How to Remove JPEG Artifacts and Fix Blurry Compressed Images

JPEG artifacts are the blocky edges, fuzzy halos, and banded skies that appear after aggressive compression. You can reduce them noticeably with the right techniques, but you cannot truly undo them — the data the encoder threw away is gone. This guide covers what each artifact type is, why it happens, what actually reduces it, and when re-exporting from your source file beats any amount of cleanup.

What JPEG Artifacts Actually Are

JPEG compression works in three stages. First, the encoder converts the image to a color space that separates brightness (luma) from color (chroma). Second, it divides the image into 8×8 pixel blocks and applies a discrete cosine transform (DCT). Third, it quantizes the resulting coefficients — this is where information gets discarded — then entropy-codes the result.

Every classic artifact maps to one of these stages:

Artifact	Compression choice that causes it	Typical appearance
Blocking	Heavy DCT quantization at low quality	Visible 8×8 grid, especially on smooth gradients
Ringing / mosquito noise	Quantized high-frequency coefficients near sharp edges	Fuzzy halos or “buzz” around text and outlines
Banding	Coarse quantization of gentle gradients	Visible steps in skies, shadows, and backgrounds
Chroma bleed	4:2:0 chroma subsampling	Color fringing where saturated colors meet edges
Blurriness	Aggressive smoothing baked into the encoder	Softened fine detail, waxy textures

The quality slider controls how much of each stage bites. At quality 90+, most images show no visible artifacts. Between roughly 60 and 80, artifacts stay subtle. Below 50, all five types appear together, and below 30 the 8×8 block structure becomes obvious even at normal viewing size.

The Honest Truth: Lost Data Does Not Come Back

Before any technique, set expectations correctly. JPEG is lossy. When the encoder quantizes those 8×8 blocks, it permanently discards coefficient precision. A cleanup tool can only estimate what the original detail probably looked like — it reconstructs plausible detail, not the real thing.

Three consequences follow from this:

1. Cleanup is approximation. Every deblocking filter, noise reduction pass, or AI model invents detail based on surrounding pixels. Good tools invent convincing detail; none recover exact pixels.
2. Artifacts compound. Re-saving a JPEG at low quality encodes already-artifacted pixels again. Each generation adds error on top of error. If you must edit a compressed image, work in a lossless format between saves.
3. Prevention beats repair by a wide margin. Exporting once from the original source at quality 80 produces a better result than exporting at quality 40 and running the best cleanup pipeline. The first option never loses the data.

This is why the single most effective way to fix blurry compressed images is to skip the problem: go back to your RAW file, original PNG export, or highest-quality master and re-export properly. Our JPEG optimization guide covers quality settings that keep artifacts invisible while still shrinking files substantially.

Prevention First: Re-Export From Source

If you have access to the original image, cleanup tools are the wrong tool entirely. Work through this checklist before touching a filter:

• Find the true source. Camera RAW, the design tool’s native file, or the original screenshot. A JPEG downloaded from social media is already two or three generations deep.
• Export at an appropriate quality. For photos destined for the web, JPEG quality 75–85 with 4:2:0 subsampling is usually indistinguishable from the source at typical viewing sizes.
• Resize before compressing. Downscaling averages out noise and reduces the visibility of any remaining artifacts. Compress after resizing, not before.
• Avoid repeated generations. Every edit-and-resave cycle through JPEG costs quality. Keep a lossless master (PNG, TIFF, or RAW) and export JPEGs from it as needed.

If your source is gone, the best remaining move depends on which artifact dominates. The rest of this guide covers those cases.

Technique Comparison at a Glance

Method	Effectiveness	Risk of new damage	Effort	Best for
Re-export from source	★★★★★	None	Low	Anyone with the original
Slight blur + sharpen masking	★★☆☆☆	Medium	Low	Mild blocking, quick fixes
Median / bilateral filters	★★★☆☆	Medium	Medium	Mosquito noise, mild blocking
Photoshop Reduce Noise dialog	★★★☆☆	Low–medium	Low	Photos with mixed noise + artifacts
AI upscalers with artifact suppression	★★★★☆	Low	Low	Heavily damaged images needing enlargement anyway
Convert JPG to PNG	☆☆☆☆☆ (repair) ★★★★☆ (prevention)	None	Trivial	Stopping further generational loss
Add grain for banding	★★★☆☆	None (additive)	Low	Banded skies and gradients
16-bit roundtrip	★★☆☆☆	None	Medium	Banding during editing workflows
Reshoot or re-source	★★★★★	None	High	Commercial and print use

No method earns five stars except prevention. That is not pessimism — it is the honest rating for anything working from already-lossy pixels.

Reduction Techniques That Work

Slight Blur Plus Sharpen Masking

The oldest trick in the book works because blocking and ringing are high-frequency errors. A very slight blur (0.5–1 px Gaussian) smears them out, and an unsharp mask restores the perception of edge crispness without restoring the artifacts.

In ImageMagick:

magick input.jpg -blur 0x0.6 -unsharp 0x6+0.5+0 output.jpg

The blur radius stays deliberately tiny. Go beyond about 1 px and you soften real detail along with the artifacts, which trades one form of damage for another. This method suits mildly compressed images where artifacts are visible only at 100% zoom. On heavily damaged files it does little, because the block structure survives a sub-pixel blur.

Median and Bilateral Filters

Median filters replace each pixel with the median of its neighborhood. Isolated outlier values — exactly what ringing and mosquito noise produce — vanish, while genuine edges survive far better than under a plain blur. Bilateral filters go further: they average only similar neighboring pixels, so flat areas get smoothed while edges stay sharp. That edge-preserving behavior matches the artifact problem almost perfectly.

In GIMP, use Filters → Enhance → Despeckle (median-based) or Filters → Enhance → Sharpen (Unsharp Mask) afterward to restore crispness. GIMP’s despeckle with a small radius (2–3 px) and “adaptive” enabled handles mosquito noise well.

In Photoshop, the closest equivalents live inside the noise reduction stack described next, plus Filter → Noise → Dust & Scratches at radius 1–2 for stubborn ringing around text.

Photoshop’s Reduce Noise Dialog

Photoshop bundles several of these ideas into Filter → Noise → Reduce Noise:

• Strength controls overall smoothing — start at 4–6 for artifacted photos.
• Preserve Details protects edges from the smoothing pass — 10–20% works for artifact cleanup since you want edges softened slightly, not erased.
• Reduce Color Noise targets chroma bleed and color speckling directly — raise this toward 60–80% when colors fringe at edges.
• Remove JPEG Artifact is a dedicated checkbox that applies a deblocking-style pass. It helps on visibly blocked images but tends to over-smooth, so use it only when blocking is obvious.

For severe cases, run Reduce Noise first, then apply a gentle Smart Sharpen with “Remove: Gaussian Blur” at 40–60% strength. Working on a Smart Object lets you dial both passes back without stacking generations of damage.

AI Cleanup Tools

Modern upscalers do more than enlarge — many include explicit artifact suppression models trained on compressed inputs. They recognize the 8×8 block signature and ringing patterns and synthesize clean detail in their place. Results vary by content: faces and architectural lines respond well, while fine organic textures like foliage sometimes come back plastic-looking.

Two practical notes:

• AI reconstruction is best when you need enlargement anyway. Upscaling 2× with artifact suppression often beats cleaning at 1×, because the model has room to rebuild edges rather than just smear them.
• Treat output as a new derivative, not a restored original. Save it as PNG or WebP so it does not re-enter the lossy cycle.

Our image upscaling guide compares upscalers and settings in depth. For AI-powered enhancement beyond upscaling — including background removal and automated editing that sidesteps artifact-prone workflows entirely — Sirv Studio handles enhancement, resizing, and delivery in one place.

Re-Encoding Traps: JPG to PNG Restores Nothing

A persistent myth says converting a JPEG to PNG “repairs” it. It does not. PNG is lossless, so it stores the artifacted pixels exactly as they are — every block, every halo, perfectly preserved forever. The artifacts are in the pixel values now, not the file format.

That said, the conversion has one legitimate use: stopping further loss. If an image must pass through several editing rounds, moving it to PNG (or TIFF) after the first JPEG decode means subsequent edits and saves are lossless. No new artifacts accumulate. The existing ones remain until you apply one of the reduction techniques above.

The same logic applies forward: if you are choosing an output format for new work, modern formats change the calculus. WebP and AVIF compress more efficiently than JPEG at equivalent visual quality, which means fewer artifacts at the same file size. Our compression fundamentals guide explains why, and our file size reduction guide covers format selection alongside quality tuning.

One more trap: lowering quality further does not “smooth out” artifacts. Re-saving an artifacted JPEG at lower quality makes them worse. And raising quality on re-save does nothing either — the decoder only sees the damaged pixels, so encoding them at quality 95 just stores the damage more faithfully at a larger size.

Fixing Banding Specifically

Banding deserves its own section because the standard smoothing approaches handle it poorly. Banding comes from too few distinct tonal values across a smooth gradient, and blurring cannot create tonal steps that are not there.

Add subtle grain. Dithering via noise breaks up the hard boundaries between bands. In Photoshop: add a new layer filled with 50% gray, set it to Overlay blend mode, apply Filter → Noise → Add Noise at 1–3% with Gaussian distribution and Monochromatic checked, then clip it to the gradient area. The eye reads the fine texture as continuous tone instead of steps. Photographers have used film-grain scans for exactly this purpose for decades.

Use a 16-bit roundtrip. If banding appears during editing rather than in the delivered file, the cause is usually 8-bit precision being pushed through curves or levels adjustments. Converting the document to 16 bits per channel (Image → Mode → 16 Bits/Channel) before tonal edits gives the math enough headroom that gradients stay smooth. Export to 8-bit JPEG at the end; the banding will be far less pronounced than if the whole edit happened in 8-bit.

Blur the gradient region selectively. For banding caused by heavy compression rather than bit depth, a selective Gaussian blur confined to the sky or background (mask off the subject) at 1–2 px can merge adjacent bands. Follow with light noise to prevent the blur from creating new, wider bands.

When to Reshoot or Re-Source Instead

Some images are past economical repair. Cleanup effort has diminishing returns, and these situations mark the point where reshooting or finding a better source wins:

• Print use. Print resolves far more detail than screens. Artifacts invisible at 100% on a monitor become obvious on paper. Cleanup rarely survives enlargement to print resolution.
• Heavy cropping. If the usable region is a small crop of an already-compressed image, effective resolution is tiny. A reshoot or a licensed stock replacement delivers better results than heroic filtering.
• Multiple compression generations. An image saved as JPEG, uploaded to a social platform (which recompresses), downloaded, edited, and re-saved carries stacked artifacts from every step. Filters address the latest layer only.
• Text and UI screenshots. Ringing around letterforms resists photographic noise reduction. Re-taking the screenshot at native resolution takes two minutes and looks perfect. Our screenshot guide covers capture settings that avoid the problem entirely.
• Commercial deliverables. Client-facing imagery justifies the cost of sourcing the original. Estimate cleanup time honestly; an hour of filtering that still looks processed is worse than asking for the RAW file.

A Practical Workflow

Putting it all together, here is the decision path for an artifacted image:

1. Can you access the original source?
   YES -> re-export at quality 75-85, resize first, done.
   NO  -> continue.

2. Is the image bound for print or heavy cropping?
   YES -> reshoot or re-source; cleanup will not hold up.
   NO  -> continue.

3. What dominates?
   Blocking/mosquito noise:
     median/bilateral filter or Reduce Noise,
     then gentle unsharp mask.
   Banding:
     selective blur + 1-3% monochromatic grain;
     16-bit roundtrip if you still have layered sources.
   Chroma bleed:
     Reduce Color Noise high, luma smoothing low.
   Overall softness + artifacts:
     AI upscaler with artifact suppression at 1.5-2x.

4. Save the result losslessly (PNG/WebP).
   Never re-enter the JPEG edit cycle.

Step 4 matters more than it looks. Whatever you produce, protect it: lossless intermediate formats mean future edits never compound the damage again.

Delivering Clean Images

Cleanup only pays off if delivery preserves it. Serving the repaired image through a pipeline that recompresses aggressively throws the work away. Two practices help:

• Serve WebP or AVIF versions generated once from your cleaned master, rather than letting multiple systems each re-compress the JPEG. Sirv’s Media Viewer pairs well with this workflow for galleries, zoom, and 360 spins, where viewers inspect images closely and artifacts would be exposed.
• Set sensible cache headers so intermediates and proxies stop re-fetching and re-processing unnecessarily — our image caching headers guide covers the configuration.

Key Takeaways

• JPEG artifacts come from quantization (blocking, ringing, banding) and chroma subsampling (color bleed); the quality setting determines how severe each becomes.
• Lossy compression destroys data permanently. All cleanup methods approximate; none restore.
• Re-exporting from the original source at quality 75–85 beats any repair technique and should always be step one.
• For repair: median/bilateral filters and Photoshop’s Reduce Noise dialog handle mild-to-moderate artifacts; AI upscalers with artifact suppression are strongest when enlargement is needed anyway.
• Converting JPG to PNG prevents further generational loss but repairs nothing.
• Banding responds to subtle monochromatic grain and 16-bit editing roundtrips, not to blur alone.
• Know when to stop: print use, deep crops, and multi-generation images justify reshooting or re-sourcing instead.

Ready to put this into practice? Create a free Sirv account to deliver your cleaned images through an optimized CDN pipeline.