Upscale Images 4× With Browser-Side AI (No Signup)

You inherited a folder of vacation photos from 2012, all 1024×768 — small enough for a slideshow on a 1080p monitor but useless for a 4K print, a poster, or a modern Retina-quality blog header. Or you're building a landing page and the perfect stock photo only comes in 600×400. Or you have one cherished photo of a relative who passed, taken on an early smartphone, and you want to print it at 12×18 inches without it looking like a screenshot of a screenshot.

This is the image upscaling problem, and it's been transformed in the last 5 years by deep-learning models that don't just enlarge pixels (which produces a blurry mess) but synthesise plausible new detail based on what they've learned from millions of training images. The best models in 2026 run in your browser tab in under 30 seconds per image, with results that would have required a $40/month Adobe subscription in 2022.

This guide explains how AI upscaling actually works, when it's magic vs when it produces creepy "hallucinated" detail, and how to use Ai2Done's Super Resolution tool for browser-side 2× / 4× upscaling.

TL;DR

• AI upscaling synthesises new pixels based on what a neural network has learned. It's not magic — it's pattern matching.
• 2× and 4× upscale work well for photos. 8× is the limit before "hallucinated" detail becomes obvious.
• Use Ai2Done Super Resolution for browser-side upscaling with ESRGAN / Real-ESRGAN models running in WebGL.
• Works best on: faces, landscapes, food, animals. Works worst on: text in images (use OCR + re-render instead), pixel art (use a dedicated tool), heavily-compressed JPGs (artifacts get amplified).
• Privacy matters: family photos, ID scans, personal memorabilia — do this locally.

Why this is harder than it looks

"Upscaling" sounds like multiplication: take 800×600, make it 3200×2400, done. The classic algorithms for this — bicubic interpolation (every image editor's default), Lanczos resampling (slightly fancier), nearest-neighbor (for pixel art) — produce mathematically correct outputs that look terrible. The reason: when you double the dimensions, you have 4× as many pixels but the same amount of information. The new pixels are necessarily averaged from neighbors, which produces softness.

AI upscaling asks a different question: "given this 800×600 image, what would a 3200×2400 photo of the same subject look like?" A neural network trained on millions of pairs of (low-res, high-res) images learns to predict plausible high-frequency detail — sharp edges, hair strands, leaf veins, fabric weave — that the math-only algorithms cannot recover because the information genuinely isn't in the source pixels.

The trade-off: AI upscalers can hallucinate. If the model has never seen a similar input, it makes its best guess based on training-data priors. Faces get extra wrinkles or moles that weren't there. Text becomes more legible but with subtly wrong letterforms. Backgrounds gain detail that didn't exist. For artistic and family-photo use, this is usually fine (and often preferable to the original's softness). For forensic or legal use, AI-upscaled images are not the original truth — they're plausible reconstructions.

The state-of-the-art open models in 2026:

• Real-ESRGAN — best general-purpose photo upscaler, excellent on faces, OK on text.
• ESRGAN (the predecessor) — slightly less aggressive, fewer hallucinations.
• SwinIR — slightly better on natural textures but slower.
• Latent Diffusion Super-Resolution (LDSR) — best quality, ~100× slower, GPU-only practically.

For browser-side use we ship Real-ESRGAN compiled to ONNX-Web running on WebGL or WebGPU where supported, falling back to CPU WebAssembly on older devices.

Method 1: Ai2Done Super Resolution (browser-side, no signup)

The Ai2Done Super Resolution tool runs Real-ESRGAN entirely in your browser:

1. Open /tools/super_resolution in any modern browser.
2. Drop your image — JPG, PNG, WebP, HEIC accepted.
3. Pick scale — 2× (faster, ~5 seconds per image on a 2022+ laptop) or 4× (slower, ~15-30 seconds).
4. Pick model — "General photo" (Real-ESRGAN x4plus, the default) for most photos; "Anime / illustration" (Real-ESRGAN anime) for cartoons and illustrations.
5. Click Upscale. A progress bar shows the inference progress; on devices with WebGPU it's near-realtime, on WebGL fallback it's ~10× slower but still usable.
6. Side-by-side preview of original vs upscaled appears. Toggle a slider to compare regions.
7. Download as PNG (default, lossless) or JPG quality 95 (smaller file).

The whole thing runs in your browser tab. Your photos never upload to a server, including ours.

On supported devices (Chrome / Edge on modern desktops with WebGPU enabled): 4× upscale of a 1024×1024 image takes ~3-5 seconds. On WebGL fallback (older devices, Safari): same operation takes ~30-60 seconds. On CPU WASM fallback (no GPU at all): 2-5 minutes per image. We auto-detect the best backend; you can override in Settings.

Method 2: Real-ESRGAN command line (for batch jobs)

For automating upscaling on dozens of images, the official Real-ESRGAN binary is the fastest path:

# Install (one-time, includes the model weights)
pip install realesrgan

# Upscale 4x with the general-purpose model
realesrgan-ncnn-vulkan -i input.jpg -o output.png -s 4 -n realesrgan-x4plus

On a desktop with a discrete GPU (Nvidia GTX 1660 or better, AMD RX 580 or better), this runs ~5× faster than the browser version because it uses the GPU directly via Vulkan instead of going through WebGPU's abstraction layer.

This is the right answer if you're upscaling 100+ photos for a print catalog or restoring an entire family photo archive. Overkill if you're upscaling 3 photos for a blog post.

Method 3: Topaz Gigapixel AI (commercial desktop, the quality benchmark)

For when results matter more than budget, Topaz Gigapixel AI ($99 one-time) is the commercial gold standard. It includes several specialised models — one for low-resolution photos, one for compressed JPEGs, one for art, one for faces — and the UI lets you mix model outputs per image region. Output quality is meaningfully better than Real-ESRGAN on photos with mixed content.

Trade-off: it's a desktop app (3 GB install), requires a beefy GPU for speed, and the per-license cost only makes sense if you upscale frequently. For occasional use, the browser tool is plenty.

How we built the browser upscaler (technical deep-dive)

The Ai2Done Super Resolution tool is built on:

• ONNX Runtime Web for model inference. ONNX is a model-format standard that lets the same model run on WebGPU, WebGL, and CPU WASM without rewriting the inference code.
• Real-ESRGAN x4plus model weights (~67 MB), quantised to fp16 to fit in mid-tier GPU memory. Lazy-loaded only when the user starts an upscale (so the page loads fast).
• WebGPU as the primary inference backend on Chrome/Edge on supported hardware (Mac M1+, Windows with discrete GPU, Android with Adreno 7xx+). Falls back to WebGL on Safari and older Chrome. Falls back to CPU WASM (multi-threaded, SIMD-accelerated) when no GPU is available.
• Tile-based inference for large images. Real-ESRGAN's native input is 256×256; for a 1024×1024 image we tile into 16 inferences, then stitch with overlap blending to avoid seams. This is what lets us upscale a 2048×2048 source on a phone without OOMing.

The interesting design choice: we deliberately do not offer face-restoration as a separate model. Tools like GFPGAN can do amazing things on damaged or low-res faces, but they often produce eerily "smoothed" results that look like the person's twin rather than the person themselves. For most users the standard Real-ESRGAN result is more faithful to the original.

FAQ

Q: Will the upscaled image look real, or "AI-generated"? A: For most photos at 2× upscale, the result is indistinguishable from a higher-resolution photo of the same scene. At 4× the AI's choices become more visible — extra detail in hair, slight smoothing of skin texture, subtle changes to backgrounds. Whether this is good or bad depends on use: for a vacation photo collage it's an improvement; for a forensic or court-admissible image, it's not the original truth.

Q: Can I upscale text in an image? A: Real-ESRGAN tries, but typically produces text that looks sharper but has subtly wrong letterforms. For images where text legibility matters (scanned documents, screenshots, handwritten notes), a much better approach is: (1) run the source through the Image to Text OCR tool to extract the text, then (2) re-render the text at high resolution in a real font. This produces actually-correct text instead of plausible-looking text.

Q: How big can my input image be? A: On modern desktops with discrete GPUs, up to ~4000×4000 input (16 MP) works fine and produces a 16000×16000 (256 MP) output. On phones and integrated GPUs, ~2000×2000 (4 MP) input is the practical limit before VRAM runs out. The tool will warn you if your input is likely to fail.

Q: How long does upscaling take? A: A 1024×1024 → 4096×4096 upscale takes:

• ~3-5 seconds on M1/M2 Mac, modern Windows with discrete GPU (WebGPU)
• ~15-30 seconds on Intel/AMD integrated GPU (WebGL)
• ~2-5 minutes on no-GPU fallback (CPU WASM)

Q: Should I use 2× or 4× scale? A: 2× when the source is already reasonable (say, 1500×1000) and you want a modest sharpness boost. 4× when the source is genuinely small (500×400 or less) and you need it for a larger display. Going beyond 4× in one pass produces visible hallucinations; for extreme upscaling, run 4× → 4× as two passes, but expect the result to diverge from the original.

Q: Why does my 4× upscale of a JPG show weird patterns I didn't see before? A: Real-ESRGAN amplifies whatever's in the source, including JPG compression artifacts (8×8 block boundaries, ringing around edges). If the source has visible compression, run our Image Compress tool first to get a clean PNG/WebP intermediate, then upscale. Better still: use the highest-quality source you have access to.

Try it now

Upscale images 2× or 4× with AI, in your browser, in seconds:

Open the Super Resolution tool →

Drop an image, pick scale, click Upscale. No signup, no upload, no per-image fee.

Related reading

• Restore old family photos with AI (in your browser) — for damaged or faded sources
• Image compression for the web: how much do you really save? — the opposite operation (smaller files for the web)
• Image OCR: extract text from any picture in 100+ languages — for text in images
• Browse the AI image tools hub and image tools

Last updated 2026-06-14. The Super Resolution tool runs 100% in your browser using ONNX Runtime Web — your photos never leave your device. Inference happens on your GPU (or CPU). We never collect, log, or analyse the files you process.