Choose a low-resolution image
Import a photograph, product image, screenshot or old photo; source pixels are read only on this device.
ImgIng / AI image upscaler
REAL-HAT-GAN · LOCAL SUPER-RESOLUTIONUpscale images locally with Real-HAT-GAN and five enhancement workflows
Each real model downloads on demand, passes byte and SHA-256 verification and is reused from browser cache. A dedicated Worker uses model-specific overlapping tiles and releases inference memory when processing ends or the task is cleared.
Official Real-HAT-GAN SharperLocal 2x / 4x upscaleModel-specific contextual tilingNo image upload
Capability quick facts
These facts describe the current product, not unshipped roadmap work.
- Processing
- Dedicated browser-local Worker; the image never uploads
- Models
- Real-ESRGAN / SwinIR / official Real-HAT-GAN Sharper with fixed bytes and SHA-256
- Scale
- 2x / 4x with model-specific contextual tiling for large images
- Memory
- Tensors are released per tile; completion, replacement or exit terminates inference
Reviewed by the DataDance product and engineering team · Published · Updated
Finish in three steps
Confirm important settings before export; the processing location is always disclosed.
Choose one of five workflows
Use Fast, Fidelity, Complete old-photo restoration, Natural dual-model refinement or high-order Real-HAT-GAN refinement, then choose 2x or 4x.
Compare and save
Drag the divider to inspect detail, then export WebP, PNG or JPG. Mobile limits final pixels to reduce memory-related browser exits.
This is not ordinary interpolation
Browser interpolation only spreads existing pixels. ImgIng runs verified Real-ESRGAN, SwinIR and official Real-HAT-GAN models to recover visible structure from blur, noise and compression damage before producing the requested dimensions. Model bytes and SHA-256 are pinned; failed verification never reaches inference.
Which of the five workflows should I use?
Fast enhance prioritises speed. Fidelity upscale protects text, architecture and source structure. Complete old-photo restoration rebuilds full-image texture and faded depth. Natural dual-model refinement combines a photographic base with a detail pass for continuous subject and background texture. High-order natural refinement runs the official Real-HAT-GAN Sharper with a larger contextual crop on capable desktop browsers.
How are large images and memory handled?
A dedicated Worker uses each model’s own window or tile-padding strategy and writes only the reliable centre of every prediction into the final image. Input and output tensors and temporary canvases are released after each tile. Completion, stop, replacement, clear and close all terminate the Worker; keeping model bytes in browser disk cache is not the same as keeping inference memory alive.
Super-resolution predicts some visual detail. Use the result for visual enhancement, not identity documents, forensics, historical evidence or OCR source verification; retain and compare the original for those tasks.
Updated 2026-09-01 · Live capability detection inside the tool is authoritative
Frequently asked questions
These visible answers match the current product behaviour and structured data.
Does AI upscale upload my image?
No. Model files download to the browser, while decoding, tiled inference, comparison and encoding remain on the current device.
Should I choose 2x or 4x?
Prefer 2x on mobile and for already large sources. Choose 4x for small inputs and capable desktop hardware. The tool blocks output beyond the current device stability limit.
Can old-photo enhancement change a person’s identity?
The current mode prioritises the original and does not use an aggressive face-reconstruction model. Truly lost detail cannot be proven to have been restored, so the interface states that boundary.
Explore ImgIng
Each task page documents real settings, limits and format advice—not keyword-swapped duplicates.