AI upscale and denoise
Enlarge a small image with neural super-resolution and optional denoising. Use AI upscale and denoise locally with practical instructions, output settings and format limits.
AI upscale and denoise
Browser super-resolution uses a trained 3× sub-pixel CNN on luminance, then resamples to the requested 2× or 4× size. Denoise and contrast use local image filters, not the desktop FSRCNN/NLM/CLAHE algorithms.
- Enlarge a small product photo
- Prepare a larger copy of a compact illustration
- Try denoising before neural enlargement
How to use AI upscale and denoise
Choose a small source
Upload one supported still image of no more than one million pixels. Keep the original available so you can judge whether the generated texture is useful.
Set enlargement and denoising
Choose two-times or four-times scaling. Set denoise strength to zero to disable it, or increase it cautiously when noise competes with useful fine detail.
Run the local model
Enable local contrast enhancement if needed and run Create output. The CPU model enlarges each dimension by the selected factor and produces a PNG.
Compare actual detail
Download the PNG and compare edges, faces and textures at similar viewing sizes. Check for artificial patterns or smoothing before using the image in a layout.
Understand the settings
- Upscale factor
- Choose upscale factor according to the result you need. Available choices: 2×; 4×. The initial selection is 2×.
- Denoise strength (0 = off)
- Denoise strength (0 = off) accepts values from 0 to 20, using whole numbers. The initial value is 3. Change it for a specific output requirement, then check the result before processing more files.
- Improve local contrast
- Turn “Improve local contrast” on only when it matches your intended result. The initial state is off. Read the workflow and output limitations below when this changes how the original content is processed.
A small image for a larger mockup
An 800 by 600 product photo needs a larger mockup copy. Select two-times enlargement and light denoising, then create the PNG. The output is 1600 by 1200 pixels; inspect product lettering and textured surfaces before accepting the enlargement.
Check your result
Browser super-resolution uses a trained 3× sub-pixel CNN on luminance, then resamples to the requested 2× or 4× size. Denoise and contrast use local image filters, not the desktop FSRCNN/NLM/CLAHE algorithms.
What to know before you start
- Browser super-resolution uses a trained 3× sub-pixel CNN on luminance, then resamples to the requested 2× or 4× size. Denoise and contrast use local image filters, not the desktop FSRCNN/NLM/CLAHE algorithms.
- Files are processed on this device. Browser memory and engine-specific limits apply. Save results before reloading.
- Browser super-resolution uses a trained 3× sub-pixel CNN on luminance, then resamples to the requested 2× or 4× size. Denoise and contrast use local image filters, not the desktop FSRCNN/NLM/CLAHE algorithms.
- Upload one file with a combined size of at most 100 MB. The completed output must fit within 200 MB.
- Files are checked by the local engine; a matching filename alone is not proof of valid contents.
- Selections and completed jobs stay available while you navigate inside this tab. Use My files to revisit results. Reloading or closing the tab clears the workspace; download results you want to keep first.
AI upscale and denoise: common questions
What do two-times and four-times mean?
The factor applies to both width and height. Four-times scaling therefore creates sixteen times as many pixels. A one-million-pixel source can produce a sixteen-million-pixel result, which is the maximum output bound for this workflow.
Does the model recover original detail?
The model estimates an enlarged image from the source pixels. It can improve perceived edges, but it cannot establish details that were never recorded. Check text, faces and fine patterns instead of treating generated texture as original information.
When should denoising be disabled?
Set denoising to zero when the source is already clean or when fine texture is more important than smoothing. Stronger denoising can remove grain along with useful details, especially before a large enlargement factor.
Why can I not upload a large photo?
The local CPU super-resolution workflow accepts at most one million input pixels and sixteen million output pixels. These model-specific bounds are smaller than the standard still-image limit and help keep processing and memory requirements bounded.
Does AI upscale and denoise upload files to another service?
Processing runs in this browser, on your device. Files and pasted input are not uploaded to a conversion server. The website and required engine assets must load first; large operations depend on available device memory.
Can I continue after changing pages?
Selections and completed jobs stay available while you navigate inside this tab. Use My files to revisit results. Reloading or closing the tab clears the workspace; download results you want to keep first.