# shoots.page photo derivatives: Pillow vs ImageMagick (vs libvips)

The job being measured is exactly what processor/decode.py does per upload: decode, EXIF orient, ICC to sRGB, 2000 px preview (q88) + 480 px thumb (q82). Watermarking measured separately.

Contents

# Results

9 real camera JPEGs at native resolution, 1 pinned core (prod container is 1 vCPU), best of 3. CPU is whole-process time, so the Python rows include ~95 ms interpreter startup that the long-lived prod server never pays.

Pipeline	CPU/photo (mean)	Worst (61 MP)	Peak RAM	SSIMULACRA2	Butteraugli p3	Preview size
Pillow, prod as-is	2.08 s	3.39 s	496 MB	82.69	0.941	477 KB
Pillow, reordered	0.69 s	0.83 s	207 MB	82.52	0.950	475 KB
ImageMagick 7 Q16, tuned	1.31 s	1.38 s	266 MB	82.24	0.953	471 KB
ImageMagick 7 Q16, naive	2.26 s	2.41 s	522 MB	82.32	0.952	473 KB
ImageMagick 7 HDRI (official build), tuned	3.49 s	3.34 s	272 MB	82.21	0.955	472 KB
ImageMagick 7 HDRI (official build), naive	8.50 s	10.84 s	1040 MB	82.33	0.952	473 KB
libvips (control)	0.53 s	1.04 s	148 MB	81.78	1.057	475 KB

Quality is a dead heat. Everything lands within 1 SSIMULACRA2 point; the q88 JPEG step dominates. Pillow scores highest against all three references, including ImageMagick’s own. Ranking is identical across references.
Pillow isn’t slow, our order of operations is. Prod decodes the full frame and converts ICC on all 24-61 MP before resizing. On the 45 MP Adobe RGB file: decode+orient 556 ms, full-res ICC 1691 ms, resize+encode 409 ms. Reordered: half-scale decode 283 ms, orient+resize 175 ms, ICC on 2000 px 102 ms, encode 30 ms.
ImageMagick only gets decent with a hint. Without -define jpeg:size= it is as slow as prod today and peaks at 522 MB. The official build is HDRI (float pixels): 2.7x slower than Debian’s Q16, up to 1 GB RAM.
ImageMagick gotcha: it copies the camera’s chroma subsampling (4:4:4 on 7 of 9 files), so default previews come out 19% bigger on average (5-32%). Forced to 4:2:0 here for a fair comparison.

# Watermark

Overlay (tiled diagonal studio name, 30% white) built once per gallery, center-cropped per photo.

	Per photo (composite + encode wm preview + wm thumb)	One-off overlay build
Pillow	41 ms	177 ms
libvips	43 ms	73 ms
ImageMagick	+242 ms (reloads overlay every process)	3.3 s

Watermarked preview is ~7% bigger (508 KB vs 475 KB).

# Setup

Corpus (Wikimedia Commons, native res): 24 MP x2 (Capture One, sRGB), 33 MP Sony A7 IV (no ICC), 45 MP Canon R5 x2 (Adobe RGB), 45 MP Nikon Z8 (Nikon sRGB ICC), 61 MP Sony A7R V (no ICC), 61 MP Lightroom portrait, plus a 45 MP Adobe RGB file with EXIF orientation 6.
Versions: Pillow 12.3.0 + numpy 2.5.3 (prod pins, Python 3.13), ImageMagick 7.1.1-43 Q16 (Debian), ImageMagick 7.1.2-32 Q16-HDRI (official AppImage), libvips 8.18.7. All encode via libjpeg-turbo at 4:2:0 with the same sRGB ICC.
Pipelines: prod = current write_derivatives() called directly. Reordered = same libs, draft() half-scale JPEG decode when the source is at least 2x the target, orient + resize first, ICC on the 2000 px raster. IM tuned = -define jpeg:size=<2x target> -auto-orient -resize 2000x2000> -profile sRGB -strip -profile sRGB -sampling-factor 4:2:0, preview + thumb in one process. IM naive = same without the hint. libvips = thumbnail() with shrink-on-load and export profile.
Quality: SSIMULACRA2 and Butteraugli (libjxl 0.12) against three independent float-precision Lanczos references from the full-res original (libvips, ImageMagick HDRI, Pillow float). The references agree with each other at 94.5, so tool-to-tool resampling differences are far below the JPEG loss.
Not covered: ImageMagick 6 (what bookworm apt ships), HEIC/RAW/PNG/TIFF inputs, mozjpeg.

End