toolgarden.xyz
中文

Image Enhance

Free browser-local Real-ESRGAN x4plus tool to reduce blur, noise, and compression artifacts with original-size, 2x, and 4x output

Enhancement settings

Real-ESRGAN HD restorationUses x4plus to restore blur, noise, and compression texture. PyTorch FP32 validates conversion; the browser runs ONNX FP16.
FP16
Model download32.2 MiBRuntimeWebGPU
Output size

Keeps the original pixel dimensions after 4x AI restoration and high-quality downsampling.

Original size--
Output size--
Supported formats
JPGPNGWebPGIFBMPSVGAVIF

Image decoding, Real-ESRGAN inference, and export all run locally in your browser; the image is not uploaded. First use downloads about 32.2 MiB of model assets, which are normally cached. AI-generated detail is an estimate and must not be used as factual or forensic evidence.

Enhanced result

AI / HD

Upload an image to enhance

Best for low-resolution photos, older images, noise, and JPEG artifacts. Inspect faces, text, and thin lines closely afterward.

About this tool

Image Enhance uses Real-ESRGAN x4plus to reduce soft edges, noise, and compression texture in low-quality photographs. The model first reconstructs the image at high resolution in the browser, then exports at the original size, 2x, or 4x.

The official PyTorch FP32 checkpoint is used only for offline conversion and output validation. The page downloads a verified ONNX FP16 model and runs tiled WebGPU inference locally; the image is never uploaded.

How to use it

  1. Upload the image to restore

    Use the original file when possible instead of a screenshot or repeatedly compressed copy. The model suits photographs, older images, and JPEG artifacts.

  2. Choose an output scale

    Original size improves appearance without enlarging dimensions; 2x suits sharing and displays; 4x creates a higher-pixel asset.

  3. Inspect details and export

    Zoom into faces, text, product texture, and thin lines, then download only after checking for invented detail.

Supported range and limits

Browser model
RealESRGAN_x4plus ONNX FP16, about 32.2 MiB
Conversion reference
Official PyTorch FP32 checkpoint, used only for offline export and numerical validation
Output scales
Original size, 2x, or 4x; the model restores internally at 4x and smaller outputs use high-quality downsampling
Runtime
ONNX Runtime Web + WebGPU with 128px overlapping tiles
Input limits
About 4 MP maximum input and 40 MP maximum output; speed depends on the GPU and image dimensions
Privacy
Image decoding, model inference, and export all run locally in the browser

When you would use it

  • Restoring a low-resolution old photo

    Reduce compression, noise, and soft edges so the image is easier to view on a modern display.

  • Improving a web image

    Process an image affected by resizing or JPEG compression and export at 1x or 2x.

  • Preparing a larger asset

    Use 4x output to increase pixel dimensions, then review it in the final layout or display environment.

What to know before you start

  • The model generates visually plausible estimates; it does not recover information the source never recorded. Do not treat enhanced identity, archival, medical, legal, or forensic images as factual evidence.
  • Low-resolution faces, text, and regular patterns are most likely to be reconstructed incorrectly and must be compared closely with the source.
  • The HD model requires WebGPU. First use downloads about 32.2 MiB of model assets, which the browser will usually cache.

Related concepts

Real-ESRGAN
A blind super-resolution and restoration model for images with unknown real-world degradation.
ONNX FP16
A portable model format using half-precision weights to reduce download and GPU memory while staying close to FP32 output.
tiled inference
Running overlapping image patches separately, then cropping their overlap and stitching them into the final output.

Frequently asked questions

Does Image Enhance upload files?
No. Image decoding, Real-ESRGAN inference, and export all run locally in the browser; files are not uploaded to a server.
Which model does it use?
The browser uses an ONNX FP16 model converted and validated from the official RealESRGAN_x4plus PyTorch FP32 checkpoint. Its download size is about 32.2 MiB.
Why does it only work in some browsers?
The full Real-ESRGAN x4plus model is compute-intensive and needs WebGPU to use the device GPU. A recent desktop Chrome or Edge browser is recommended.
Why does original-size mode still run a 4x model?
x4plus natively creates a restored 4x result. Original-size mode then downsamples it at high quality to improve noise and compression texture without changing pixel dimensions.
Can I use the result to identify document or evidence details?
No. The model can generate plausible but incorrect faces, characters, and texture. Enhanced output is for visual improvement only, not factual or forensic evidence.
How is Image Enhance different from Image Upscale?
Image Upscale offers pixel copying, smoothing, sharpening, and a lightweight AI option for changing dimensions. Image Enhance specifically uses Real-ESRGAN to restore blur, noise, and compression artifacts in real photographs and can preserve the original dimensions.
Can enhancement recover text that was unreadable?
Not reliably. The model estimates plausible edges and texture and may produce a clear-looking but incorrect character, so it must not be used to recover evidence, numbers, or other fact-critical information.
Why is WebGPU required?
Real-ESRGAN x4plus is much more compute-intensive than normal interpolation. WebGPU uses the device GPU for tiled inference; CPU-only WebAssembly would be impractically slow.

Related tools