Akhouri Anmol Kumar

Posted on Oct 16

Free Online Image Tools Are Not Free. Here Are the Receipts.

#webassembly #javascript #privacy #webdev

Description: Resolution caps, "we delete it after 2 hours (trust us)", paywalled buttons, and why your browser can now do all of this by itself.

You know the ritual.

You need one transparent PNG of a product photo. Five minutes, tops. You google
"remove background online", click the first result, upload, and watch the progress bar
do its little dance.

The preview? Gorgeous. Clean edges. You hit download.

And you get a postage stamp. 🖼️

Let's talk about why "free online image tools" is mostly a vibe, not a price tag, and
about the genuinely interesting engineering reason this whole category is overdue for a
rewrite.

🧾 Receipts policy: every limit I quote below comes from the tool's own
documentation or pricing page. No hot takes without sources.

💀 Sin #1: The bait-and-downscale

remove.bg is the poster child, and to their credit they say it out loud in their
own FAQ: the free website version is for low-resolution images, up to 0.25 megapixels.
Full resolution (up to 50 MP) lives behind premium plans and credits.

Do the math with me:

0.25 MP  =  250,000 pixels  ≈  625 × 400

A modern phone photo is 12 MP. That's roughly 48× more pixels than what the
free tier hands back. The preview on screen looks sharp because your screen is small.
The file is where the shrinkage happens.

Fair play: servers cost money, GPUs cost more, and freemium is a legit business
model. But "free" and "free at thumbnail size" are not the same product.

🕵️ Sin #2: Your photos go on a little vacation

Here's the thing most people never think about. For the typical online tool, the
architecture is:

 your photo ──HTTPS──▶ their server ──▶ (processing) ──▶ their server ──▶ you
                         ▲
                 a copy exists here.
                 for "a while".

Take iLoveIMG: the files go to their servers, and they say uploads get deleted
within about two hours. That's a reasonable policy, honestly. But notice what it
actually is: a promise, not a property of the system. You can't verify it. You're
trusting a privacy page.

For a meme? Whatever. For an ID scan, a client's unreleased product shots, your
kid's school photo, a screenshot with a token in it? That's a bigger ask than the
"drag & drop" button makes it look like.

🔒 Sin #3: Paywall roulette

You open Canva, find the Background Remover button right there in the Edit panel,
click it, and get the upgrade prompt. The feature is a Canva Pro thing. You get a
limited free taste, and then it's subscription time.

Again, not evil, just the pattern: the button is free to look at.

🧱 Sin #4: Limits that hit exactly when you need it

TinyPNG is a great compressor and I will defend that. But the free web version
tops out at 20 images per session, 5 MB per image, and 3 conversions. Paid
tiers raise that to 75 MB (Pro) and 150 MB (Ultra).

RAW-ish export from a camera? A big screenshot? Over 5 MB, no entry. 🚪

🔁 Sin #5 (the one nobody counts): the Tab Spiral

Real task: "Make this product photo web-ready, transparent, under 200 KB, as WebP."

Real workflow with single-purpose tools:

Site A  remove background   → upload #1 → download
Site B  resize              → upload #2 → download
Site C  compress            → upload #3 → download
Site D  convert to WebP     → upload #4 → download

That's 4 uploads, 4 privacy policies, 4 free-tier limits, and 3 extra lossy
re-encodes (every JPEG save throws away a bit more detail; it's called generation
loss). You didn't edit an image. You ran a relay race with your pixels. 🏃

🤔 So why does it work like this?

Historically: the browser couldn't do the heavy part. Segmentation models needed
GPUs and big runtimes, so the model lived on a server, and your photo had to travel
to it. Server costs → limits → upsell → "download HD for 1 credit."

That excuse has been quietly expiring for a few years:

Then	Now
JS is too slow for ML	WebAssembly runs near-native code in the tab
No runtime for neural nets	ONNX Runtime Web runs real models in-browser
Models are 500 MB+	Good segmentation/inpainting models ship in the 5 – 180 MB range
Can't cache big files	Cache Storage keeps a model after the first download

Meaning the "server" in the diagram above is now… optional:

 your photo ──▶ [ your browser tab: WASM + model ] ──▶ you
                         ▲
                 nothing leaves.
                 there is no "their server" step.

Don't trust me. Verify it. 🔍

Any tool that claims "100% client-side" should survive this 20-second test:

Open DevTools → Network tab.
Clear it. Load an image and run the tool.
Look for any request that carries your image bytes.

A real on-device tool shows no image upload. (You'll see one-time downloads of the
model files, because the AI has to come from somewhere, but those flow to you,
not from you.) Try this on whatever you use today. It's genuinely eye-opening.

Credit where due: Squoosh (from the Chrome team) has been proving in-browser
image work is viable for years. The tech is not exotic anymore. It's just that
nobody bundled the whole workflow in one place.

🧪 The nerdy bit: why most cutouts look "off" (even the AI ones)

This is the part I find fun, and it's where "just run a model" stops being enough.

Problem A: the mask is tiny

Segmentation models run at a fixed size, commonly around 1024×1024 (lightweight
ones: 320×320). Your photo is bigger. On a 4032×3024 shot, one pixel of the mask
covers roughly a 4×3 block of real pixels.

Stretch that mask up naively and you get the classic crime scene: soft, blobby edges,
hair turned into a helmet. 🪖

Fix: use the photo itself as a guide. A guided filter (He et al.) snaps the
blurry mask onto the real edges in the full-res image:

// idea in one breath:
// alpha  = upsample(modelMask)                // blurry but in the right place
// guide  = luminance(photo)                   // sharp, knows where edges really are
// alpha' = guidedFilter(guide, alpha, r, eps) // edges now follow the photo
// then clamp alpha' to a narrow band around the model's own edge
// so it can sharpen the boundary, but never invent or delete regions

Problem B: the halo 🌈

Edge pixels are a mix of subject and old background. Cut the subject out of a green
screen-ish photo and drop it on white: green fringe. Every time.

Fix: estimate the true foreground colour at edge pixels (Blur-Fusion foreground
estimation, Forte & Pitié) and use that instead of the contaminated original. Interior
pixels stay byte-for-byte original; only the fringe gets repaired.

Problem C: memory is the real boss fight 💣

A 12 MP RGBA buffer is 48 MB. (12,000,000 × 4 bytes.) Now count how many full-size
copies a naive "remove object" pipeline makes: source, mask, blurred mask, model
output, blend buffer... phones start sweating at ~500 MB and then the OS just ends
your tab.

Fix: never process the whole photo. Crop the region around what you're editing
(plus context), run the model on that, blend it back, and leave every other pixel
untouched.

whole-photo pipeline:  cost ∝ size of the photo   → 💥 on phones
region pipeline:       cost ∝ size of the edit    → ✅ boring, which is the goal

Good engineering is making the scary thing boring.

😅 Okay, disclosure time

Everything above is the checklist I kept running into, until I got annoyed enough to
build against it. I'm the dev behind APIC Web,
a browser-based image toolkit, so yes, this is where I tell you about it. But you've
read this far, so let me at least be straight about what it is and isn't.

What it is: everything in the "Tab Spiral" in a single tab.

Tool	What it does
✂️ Remove Background	On-device AI cut-out (IS-Net from the DIS research; Light/Fast/Max tiers), guided-filter edge refinement + colour-fringe removal, keeps your original resolution
🧽 Remove Object	Paint over it, AI fills it in (MI-GAN, Picsart research, MIT-licensed); only the region around your brush is processed
📐 Resize	Clean multi-step downscaling
🔄 Convert	PNG, JPEG, WebP, BMP, ICO, PDF
🗜️ Compress	Type a target size in KB; it binary-searches the highest quality that fits
🎛️ Edit + Advanced	Filters, light/colour controls, crop, undo/redo

Your image is processed in your browser, and the Network-tab test above is the
proof, not my word.

The honest tradeoffs (because you'd find them in 5 minutes anyway)

🔑 Sign-in: the tool panels sit behind a free Google sign-in. Guests can I store account basics (name, email), never your images.
⬇️ First-run download: the AI models download once (~5 MB Light, 44 MB Fast, 176 MB Max for Remove BG; ~28 MB for Remove Object) from a public model host, then get cached in your browser.
📱 Phones play by phone rules: weaker devices default to the lightweight model, and huge photos (above roughly 12 MP) get downsized on phones only, specifically so the tab doesn't crash. Desktop is never capped.
🧠 It's not Photoshop. No layers, no templates, no brand kits. It's built for the 90% of "I just need this one image fixed" moments.

The scorecard

	Typical free tool	APIC Web
Free output resolution	often capped	original resolution
Where your photo goes	uploaded to a server	stays in your tab
Verify it yourself?	trust the privacy page	Network tab
Background removal	paywall / credits	included
Workflow	4 sites, 4 uploads	one tab
Re-encodes along the way	many	one export

🚀 Try to break it

I'd rather you roast it than ignore it:

👉 akhouri-anmol-kumar.github.io/APIC-web

Throw your worst photo at it: messy hair, a glass bottle, a low-contrast product,
a 12 MP phone pic on a potato Android. Open the Network tab while you do it. If it
fails, tell me in the comments (the phone-crash fix I mentioned came straight
from user complaints, so I take them seriously).

And if you've got your own "free tool betrayed me" story, drop it below. I'm
collecting them. 👇