You Can Increase the Resolution of an Image, but Your Source File Decides How Far

The photo is 480 pixels wide. The banner slot wants 1600. There is no negative, no second shoot, and no bigger copy hiding in a folder somewhere. So you type the thing everyone types into a search box: how do I increase the resolution of an image that was never big enough in the first place?

Three routes actually do this. A model running inside your browser, a model running on a GPU somewhere else, and Adobe's Enhance panel if you shoot raw. Each one caps out in a different place, and every cap is published if you go looking for it. We read all three sets of specs, including the 384 pixel gate on our own free tool that quietly swallows most of a 2x button.

Which route you want depends far less on the multiplier you click than on the file you hand it. Here is what each one gives back, and where it stops giving.

Key Takeaways

• Resolution is pixel count. An upscaler predicts the pixels a larger version would contain, and it cannot recover detail the camera never captured.
• A 2x pass doubles each edge and quadruples the pixels. A 4x pass gives sixteen times the pixels of the original.
• The free browser upscaler runs Swin2SR on your own device at a fixed 2x with no account, accepts a 384 pixel long edge, and returns up to 768.
• The cloud Image Upscaler runs Real-ESRGAN with a scale control from 1.0 to 4.0 and no 384 pixel gate.
• Adobe's Super Resolution gives 2x the width, 2x the height, and 4x the total pixel count, written out as a new DNG.

What Increasing the Resolution of an Image Actually Changes

Resolution is the number of pixels in the file. Nothing else in the image's metadata moves the needle on how much detail is there, and no setting reaches back in time to sharpen a lens that was already soft.

That leaves one honest description of what an upscaler does. It looks at the pixels you have, predicts what a larger version of the same picture would plausibly contain, and writes those predicted pixels out. Reconstruction, not recovery. The Real-ESRGAN model behind most modern upscaling was trained entirely on synthetic degradations, and its authors' paper describes building a high-order degradation process to imitate what real-world damage looks like. It learned the shape of blur, then learned to argue backwards from it.

The multiplier confuses people more than the mechanism does. A 2x button sounds like twice the image. It is twice each edge, which means four times the pixels.

Worth keeping in mind when you look at that jump: a bigger pixel count says nothing about how much of it carries real information. Sixteen times the pixels from a blurry source gives you a large blurry file with cleaner edges.

How to Increase Image Resolution in Three Steps

Atlas Cloud runs two ways into the same job. One sits with the rest of the free AI tools and never uploads your file. The other is a hosted model you can call from code when one image turns into four hundred. Same task, different ceilings, and the browser one is the faster answer if your source is genuinely small.

Method 1: AI Upscale a Photo Free in Your Browser

The free AI image upscaler loads a 54MB Swin2SR model into the page and runs it on your own hardware through WebGPU. No account needed, and nothing gets uploaded. Your image stays on the device, and what downloads is a full resolution PNG with nothing stamped on it.

Three steps, and the first run is the slow one:

1. Add your image by clicking, dragging, or pasting with Ctrl+V.
2. Run the 2x pass. Budget about 60 seconds the first time, because that run includes the one time model download. Your browser caches it afterwards.
3. Compare original against result in the side by side view, then save the PNG.

The model is worth knowing about. Swin2SR came out of ECCV 2022 Workshops and was built for compressed image super resolution, which is exactly the condition most small pictures arrive in after a few rounds of messaging apps and re-saves.

One honest limit before you plan around it, and the tool page states it plainly: this tool "feeds the model a maximum long edge of 384 pixels, which caps the output at 768." Anything bigger gets scaled down before the model sees it. Feed it a 1200 pixel photo and what comes back is smaller than what you started with.

Method 2: Increase Image Resolution Through the API

When the job is a product feed rather than a single photo, the hosted Image Upscaler model takes over. It runs Real-ESRGAN, exposes a scale control from 1.0 to 4.0, and has no 384 pixel gate in front of it. One image goes in per request, so volume comes from looping the call rather than from a bulk upload field.

Step 1: Get your API key. Create a key in the Atlas Cloud console and store it as an environment variable, never in client-side code.

Step 2: Check the API docs. Endpoints, parameters, and authentication live in the API documentation.

Step 3: Make your first request. Submit the image and the scale you want:

BashCopy
1curl -X POST https://api.atlascloud.ai/api/v1/model/generateImage \
2  -H "Content-Type: application/json" \
3  -H "Authorization: Bearer $ATLASCLOUD_API_KEY" \
4  -d '{
5    "model": "atlascloud/image-upscaler",
6    "image": "https://example.com/product-shot-600px.jpg",
7    "outscale": 4.0,
8    "output_format": "png"
9  }'

The response carries a prediction ID. Poll it until the status reads completed:

BashCopy
1curl https://api.atlascloud.ai/api/v1/model/prediction/<prediction_id> \
2  -H "Authorization: Bearer $ATLASCLOUD_API_KEY"

The completed response returns the hosted image URL. Set output_format to png when the result feeds another pass, since a JPEG round trip adds its own compression on top of whatever the source already carried.

That submit-and-poll shape is the part that pays off later. Atlas Cloud runs one API across the whole catalogue, so one key, one auth header, and one polling loop cover every model. Swap the model string and you are calling something else entirely, which is the same pattern we walk through for portraits in the AI headshot API guide.

Which Upscaling Route Fits the File You Have

Pick by source, not by the biggest number on offer. A 300 pixel avatar and a 3000 pixel product photo want different machines, and the second one will be disappointed by the first.

That third row is the one people miss. The Tencent model takes a mode of percent, aspect, or fixed, so instead of asking for a multiplier you can name the long edge you need and let it land there. Its four engines split by material: standard for generic work, super for AI generated scenes, ultra for more detail, and fidelity aimed at ecommerce imagery where the product has to stay honest.

Neither hosted model publishes a pixel ceiling, which is the one number people want before they start a batch. What is checkable is the file size gate in front of the playground: its uploader validates images against a 30MB limit and refuses anything heavier before the model sees it. Going through the API you hand over a URL instead of a file, so that particular check never runs, and the model documentation says only that very large inputs may be rejected before processing.

Both hosted models return a URL rather than a file, and both bill per image rather than per pixel, which makes a 4x pass on a small file and a 4x pass on a large one cost the same.

How Big Does the Final Image Need to Be?

Before choosing a multiplier, work backwards from the slot. Wanting to make an image bigger is not a target. 1920 pixels wide is a target, and it tells you which route can even reach it. A forum avatar and a full bleed 4K wallpaper are two different problems, and only one of them is solvable from a 480 pixel source.

Read that chart as a shopping list. Filling a 1080p hero from a 600 pixel source needs the 4x route, and even then you are asking the model to invent three quarters of what lands on screen. Filling an avatar slot needs almost nothing, which is why the browser tool covers that case completely.

Print runs on different arithmetic, since the pixel target moves with physical size and printer, and the resolution tag in the file has its own trap waiting. That whole path, including the print size tables, sits in the guide on increasing DPI properly. Posters are their own particular headache, handled in the walkthrough on generating a poster with AI.

Increase Photo Resolution in Photoshop or Lightroom

Shoot raw and there is a fourth door. Adobe's Enhance panel in Camera Raw holds two separate features that get talked about as one thing, and only one of them changes the pixel count.

Adobe's own Camera Raw documentation puts the Super Resolution figure plainly: the enhanced image "will have 2x the width and 2x the height of the original image or 4x the total pixel count." The same page calls it "especially useful for increasing the resolution of a cropped image," which is the case it was built for. You cropped hard, the frame is now small, and the sensor data is still good.

One trap costs people an afternoon. Super Resolution is unreachable through Photoshop's Camera Raw Filter. Open the file with File then Open instead, and JPEGs or TIFFs need the Camera Raw preference for handling those formats switched on before they will route through the dialog.

Plain resampling in the Image Size dialog is the non-AI baseline, and it does something meaningfully different: interpolation spreads the pixels you have across a bigger grid. It gets you the dimensions. It does not rebuild edges.

Increase Image Resolution Without Losing Quality

People go looking for a way to upscale an image without losing quality, and the phrase hides an assumption. Nothing is lost during an upscale, because nothing from the original is thrown away. What people are actually noticing is that the invented pixels look different from photographed ones, and at 100% they look invented.

Three habits keep that gap small.

Crop tight before the pass, so the model spends its capacity on the subject instead of a wall. Feed the original file rather than a screenshot of it, since a screenshot has already been through one round of scaling and compression. Keep a single subject in frame when you can, because a model resolving one face does better than a model resolving a crowd.

The failure mode has a name in the literature. Real-ESRGAN's authors explicitly handle ringing and overshoot artifacts, the halo and the hard bright line that appear along strong edges when a reconstruction gets confident. On skin and fabric they read as a slightly waxy surface. On architecture they read as a glow.

Stacking passes makes all of it worse. A second 2x on an already upscaled file feeds the model its own guesses as evidence, and it treats them as photographed detail. One pass at the scale you need beats two passes that add up to the same number.

When Image Upscaling Makes a Photo Worse

Some files are better left alone, and knowing which ones saves a round trip.

Small text is the case people expect to fail, and our test did not fail the way we expected. A 360 pixel shipping confirmation with 7 to 9 pixel type went through the browser tool at 2x and every word came back legible, the footer a touch softer than the original but intact. What the pass did not do is the thing people actually want from it. The letters got bigger without getting any sharper… For text, re-export from the source document instead.

Logos and icons should never go near an upscaler if a vector original exists anywhere. An SVG or an EPS scales to any size with no invention involved at all, which is the whole point of vector artwork.

Already sharp images have little to gain. The tool page says as much, and the reason is structural: super resolution models earn their keep by undoing degradation, and a clean file has none to undo.

Then there is subject matter the model was never specialised for. The browser tool runs classical super resolution weights with no anime or document specific training, so line art and scanned pages are a coin flip rather than a promise.

Frequently Asked Questions

Can you resize an image without losing quality?

Resizing down and resizing up are different jobs. Shrinking an image loses data but rarely looks worse, since you are discarding pixels rather than guessing at them. Going the other way with a plain resize dialog spreads existing pixels across a larger grid by interpolation, which holds the composition and softens the detail. To resize an image without losing quality in the upward direction, you need a model that rebuilds edges instead of stretching them, which is what an AI upscaler is for.

How do I increase the resolution of a photo on my phone?

Open the free browser upscaler in your phone's browser. The model downloads once and runs on the device, so nothing leaves the phone. One caveat matters on mobile: almost any modern phone photo is far larger than a 384 pixel long edge, so it gets scaled down before the pass and comes back at 768. Crop to the part you actually need first, or use the hosted model when the full frame has to stay big.

Can a low resolution image become a high resolution one?

It can gain pixels and it can gain apparent sharpness. Whether it gains information depends on what survived in the original. A 400 pixel photo taken in good light with a steady hand upscales well, because the edges are all still there and the model only has to extend them. A 400 pixel frame grab from a shaky video has lost the edges themselves, and no amount of low resolution to high resolution processing puts them back. Judge the source before you judge the tool.

Is increasing image resolution free?

The browser route is free outright: no account, no daily cap, no credits to spend, no watermark on the way out. The hosted models bill flat per image: the Atlas Cloud Image Upscaler at $0.0100 per image and the Tencent Image Upscaler at $0.024 per image, both list prices from each model's own documentation with no discount applied. A 4x pass and a 1x pass cost the same.

How many times can I upscale the same image?

Technically as often as you like, and practically once. Each pass treats the previous pass's predictions as real detail, so artifacts compound and skin turns plastic faster than the resolution number climbs. If one 2x is not enough, go straight to a single 4x on the original file instead of running 2x twice.

Conclusion

Every route here is doing the same thing under different constraints: reading the pixels you have and writing more of them. The browser tool trades reach for privacy and cost, staying entirely on your device and topping out at 768 pixels. The hosted models trade that privacy for room to run, taking whatever you can host and scaling up to 4x or naming an exact target edge. Adobe's Enhance panel sits apart from both, waiting for a raw file and paying off best on a hard crop.

None of them will rescue a photo that lost its edges before you ever opened it. That is the real ceiling, and it was set at the moment of capture. So the practical way to increase resolution of an image is to start from the sharpest, smallest, tightest crop you own, pick the route that matches it, and run one pass rather than three. Check the result at 100% before you commit, because the number in the file name is not the thing your reader sees.