Online upscalers and denoisers usually need your image on their server. For many photos that does not matter much. For client work, product shots before launch, or any image you would rather keep to yourself, it does.
LabelOp Retouch runs both tools in the browser tab. The models download once, and from then on the processing happens on your device.
This guide covers what the two tools do, how to use them, what order to apply them in, and when to leave an image alone.
Short answer
- Upscale (U) enlarges a photo 2x or 4x with NERVE 4x, a model released in 2026 (Apache-2.0, about 7 MB).
- Denoise (N) removes noise, JPEG artifacts and mild blur with GaterV3 Restore, released in 2025 (CC-BY-4.0, about 86 MB). A strength slider from 0 to 100% sets how much it does.
Both tools work on the image in 256 px tiles and blend the overlapping edges, so large photos do not have to fit through the model in one go. Each result goes on a new layer, and the original stays underneath.
Your images are never uploaded. NERVE downloads from Hugging Face. GaterV3 comes from the Retouch site itself, because its author publishes it only as a GitHub release, which browsers cannot download directly. Both are cached after the first download and then work offline.
What these tools can and cannot do
Before you start, it helps to be clear about what an AI upscaler does.
An upscaler does not recover lost detail. It predicts what a sharper, larger version of the image would plausibly look like. On textures, edges and surfaces, that often looks very convincing. On fine text, small logos or tiny faces, the model can make up details that were never there. Check those areas at 100% zoom before you use the result.
A denoiser has the opposite trade-off. It removes grain and compression blocks, but at full strength it can also smooth away real texture such as fabric, skin pores or wood grain. That is what the strength slider is for.
Mild blur is within GaterV3's scope. Heavy motion blur and badly out-of-focus shots are not.
Which order: denoise first, then upscale
If a photo is both small and noisy, denoise first. An upscaler enlarges everything it sees, compression blocks and sensor noise included. Cleaning those up at the original size gives the upscaler a better image to work from, and the denoise step runs on fewer pixels.
A typical sequence:
- Crop (X) to the part you actually need, which also saves time on both models.
- Denoise (N) at the lowest strength that clears the artifacts.
- Upscale (U) 2x or 4x.
- Fine-tune in Adjust (A) with exposure, contrast, saturation and Sharpen, or try Auto enhance.
- Export.
How to upscale a photo
Open the photo in Retouch (Open, drag and drop, or paste with Ctrl/Cmd+V). Select the layer you want to enlarge and press U.
- Choose ×2 or ×4.
- Check the Output line, which shows the final width, height and megapixels before anything runs.
- Click Upscale ×2 or Upscale ×4. On first use, approve the model download (about 7 MB).
- Progress is shown per tile ("Upscaling tile 3/48"), and you can stop at any time.
The result is a new layer named after the source with "x2" or "x4" added. The document grows to the new size, and the original layer is hidden underneath.
The model always enlarges by 4x. For 2x, each tile is scaled down as it is placed, so the full 4x image never has to sit in memory.
Upscaling keeps transparency and editable masks. That means you can cut out a product first and upscale the cutout afterwards.
Size limits
Browsers limit how big a canvas can be. Retouch checks before running, so you do not wait through a long job that would fail at the end. The output has to stay within about 67 megapixels (8192×8192). Some examples:
| Source | ×2 output | ×4 output |
|---|---|---|
| 1000×750 | 2000×1500 (3 MP) | 4000×3000 (12 MP) |
| 2000×1500 | 4000×3000 (12 MP) | 8000×6000 (48 MP) |
| 3000×2000 | 6000×4000 (24 MP) | 12000×8000 (96 MP), too large |
If the output is too large, the panel shows a warning with a Resize image first button. Usually the better choice is ×2, or cropping first.
How to denoise a photo
Press N to open Denoise.
- Set Denoise strength. Lower values keep more of the original texture, and 100% applies the model's full output.
- Click Denoise image. On first use, approve the model download (about 86 MB).
- Compare with Before / after (C) and drag the split across a textured area.
The result is a new layer named after the source with "denoised" added. Transparency and the editable mask are kept, and normal undo works.
At strengths below 100%, each processed tile is blended with the original pixels on the GPU before the tiles are put back together. At 0% nothing runs and no model is downloaded.
A good habit is to start around the middle of the slider. Raise it only until the blocks or grain you care about are gone.
Sharpen and Auto enhance
The Adjust panel (A) has a single Sharpen slider next to exposure, contrast and saturation, with a live histogram. Sharpen runs in the same GPU pass as the preview and needs no model download. It also skips transparent neighboring pixels, so cutout edges do not get dark fringes.
Auto enhance looks at a small copy of the image and sets a starting grade. Transparent padding is ignored. You can then adjust the sliders from there.
Use Save as layer if you want the adjustments on a separate layer. The Export dialog also lets you include or leave out adjustments.
Many photos at once: Batch
Open Batch in the left sidebar and pick Upscale ×2, Upscale ×4 or Denoise. Add PNG, JPEG, WebP or AVIF files. Retouch processes them one after another on your device, and you download each result or click Save ZIP.
Batch results are saved as PNG. Batch denoise always uses full strength. For finer control over strength, process the image in the editor.
When not to upscale or denoise
Images for a training dataset
If you build computer-vision datasets, be careful with these tools. A model learns from what it sees. If training images are upscaled or denoised but the camera in production is not, the training data no longer looks like the real input.
Detail an upscaler made up is especially risky for small objects. A label might then describe something the camera never captured. Keep the originals for training and labeling, and use enhanced copies only when your production pipeline applies the same step.
For more on this, see small object detection labeling and our guide to preparing object images for a computer vision dataset.
Evidence and documents
If the exact pixels matter, as with a receipt, a scan for records or a photo used as evidence, an AI-enhanced version is a different image. Label it as enhanced, or keep the original.
Images that are already sharp and clean
Upscaling a sharp 4000 px photo 2x mostly creates a bigger file. Check what size you actually need. The export dialog can also scale down or fit the image to a preset.
Exporting the result
Press Ctrl/Cmd+S. PNG gives the best quality, and WebP or AVIF give smaller files with transparency kept (AVIF only if your browser can encode it). JPEG works everywhere but turns transparent areas white. Size presets cover common social and marketplace sizes, or you can keep the original proportions.
Every export is re-encoded, so camera EXIF data, including GPS location, is removed.
Requirements and limits
- Browsers. Chrome and Edge 113 or later get full WebGPU support. Safari 26 and later currently use a WebAssembly fallback. Firefox uses WebGPU where it is available and the fallback otherwise. Tiled models run much slower without WebGPU.
- Desktop first. Mobile and touch input are not optimized yet.
- Input limits. Files over 50 MB, or with a side longer than 16,384 px, are refused. Above 24 MP the editor asks whether to downscale.
- Recovery saves a flat image. If the tab closes unexpectedly, the editor can restore a flattened image, but not your layers or undo history. Export when you are done.
Part of the LabelOp family
Retouch sits alongside LabelOp's other browser tools: the LabelOp labeling platform, Tracker for video annotation, and Moments for finding scenes in video. Like them, it does the heavy work on your own device whenever possible.
To try it, open LabelOp Retouch, drop in a photo and press N or U. You do not need an account.
FAQ
Is my image uploaded when I upscale or denoise it?
No. Both models run in your browser tab on your device. The only downloads are the models themselves, once, and the browser caches them for offline use.
How large can the upscaled image be?
The output has to stay within about 67 megapixels, the browser canvas limit. Retouch shows the output size before running and warns you if it is too large, so you can choose ×2 or resize first.
Does the upscaler recover real detail?
No. It predicts plausible detail. That looks convincing on textures and edges, but it can invent details in small text, logos or faces, so check those at 100% zoom.
Should I denoise before or after upscaling?
Before. Upscaling enlarges noise and compression artifacts too, so clean them up at the original size first. Then upscale the cleaner image.
Can I process a folder of images?
Yes. Batch mode handles Upscale ×2, Upscale ×4 and Denoise on many images, one after another on your device, and exports PNGs individually or as a ZIP.