Dropvert

AI Image Upscaler

4× resolution upscale, in your browser via Real-ESRGAN.

Drop files anywhere or click to browse

JPEG, PNG, or WebP

WebGPU not detected.

The upscaler will fall back to WASM, which works but is roughly 5–10× slower. For best performance, use Chrome 113+, Edge 113+, or Safari 18+ on macOS / iOS 18. Firefox needs dom.webgpu.enabled in about:config.

Upscale photos and illustrations 4× using Real-ESRGAN AI — entirely in your browser. The model runs on your GPU via WebGPU; nothing is uploaded, and the same image converted on Upscayl Online or ImgUpscaler would otherwise pass through their servers. Your photo never leaves your device.

How it works

4-step walkthrough

  1. 1

    Drop the image

    JPEG, PNG, or WebP. Dropvert decodes it locally and shows the dimensions plus a preview.

  2. 2

    Pick a model variant

    General (~64 MB): best for photographs, portraits, and natural scenes. Anime / illustrations (~17 MB): smaller model, optimized for cartoons, pixel art, and vector-style graphics. The model downloads once and is cached by your browser for future runs.

  3. 3

    Click Upscale 4× and wait

    Real-ESRGAN runs tile-by-tile so even very large images don't exhaust GPU memory. A progress bar shows tile inference and final compositing. Output is always 4× the input dimensions, regardless of image size.

  4. 4

    Download as PNG

    PNG output preserves the upscaled detail without lossy re-encoding. If you need JPEG/WebP, run the result through Compress Image after.

Why use Dropvert

Local-first, free, no upload required

  • Browser-side AI — your photo never leaves your device. Most upscalers require an upload to a server.
  • Powered by Real-ESRGAN, the open-source state-of-the-art upscaler used by Upscayl, Cupscale, and others.
  • WebGPU acceleration when available — typically 5–10× faster than WASM-only browsers.
  • Tile-based inference handles inputs up to 4K and beyond without OOM.
  • Free, no signup, no per-day quota, no watermark on the output.

Frequently asked questions

8 answered

Why does the first run take so long?
The model downloads once on first use — about 64 MB for the General variant. After that, it's cached in the browser's storage and subsequent runs skip the download. The General model fully cached + a 1 MP input takes 30–90 seconds on a recent laptop with WebGPU.
What's the difference between General and Anime models?
Real-ESRGAN ships two variants. The General model is trained on a wide mix of natural images — good for photos, screenshots, scanned documents. The Anime model is trained specifically on illustrations, cartoons, and stylized art; it preserves clean edges and flat color regions better than the General model. As a rule of thumb: photos → General; anything drawn / illustrated → Anime.
Why 4× and not 2× or 8×?
The Real-ESRGAN x4 model is what we ship. 2× is mostly a downscale of the 4× output and doesn't add new detail; 8× would need a larger or different model. If you only need 2×, take the 4× output and resize it down — the result is sharper than running a separate 2× model.
Why is the output always PNG?
Upscaled images often have fine detail that JPEG's lossy compression would visibly destroy. PNG keeps every pixel of the upscaled output exact. If file size matters more than fidelity, run the PNG through Compress Image afterward.
Can I use this on my browser?
WebGPU support: Chrome 113+, Edge 113+, Safari 18+ (macOS / iOS), Firefox Nightly with the flag enabled. The tool falls back to WASM execution on older browsers — works but is 5–10× slower. The page detects your browser and shows a banner if WebGPU isn't available.
How private is this?
Fully. The image stays in your browser tab. The model file downloads from Hugging Face's public CDN (Hugging Face sees the request for the model bytes, but not your image). No analytics on your image, no logging of inputs or outputs.
Are there any limits on input size?
Practically: 4 megapixels (e.g. 2048×2048) is comfortable on most hardware. Beyond that, expect inference times of 2–5 minutes and consider using Resize Image to bring the input down first. The browser will not crash but the tab may become unresponsive.
Can I get a refund-style fallback if it doesn't look good?
AI upscaling improves perceived sharpness but invents detail that wasn't there. For some content (already-noisy photos, low-quality JPEGs with heavy artifacts), the output can look worse than a plain bicubic resize. If you don't like the result, the tool produces a fresh download — just don't use it. Original is unchanged in your file system.

Related tools

4 suggestions

Compress ImageResize ImageRemove BackgroundSVG Trace