth more than the numbers

1. Sample pairs do not generalise. A second example in the same repo takes a 1670×1152 JPEG and returns a 1024×1024 WebP — different geometry, which means its before/after slider is not pixel-aligned. If you are judging an enhancer, check whether the pair you are shown preserves the frame; a geometry change can flatter any comparison.

2. This is not an NVIDIA DLSS runtime. The name is confusing, so it is worth stating plainly: nothing here installs, patches or replaces an nvngx_dlss* file, and no game is involved. It re-renders a supplied image and hands it back.

Why hashes

A before/after image on a landing page is unfalsifiable: you cannot tell whether the "after" was produced by the tool, by a different model, or by hand. A file hash is checkable years later by anyone who kept the file. That is the whole reason the example ships with compare.html — a single self-contained page with a slider and both hashes, so the pair can be re-verified without trusting the README that describes it.

If you build anything in this space, the cheapest credibility you can add is exactly this: keep the input, keep the output, publish both hashes, and say which process produced them.

Reproducing it 

git clone https://github.com/wangsheng1991/dlss5-visual-enhancer
cd dlss5-visual-enhancer
node bin/enhance-example.mjs --sample sample1 --out ./out
shasum -a 256 out/input.jpg out/output.webp
Enter fullscreen mode Exit fullscreen mode

Node 18+, zero dependencies. Your hashes should match the table above; if they do not, the example was re-generated and the README should say so.

If you want to run your own screenshot through it rather than a cached sample, that is the signed-in flow on the site — one credit per task, and new accounts start with free credits. That is where the 1536-pixel edge cap and 2/4 factors apply.

Evidence

Same picture, same size, drawn again

"Enhance" is usually read as "make the small one big". This step is not that: it re-renders the frame you give it — same camera, same dimensions, with a 1536-pixel longest-edge cap and 2/4 target factors. Your own screenshot takes one credit, and new accounts start with free credits.

Enhance your own screenshot — free credits on signup: https://www.dlss5nvidia.com/visual-enhancer?ref=enhance-actually-returns

Top comments (0)