Image Resizing: File Size, Formats and Metadata
On this page
“Make this image smaller” sounds like a lovely little feature.
Then I have to ask whether smaller means fewer pixels, fewer bytes, a square canvas, or just please make this thing upload without complaining.
That's what I like about small utilities. The button looks simple until the implementation has to decide what the user actually meant.
make the target fail visibly
A one byte JPEG budget is a useful deliberately impossible request. In test_convert.py, convert(source(), "JPEG", byte_limit=1) returns its smallest tried candidate with target_met = false; the receipt's byte count must equal the actual returned payload length.
A separate fixture fits a transparent 160×90 image inside 40×40: the output is 40×22, with a white JPEG background and the private PNG metadata removed. Exact compressed bytes depend on the encoder, so the receipt is more useful than a promised percentage saving.
After setup, python -m pytest -q test_convert.py exercises both cases without starting the web interface. Use receipt["bytes"], receipt["dimensions"] and receipt["target_met"] to decide whether the output meets your actual limit.
geometry first
A 1600 × 900 image fitted inside 400 × 400 should become 400 × 225.
If I also want a square, now I need to pad, crop or stretch it. Those are completely different results, so I don't want one generic resize option hiding all three.
def fit_inside(width, height, max_width, max_height):
values = (width, height, max_width, max_height)
if any(type(v) is not int or v <= 0 for v in values):
raise ValueError("dimensions must be positive integers")
scale = min(1.0, max_width / width, max_height / height)
return max(1, int(width * scale)), max(1, int(height * scale))
No upscaling, preserve aspect ratio and stay inside the bounds.
measure the file you're actually downloading
If I need a JPEG below 50 KB, I don't want to estimate at one quality and then re encode differently at the end.
decode
→ orient
→ resize
→ resolve transparency
→ encode candidate
→ measure actual bytes
Then choose the highest tested quality that meets the byte target.
Encoded size isn't guaranteed to move perfectly with every quality step, so the search stays bounded and measures real outputs. If none fit, say so instead of quietly shrinking the dimensions.
PNG compression effort and JPEG quality also aren't the same thing. sharp's output documentation is a good reminder that one universal quality slider can hide real codec differences.
metadata removal isn't invisibility
Orientation is a good example.
Strip EXIF too early and an upright photo can export sideways. Apply orientation first, then remove the tag if that's the selected policy.
Metadata removal can drop GPS and descriptive fields, but it cannot remove a street address visible in the pixels. I prefer saying exactly what was stripped rather than calling the result “anonymous”.
Transparency and animation need the same honesty. JPEG requires a chosen background for alpha. Animated input either needs a supported animation path or an explicit first frame export.
Silent behaviour is the thing I'm trying to avoid.
give the output a receipt
For each result I want final format, dimensions, encoded bytes, frame count, metadata policy and whether the requested limit was actually met.
The fixture set covers alpha, grayscale, EXIF rotation, text metadata, animation, truncated files, repeated names and impossible size targets.
Implementation checks and recorded findings
Across the generated image families, JPEG outputs averaged about 19 KB, WebP 31 KB and PNG 116 KB under the tested settings. These are encoded size observations, not a quality matched format ranking. Some size targets could not be met; the conversion receipts retain those outcomes.
Generated image fixtures; actual encoded bytes and metadata checks, not a photographic perceptual quality or browser study.
| Recorded metric | Mean | Seed standard deviation |
|---|---|---|
| JPEG · mean bytes | 19114.2 | 15.543 |
| JPEG · targets met | 4 | 0 |
| PNG · mean bytes | 115909 | 13.122 |
| PNG · targets met | 2 | 0 |
| WEBP · mean bytes | 31440 | 22.163 |
The comparison record includes the 5 recorded runs, measured values, source hashes and dependency versions. Variation is reported across the declared seeds; it does not establish generalisation beyond this workload.
what I built from this
The Pillow backend now decodes input, applies EXIF orientation, fits without upscaling, handles JPEG transparency against a chosen background and measures exact encoded candidates.
It selects the highest tested quality meeting the limit or reports that the target wasn't met. Fresh pixel images strip source metadata; ICC preservation is explicit. The browser form returns the same bytes described in its receipt.
See convert.py or the module README.
python3 convert.py input.png output.jpg --format JPEG --bounds 400 400 --byte-limit 50000
The local form runs at http://127.0.0.1:5051. Uploads are capped at 25 MB and decoded input at 25 million pixels.
Animation is intentionally first frame only. There is no crop editor, colour space conversion or batch UI yet, and WebP depends on the installed Pillow build.
For such a tiny tool, there are already enough decisions hiding behind “make image smaller” to keep me entertained :)