Upscale an image
Planned as a local AI tool: enlarge an image with detail that plain interpolation cannot produce.
Not available yet
This tool is on the roadmap. The page is here so the workflow, formats and related tools are documented — but nothing is processed and no result is produced.
Conventional resampling can make an image bigger but not sharper — it has no information to add, so it spreads the existing pixels more thinly. Super-resolution models predict plausible detail instead, which is why an AI upscale looks different in kind from a bicubic one.
Super-resolution models are small enough to be a realistic local-AI candidate, which is why this is planned as Tier B rather than a cloud call. It ships when on-device inference is dependable.
What you get
2× and 4× targets
The standard super-resolution factors, chosen per image.
On-device inference planned
A small enough model to run in a browser with WebGPU or WebAssembly.
Tiled processing
Large images processed in tiles to keep memory use bounded.
How it works
- 1
Capability check
WebGPU or WebAssembly availability is confirmed before the feature is offered.
- 2
Tiled inference
The image is processed in overlapping tiles inside a Web Worker.
- 3
Reassemble
Tiles are recombined onto the canvas and the result enters the history.
Specifications
- This tool is not implemented yet. Nothing is processed, no image is uploaded, and no result is produced. The page documents the intended workflow and is excluded from search indexing until the feature actually works.
- Planned as Tier B (local AI): ONNX Runtime Web with tiled inference.
Why use KaiConvert
- No upload for local tools
- Works offline once the page is loaded
- Cloud AI is always clearly labelled before you start