Image Upscaler
Enlarge photos and illustrations 2× or 4× on your device. Process selected originals, compare actual saved pixels and download PNG, JPEG or WebP.
Choose JPEG/PNG originals, choose 2× or 4× and enlarge selected images. Expected dimensions appear before the model downloads. Save individual results or make a ZIP.
Up to 8 originals,8 MiB each;2 million pixels/2048 px per source for 2×. For 4×, at most 500000 source pixels. Images stay on this device. A current desktop browser with roughly 768 MiB free memory is recommended.
Photo uses a fixed learned luminance model;4× runs two 2× passes and may amplify artifacts. Suggested detail is not recovered evidence.
Advanced settings
Choose static JPEG/PNG originals. Model assets download only when learned processing starts.
How to use Image Upscaler
Choose originals and an enlargement
Add static JPEG/PNG images or try two local examples. Choose 2× or 4× width and height. A 320× 240 photo becomes 640× 480 at 2× or 1280× 960 at 4×. Expected dimensions, tile counts and estimated image buffers appear before the model downloads. Select the originals to process; move them up or remove unwanted files.
Enlarge selected images processes files one at a time and leaves successful results available if another original fails. Each job stops after 180 seconds. Cancel stops the current worker and retains checked originals and completed downloads. Clear removes images and returns to the file field while keeping processing choices.
Optional image choices
Advanced settings starts closed. Photo uses the existing verified learned brightness model.4× runs that same 2× pipeline twice; it is not a separate native 4× model. Illustration uses smooth colour/alpha interpolation, and Pixel art repeats pixels. Those two choices are conventional maths, with no learned model download. Denoise averages similar neighbouring colours before enlargement; sharpening blends an unsharp filter after enlargement. Both start at zero.
PNG keeps lossless pixels and alpha. JPEG flattens alpha onto the chosen solid background. WebP uses its real browser encoder and can retain alpha. Unsupported encoders produce an error rather than a renamed PNG. Quality affects JPEG/WebP. Portable presets save only the processing choices, without originals, filenames or preview positions. Loading a preset or closing Advanced keeps completed downloads unchanged. Reset restores advanced defaults while retaining 2×/4×.
Inspect and save
Save individual completed images or Prepare ZIP then Download ZIP. Completed images record the factor, model and format that actually produced them. Changing current controls does not rename an earlier file. The split slider overlays the original at matching size on the actual encoded download. Matching pixel crops support source X/Y,100% or 200% zoom and arrow-key panning. Preview controls do not alter saved pixels.
The source consistency score reduces the actual encoded download to original dimensions and compares sampled RGB pixels on white.100 means no sampled difference. RGB and alpha root-mean-square error are shown separately. This is a consistency check, not a quality score or a comparison with a missing high-resolution original. A smoothed result can score highly while losing useful detail.
Example: enlarge two project photos
Try the two local examples, keep 2× and choose Enlarge selected images. Inspect the saved output and matching crops, then prepare a ZIP. To try 4×, change the enlargement and process again. Each side becomes four times larger and the pixel count becomes sixteen times larger than the original.
For a pixel-art icon, choose Pixel art in Advanced and leave both filters at zero to keep exact repeated colours and alpha. For a JPEG on a white page, choose JPEG and #ffffff. Compare the actual download before replacing the source in another project. Keep the original.
About image upscaling
Upscaling adds pixels. This compact learned model estimates brightness detail in opaque areas and uses conventional colour/alpha interpolation. Near transparency it uses conventional colour interpolation to avoid fringes. The model predicts 3× brightness internally, resampled to 2× per pass.4× uses two passes and can amplify artifacts. Generated detail cannot establish identity, authenticate evidence or prove obscured text.
Each source is at most 8 MiB,2048 px per side and 2 million pixels. Output caps 8 million pixels;4× therefore accepts at most 500000 source pixels. Up to 8 originals share 40 MiB compressed,4 million source pixels and 48 MiB prepared data. A selected batch caps 24 million output pixels. Tiled predictions are processed and released individually; output buffers are still bounded. A model pass caps 144 tiles, with 192 tiles across both 4× passes. Conservative memory checks can reject a smaller image on a low-memory device. Displayed estimates cover image buffers; the runtime and browser require additional memory.
Completed encoded images share 96 MiB; ZIP construction and retained downloads share 150 MiB. A current desktop browser with roughly 768 MiB free memory is recommended. Processing and comparisons stay local in browser memory. Only the fixed verified model/runtime assets can be cached; Clear cached model removes this tool’s model cache without removing other tools’ caches or saved downloads. Reloading clears page images and results.
Camera/location metadata and original colour profiles are not copied. This is an 8-bit image workflow, not HDR or colour-proofing. Use Batch Image Resizer for general resizing. Model provenance, license texts and runtime notices.
Stars and feedback
Loading…
What should we improve next?
Help shape the next update to Image upscaler. Tell us what would make it better for you.
Prefer email? feedback@tooltulip.com
Tool reference #169