How to Fix Pixelated Photos (and Which Kind You Have)
Pixelated can mean too few pixels, JPEG compression blocks or a deliberate mosaic. How to tell them apart, what AI can fix, and what it only guesses.
On this page
"Pixelated" covers three different problems: a photo with too few pixels shown too big, a JPEG damaged by heavy compression, and a deliberate mosaic over a face or text. The first two can be improved with AI, but a mosaic can't be reversed, and in every case the best fix is often a better copy of the file rather than a better tool.
If your photo is speckled rather than blocky, that's noise; see how to fix grainy photos. If it's smeared or soft, start with how to fix blurry photos. This post is about the squares.
Which kind of pixelated is it?
Open the photo at 100% zoom and look at a smooth area (sky, a wall, a cheek) and at a sharp edge (a collar, a line of text).
Scroll horizontally to see all columns.
| What you see | What it is | What helps |
|---|---|---|
| Even squares of flat color across the whole image, and small pixel dimensions | Too few pixels, enlarged | A bigger copy, or upscaling |
| A faint grid of tiles in smooth areas, smudgy halos around edges, colors stepping in bands, even on a large file | JPEG compression blocks | A less-compressed copy, or deblocking |
| Big, deliberate squares over only one part, like a face, a license plate or a name | A privacy mosaic | Nothing recovers it |
There's a fourth lookalike: a scanned newspaper or magazine photo made of tiny dots in a regular pattern. That's halftone printing, not pixelation, and it's a restoration job.
The first kind is the one our upscaling guide covers: working out how many pixels you need and enlarging to get there. The rest of this post is mostly about the second kind, because it's the one people misdiagnose. A 4000-pixel photo can look as blocky as a thumbnail, and making it bigger just makes bigger blocks.
Why JPEG makes blocks
JPEG doesn't store a photo as one picture. According to the standard itself, ITU-T T.81 (section 4.3), the encoder groups the image's samples "into 8 × 8 blocks," transforms each block with the discrete cosine transform (DCT) "into a set of 64 values," then quantizes each of those values against a table. Quantization means rounding, and the rounding is where detail is thrown away. T.81 is explicit that the DCT-based modes are lossy.
Each 8×8 block is rounded on its own. Under light compression you never notice. Under heavy compression, neighboring blocks stop matching at their borders, and the image turns into a mosaic of little tiles. Fine edges suffer differently: the rounding leaves ripples beside them, the halos you see around text and hair.
Every app that re-saves the photo as a JPEG can run it through that process again. A picture that went from a camera to a chat app to a social post to a screenshot has been through it several times, and each pass can only discard more.
Step 1: find a better copy before fixing anything
This is the step most guides skip, and it's often the biggest improvement you can make. Compression damage is detail that was rounded away; somewhere, a copy may exist that still has it.
- Ask the person who sent it for the original file itself, not a forward of a forward or a copy saved out of a chat.
- Check shared albums. Apple's Shared Albums documentation says that on version 26 or earlier of its operating systems, shared photos "are reduced to 2048 pixels on the long edge" (Apple says version 27 or later keeps full resolution). The original may still be on the photographer's phone.
- Check your backup settings. Google says photos backed up in Storage saver "are compressed to save space," and anything over 16 MP is resized. An older hard drive or the original phone may hold a better copy.
- Never work from a screenshot if the photo itself exists somewhere. A screenshot captures the display, then gets compressed again.
- For prints, rescan. A photo that only exists as a small, blocky scan can be scanned again properly; see how to scan old photos.
Then check the pixel dimensions of what you have. If it's large and still blocky, you have compression damage. If it's small, you have a resolution problem, possibly with compression on top.
Step 2: fix compression blocks, gently
Removing JPEG blocks is called deblocking, and it's a harder problem than it looks. A 2015 research paper on reducing compression artifacts with a deep network sums up the problem: lossy compression causes "blocking artifacts" along with "ringing effects and blurring," and older methods either smoothed the blocks away and blurred the photo, or kept it sharp and left the ripples. Their network did better on both, and was tested on real compressed images from Twitter.
If you already use Photoshop, it has a dedicated JPEG artifacts removal filter, a Neural Filter that Adobe says rids an image of "flaws caused by compression," with Low, Medium and High strengths. Start at Low.
That's the honest shape of every deblocking tool, AI or not:
- It smooths the grid by guessing what each block looked like before rounding.
- It cannot recover rounded-away detail. What comes back in a cheek or a sky is a plausible texture, which is usually fine.
- Text, logos and fine patterns are the risk. Letters squashed into a few blocks can come back as different letters. Faces can shift slightly. Compare those areas with the original.
A few things make it worse:
- Don't sharpen first. Sharpening raises contrast at edges, and the border between two blocks is an edge. The grid gets louder.
- Don't enlarge with a plain resize first. Bigger blocks are still blocks, and now the tool has more of them to clean.
- Don't re-save as JPEG between steps. Keep working copies as PNG until the final export.
- Run one pass, then stop. Repeated AI passes stack their guesses.
Once the blocks are gone, if the photo is also too small for where it's going, upscale it. Upscaling a clean small image is a much easier job than upscaling a blocky one.
Deliberate pixelation can't be undone
Sometimes the squares are the point: a face blurred out in a news photo, a plate mosaicked before posting. No tool can recover what was there, because the mosaic replaced a whole area with a few averaged colors.
AI "depixelation" doesn't change that. The authors of PULSE, a face upsampling model, address this directly in their project README: it "makes imaginary faces of people who do not exist," and "it will not help identify or reconstruct the original image." A sharp face from a mosaic is a new face that happens to fit the squares.
The reverse warning also holds. If you're the one hiding something, don't rely on pixelation. Security firm Bishop Fox built a tool that recovered pixelated text by pixelating guesses one character at a time and comparing them. Its advice: "use black bars covering the whole text."
Fixing a pixelated photo in 43frames
43frames has no dedicated deblocking control, and we won't call either tool one. Here's what each actually does with a pixelated photo.
Small, with visible squares: photo upscaling. Choose 2× or 4×. The default, Keep it faithful, runs Topaz's High Fidelity V2 model and costs 2 credits. Topaz describes that model as best for high-quality images where you want to keep the original's character, so expect it to enlarge what's there, compression blocks included. The Advanced option Add detail (4 credits) redraws more texture; we haven't tested whether that clears blocks, and the app warns that it may change faces and text. It's meant for small images; on a larger photo, the app may reduce 4× to 2× "to keep this photo within the supported size."
Larger, with compression blocks: photo restoration. Restoration works from a larger copy of your upload and redraws the whole photo, instructed to return "a clean, sharp, professionally scanned original print" and to remove capture defects including noise, scan lines and halftone or moiré patterns, while keeping the people, faces and scene as they are. It returns 4800 pixels on the long side and costs 8 credits. The instruction is written for old photographs, JPEG blocks aren't named in it, and we haven't measured it on them specifically, so treat it as an experiment and try it on one photo first. It's generative: compare every face and any lettering with your original before you keep it.
A deliberate mosaic: neither tool, for the reasons above.
Both tools take JPEG, PNG or WebP up to 10 MB. New accounts get 10 starter credits once, enough to try either on a photo, and more credits come as one-time packs on the pricing page. Keep your original file untouched, whatever you run.
FAQ
Why does my photo look pixelated even though it has lots of pixels? That's usually JPEG compression. The format rounds away detail in 8×8 blocks, so a heavily compressed file shows tiles at any size. A less-compressed copy fixes it; more pixels don't.
Can AI unpixelate a blurred-out or mosaicked face? No. It can draw a face that fits the squares, not the real one. PULSE's authors say their model makes imaginary faces and won't reconstruct the original.
Should I sharpen a pixelated photo? Not first. Block borders are edges, so sharpening makes the grid more visible. Reduce the blocks or upscale, then sharpen lightly.
Which 43frames tool should I use for a pixelated photo? Upscaling (2 credits, 4 with Add detail) for small images with visible squares; restoration (8 credits) is worth one test for compression blocks on a larger photo, though we haven't measured it on JPEG blocks. Compare faces and text afterward. Neither undoes a deliberate mosaic.