How few colours can you actually see? A 2AFC test, and a request for replication

[saitoha]

Split out from #196 so the palette-size discussion there is not buried under result reports.

#196 has been an argument about how large a SIXEL palette should be, conducted almost entirely from the encoder side. What has been missing is the other half: what a person actually perceives on a real terminal. I built a test to measure it, ran it on myself, and I would like other people to run it too — the scripts are linked below.

The short version

On 640×480 photographic content encoded with libsixel's current develop quantiser and dither, I cannot tell a 128-colour image from a 256-colour one. Not by looking at which is prettier, and not by deliberately hunting for the artefact I have spent a decade learning to spot.

The same procedure — same images, same display, same ten-second window — picks out a 32-colour palette at p = 0.035 and a 16-colour palette at p = 0.000046. So this is not a blind test that fails to detect anything; it detects small palettes easily and stops detecting somewhere between 64 and 128 colours.

What that does and does not mean. It does not show that a 256-colour limit is the right ceiling for SIXEL. It shows that for photographs — the content most often shown off in these discussions — quality saturates well below 256 colours once you have a decent adaptive palette and a decent dither. If the case for 1024 or 4096 colours is going to be made, it has to be made on the content where a small palette actually hurts: smooth gradients, flat-shaded CG, UI screenshots, scientific colour maps. I think that is where this discussion should go next, and I would rather get there with measurements than with assertions.

I am one observer on one laptop panel, which is not enough to conclude anything. Hence the request.

Results so far

All runs use the same 100 photographs. xterm on XQuartz, 43.6 ppd (the 1280 px image measures 256 mm on screen; viewing distance ~500 mm):

Comparison	Judging appearance	Deliberately hunting the artefact
256 vs 16	—	70 / 99 (70.7 %), p = 0.000046
256 vs 32	—	61 / 100, p = 0.035
256 vs 64	48 / 100, p = 0.76	60 / 98 (61.2 %), p = 0.033
256 vs 128	47 / 100, p = 0.62	51 / 100, p = 0.92

iTerm2, dot-by-dot, 87.3 ppd, same observer and distance, judging appearance: 256 vs 64 → 51/100, 256 vs 128 → 51/100.

Denominators below 100 are the trials actually answered; a trial with no keypress inside ten seconds is recorded as a timeout and excluded rather than scored as an error.

Three things come out of this.

The method has sensitivity. 16 colours at p = 0.000046 settles the obvious objection that ten seconds and a side-by-side layout are simply not enough to see anything. They are enough. The 128-colour result is a null of the phenomenon, not of the instrument.

How you attack the task matters at 64 colours and not at 128. Holding observer, images, display and distance fixed and changing only my strategy moved 64 colours from 48 to 60 (Fisher p = 0.066) and moved 128 colours from 47 to 51 (Fisher p = 0.67). What I was doing in the second pass was specific: find the small-area outlier colours and check whether their saturation has collapsed. That technique pays at 64 colours. At 128 there are too few images left containing a region where it applies.

Nothing survives at 128 colours. Pooling every 128-colour run — both strategies, both rendering paths — gives 149 / 300, p = 0.95, 95 % CI [43.9 %, 55.5 %]. For this observer on this display the true rate is within about six points of chance in either direction.

So my detection threshold sits between 64 and 128 colours, and that is a threshold for deliberate expert search. For ordinary appearance judgement it is already gone by 64.

Why you should not weight my numbers heavily

I am close to the worst available observer for this. I have worked on colour quantisation for over a decade; I know where a shrinking palette breaks first, and by the 16-colour run I was no longer judging appearance at all. That makes me a decent upper bound on detectability — if someone who knows exactly where to look cannot beat chance at 128, a person looking at a picture almost certainly cannot — and a poor witness for the appearance question that the test actually asks.

It also means the ladder above is not a clean psychometric function: the 64- and 128-colour appearance runs happened before I had adopted the search strategy, and the 16- and 32-colour runs after. That is why I re-ran 64 and 128 under the search strategy, and why both columns are shown.

What the discussion needs is naive observers, on ordinary displays. That is the ask.

Method

• 100 trials. Each trial shows the same photograph twice, side by side (1280×480 total): once encoded with 256 colours, once with the reduced palette.
• The side carrying the 256-colour image is an exact 50/50 shuffle (fixed seed, maximum run length 4), so holding down one key scores exactly 50.
• Press 1 for the left image or 2 for the right one — whichever looks better. You are never told which one has more colours. Please guess when you cannot tell; a timeout only costs statistical power.
• Ten seconds per trial. A summary line is printed at the end.
• Sources are 100 distinct photographs from the Wikimedia Commons Featured pictures collection, all under CC0, public domain, CC BY or CC BY-SA. Attribution is embedded in each script's header and listed at the bottom of this post.
• Each source is centre-cropped to 640×480 at 1:1 and never resampled, so what you judge is the encoder's output and not a scaler's.

Encoding, on the develop branch:

img2sixel --threads=2 \
          --diffusion=lso2 \
          --quantize-model=kmeans:inittype=pca:binning=none:merge=ward \
          --clustering-colorspace=oklab --working-colorspace=cielab \
          --colors=<256|128|64|32|16>

That is a k-means palette (PCA init, no binning, Ward merge) clustered in OKLab with libsixel's lso2 variable error-diffusion tables — the current best encoder path, not the historical median-cut default. The claim under test is "256 colours are enough given a good quantiser and a good dither", not "256 colours are enough however you encode".

The objective difference is not zero. Over a 20-image sample, PSNR between the 256-colour version and the reduced one:

Palette	PSNR vs 256 colours
128	35.1 dB (30.5–40.0)
64	33.4 dB (28.3–38.5)
32	31.3 dB (25.8–37.5)
16	28.6 dB (23.1–35.6)

The question is whether that difference survives the trip through your display and your eye.

Why every test uses the same 100 photographs

The four scripts show the same images. The small-palette tests are not a second measurement of the same thing — they are a positive control.

If everyone scores 50 on the 128-colour test, the first reasonable objection is that the task is impossible. The only way to answer it is to show that the same method, on the same images, with the same observer, does detect a difference when a larger one is present. Holding the image set constant is what makes that argument work: with different photographs in each test, "detected at 16, not at 128" could just as easily be a difference in content difficulty.

The cost is carry-over from seeing the same photographs more than once, and it is worth being explicit about which way it pushes. Remembering which side you pressed is unlikely — these are near-coin-flip judgements — and the scripts use independent left/right sequences, so a repeated side-choice carries little information. Remembering where an image is weak is much more likely, and it makes you more sensitive on later runs. That is why the suggested order is smallest palette first: it puts the run that matters, 256 vs 128, where familiarity is working in favour of detecting a difference rather than against it.

If you prefer a different order, that is fine — just say which order you used.

Running it

curl -LO https://github.com/saitoha/libsixel/releases/download/perception-test-2026-08/test-p256-vs-p128-commons.bash
bash test-p256-vs-p128-commons.bash

Script	Comparison	Size
test-p256-vs-p128-commons.bash	256 vs 128	71 MB	the one that matters
test-p256-vs-p16-commons.bash	256 vs 16	54 MB	sensitivity check
test-p256-vs-p32-commons.bash	256 vs 32	59 MB	optional
test-p256-vs-p64-commons.bash	256 vs 64	64 MB	optional

Fifteen minutes: run 128. Half an hour: run 16 first, then 128 — the 16-colour run is what makes a null at 128 mean something.

The scripts read their own image payload from $0, so they must be run as files; curl … | bash will not work. Ctrl+C restores the terminal.

Requirements:

• A terminal with SIXEL support, at least 256 colour registers, and per-image (private) colour registers.
• At least 1280×496 px of usable text area — if the image makes the screen scroll, the two halves will not line up.
• bash, base64 -d, gunzip.
• xterm: xterm -ti vt340, with XTerm*maxGraphicColorRegisters: 256 or more and privateColorRegisters: true (the default). iTerm2 works too, but renders SIXEL dot-by-dot, so on a HiDPI panel the image appears at half its nominal size.

For terminal authors — 256 + 128 exceeds the 256-register limit, so each trial is drawn as two SIXEL images at the same cursor origin: a 1280×480 image with P2 = 1 (transparent background) painting the right half, then a 640×480 image painting the left half, each with its own palette. If your terminal shares one register file across images and stores graphics as indexed pixels, the second image's palette redefinition will corrupt the first and your result will be meaningless.

Sanity check: the first trial should show two clean, essentially identical photographs. If either half is garbled, blocky, or vertically offset, stop — your terminal is not rendering the test correctly, and that is itself worth reporting.

What to report

The score line, plus the two things that decide whether the difference is even resolvable for you:

1. Angular resolution (ppd) — or the raw numbers: the physical width of the 1280 px image on screen, and your eye-to-screen distance.

   # usage: <image width in px> <measured width in mm> <viewing distance in mm>
   python3 -c 'import math,sys; px,w,d=map(float,sys.argv[1:4]); print("%.1f ppd" % (1/math.degrees(2*math.atan(w/px/2/d))))' 1280 300 600

2. Visual acuity, corrected as you ran the test — decimal, Snellen or logMAR.

And, because my own data says it is worth more than ten points: how you did the task. Were you judging which picture looked nicer, or hunting for a specific artefact? If you work on image processing, say so.

Useful extras:

• Display model, panel type, and bit depth — true 8-bit versus 6-bit + FRC changes dither visibility a lot
• Colour gamut and profile, plus any night-shift or colour-management filter that was active
• Display scaling, which cuts both ways: fractional or 2× scaling resamples the image and can smear the dither, while a terminal rendering dot-by-dot on a HiDPI panel shows the image at half its nominal size and raises your effective ppd. Measure the on-screen width rather than assuming.
• Ambient lighting, OS, terminal emulator and version

Script:          p128 / p64 / p32 / p16
Order:           which ones you ran, in what order
Strategy:        judged appearance / hunted for the artefact
Result:          256colors: __ / ______: __ / timeout: __
ppd:             __ ppd   (1280 px = ___ mm on screen, viewing distance ___ mm)
Visual acuity:   ___ (corrected / uncorrected)
Background:      do you work on image processing / quantisation?
Display:         model, panel type, bit depth
Colour:          gamut / profile / filters active
Scaling:         1:1 or ___ %
Ambient light:   dark / dim / office / bright
Terminal:        emulator + version, OS
Notes:           anything you noticed

How to read your score

Chance is 50. Exact two-sided binomial test, with the Clopper–Pearson 95 % interval for your true rate:

Score / 100	two-sided p	95 % CI	Reading
35	0.0035	25.7 – 45.2 %	clear preference for the reduced image
39	0.035	29.4 – 49.3 %	reliable preference for the reduced image
45	0.37	35.0 – 55.3 %	indistinguishable
50	1.00	39.8 – 60.2 %	indistinguishable
55	0.37	44.7 – 65.0 %	indistinguishable
60	0.057	49.7 – 69.7 %	borderline
61	0.035	50.7 – 70.6 %	reliable preference for 256
65	0.0035	54.8 – 74.3 %	clear preference for 256
70	0.0001	60.0 – 78.8 %	strong preference for 256

Anything inside 40–60 is what "I cannot tell these apart" looks like. A single participant can only bound their own rate to about ±10 points, so pooling is where the answer comes from: ten people is a thousand trials, enough to resolve a rate as small as 53 %.

Read the table in both directions. A score meaningfully below 50 is a real possible outcome, not noise — see the prior art below.

Prior art: this has essentially been done before

Doolittle MH, Doolittle KW, Winkelman Z, Weinberg DS. Color images in telepathology: how many colors do we need? Hum Pathol. 1997;28(1):36–41. doi:10.1016/S0046-8177(97)90276-8 (PMID 9013829)

Thirty pathologists each judged 30 image pairs of microscopy images: a 24-bit truecolor original against an 8-bit (256-colour) version produced with an adaptive colour reduction algorithm plus diffusion dithering — essentially what img2sixel does now. Two questions per pair: can you detect a difference, and, regardless, which is better?

• Conscious detection of a difference was not statistically significant.
• When forced to choose, there was a significant preference for the 8-bit images (p < .005).

On diagnostic microscopy — content where you would expect colour fidelity to matter most — 256 colours with a good quantiser and a dither was not merely indistinguishable from 16.7 million, it was preferred. The authors attribute this to the apparent crispness a dithered, reduced palette gives an image.

That is a 1997 result on a 1997 display and says nothing directly about SIXEL. But it means "more colours is obviously better" has already failed an experimental test once, and it tells you what to watch for here: a dithered, reduced palette adds high-frequency structure that reads as extra acutance, and an image with more apparent local contrast can genuinely look better in a forced choice. If several participants land in the 35–45 band, that is a finding in its own right.

What would change my mind

Angular resolution is the obvious moderator, and the arithmetic is simple: the finest structure an error-diffusion dither can produce alternates every pixel, so its fundamental frequency is half your display's ppd.

Setup	ppd	finest dither structure
96 ppi 1080p monitor, 500 mm	33 ppd	16.5 cycles/deg
14″ MBP XDR via XQuartz at 2×, 500 mm	43.6 ppd	21.8 cycles/deg
27″ 1440p (~109 ppi), 600 mm	45 ppd	22.5 cycles/deg
27″ 4K (~163 ppi), 600 mm	67 ppd	33.6 cycles/deg
14″ MBP XDR 1:1, 500 mm	87 ppd	43.6 cycles/deg

Human contrast sensitivity peaks around 3–6 cycles/deg and falls off steeply above ~20, especially for low-contrast differences. Both of my measurement points, 43.6 and 87.3 ppd, sit above that knee and both are null — consistent with the geometry explanation, but they do not test it, because the interesting region is lower.

The measurement I want is somebody on a conventional, non-HiDPI monitor at around 33 ppd. If the difference is visible there and not on Retina panels, we know exactly which viewing conditions the palette-size argument applies to. A score of 65+ on a low-ppd display would be far more informative than ten more nulls from Retina laptops. Please run this on your least impressive monitor.

A result on non-photographic content would change my mind faster still, and I do not have that test yet.

Limitations

• This measures preference, not raw discriminability, and they are not the same number. Someone who can see a difference but has no consistent preference lands near 50 either way. Doolittle et al. separated them by asking both questions per pair. In my own runs the gap is twelve points at 64 colours: 48 judging appearance, 60 hunting the artefact. If you switch to hunting, you are answering the other question — worth doing, but say which you did. A proper ABX or triangle design would measure discriminability cleanly; I can build one if there is appetite.
• Photographs only. Gradient ramps, flat-shaded CG, UI screenshots and scientific colour maps are exactly where a small palette hurts most, and they are deliberately absent. A null here bounds the photographic case and nothing else.
• Dithering is on, using lso2. This says nothing about undithered output or about the old median-cut path.
• Content is fixed at 640×480. A larger image gives the quantiser more pixels but not more colours, so the effect may differ at 1920×1080.
• The image set was selected without reference to the outcome variable. Candidates were filtered on licence, source dimensions (1400–3600 px wide, so a 1:1 centre crop frames comparably), mean saturation (excluding monochrome photographs, which carry almost no chroma-quantisation signal), entropy, and a cap of two images per photographer; non-photographic content was then removed by eye. Nothing was selected or rejected on the basis of how much the two encodings differ, since that is the quantity under test.

Where this goes

I will collate results in this thread and post a summary back to #196. If the 128-colour null replicates across displays and observers, then for photographic content the encoder's 256-colour ceiling is not the binding constraint, and the case for 1024 or 4096 has to be built on the content types this test deliberately excludes. Defining those cases concretely, and measuring them, seems to me the useful next step — and I would rather the palette-size decision rest on that than on anyone's intuition, mine included.

---

Image credits — 100 photographs from Wikimedia Commons Featured pictures (click to expand)

Every image has been modified: centre-cropped to 640×480 at 1:1 (no resampling) and re-encoded with a reduced colour palette. No other alteration was made. The modified versions are distributed under the licence of the original; adaptations of CC BY-SA 2.0 / 2.5 / 3.0 originals are distributed under CC BY-SA 4.0, as permitted by the "later version" clause of those licences. The images are provided without warranties, as set out in Section 5 of the applicable Creative Commons licence.

The shell code of the test scripts is under libsixel’s own licence (MIT). The photographs are included as a collection, not as an adaptation of that code, so the CC BY-SA images do not place the code under CC BY-SA.

The same attribution is embedded in the header of every test script, so it travels with the file.

#	Image	Author	Licence
1	001 Humpback whale breaching in Ballena Marine National Park Photo by Giles Laurent.jpg	Giles Laurent	CC BY-SA 4.0
2	001 Volcano eruption of Litli-Hrútur in Iceland in 2023 Photo by Giles Laurent.jpg	Giles Laurent	CC BY-SA 4.0
3	02.Trinidad (59).JPG	Elemaki	CC BY-SA 3.0
4	04-09-12-Schaupflügen-Fahrenwalde-RalfR-IMG 1232.jpg	Ralf Roletschek	CC BY-SA 2.5
5	13-06-07 RaR MC Fitti 03.jpg	Achim Raschka (talk)	CC BY-SA 4.0
6	1928 Model A Ford.jpg	Richard Smith	CC BY 2.0
7	2 x Saab JAS-39 Gripen and 2 x Aero L-159 of the Czech Air Force, based at AFB Čáslav, inflight.jpg	Milan Nykodym	CC BY-SA 2.0
8	2004-tsunami.jpg	David Rydevik (email: david.rydevikgmail.com), Stockholm, Sweden.	Public domain
9	2007 Matterhorn.jpg	Alps	Public domain
10	2007 Nagaoka Festival 004 Edit.jpg	Kropsoq	CC BY-SA 3.0
11	20070818-0001-strolling reindeer.jpg	Alexandre Buisse (Nattfodd)	CC BY-SA 3.0
12	2009K4760 - Жовква (Львівська).jpg	Мирослав Видрак	CC BY-SA 4.0
13	2010-kodiak-bear-1.jpg	Yathin S Krishnappa	CC BY-SA 3.0
14	2011-03-09-fort-du-lomont-10.jpg	Thomas Bresson	CC BY 3.0
15	2012.06.07.-001-Erlichsee Oberhausen-Rheinhausen-Spitzenfleck-Maennchen im Flug-crop.jpg	Andreas Eichler	CC BY-SA 3.0
16	2012.08.10.-5-Mannheim Vogelstang-Kleiner Kohl-Weißling-Paarung-crop.jpg	Andreas Eichler	CC BY-SA 3.0
17	2014 Modding komputera 01.jpg	Halicki	CC BY 3.0
18	2016 Falcon 9 at Vandenberg Air Force Base.jpg	SpaceX	CC0
19	20160805 Inle Lake 7434.jpg	Jakub Hałun	CC BY-SA 4.0
20	2017 - Київ - Місячний вечір на Замковій горі.jpg	Moahim	CC BY-SA 4.0
21	20170624 Schneeberglandrallye DSC 7444.jpg	Granada	CC BY-SA 4.0
22	20180105 Men's handball Austria - Czechia Janko Bozovic 850 9203.jpg	Granada	CC BY-SA 4.0
23	2019-11-24 Men's World Cup at 2019-20 Luge World Cup in Igls by Sandro Halank–031.jpg	Sandro Halank, Wikimedia Commons	CC BY-SA 4.0
24	2021-01-15 Eberspächer Luge World Cup Oberhof 1DX 4602 by Stepro.jpg	Steffen Prößdorf	CC BY-SA 4.0
25	25 Cassutt (N-6291N) Formula One Racer 2014 Reno Air Races photo Don Ramey Logan.jpg	Don Ramey Logan	CC BY-SA 4.0
26	28-090504-black-headed-bunting-at-first-layby.jpg	Mark S Jobling	CC BY 3.0
27	2909 vallon moy res.jpg	Berrucomons	CC BY-SA 3.0
28	30th St. Moritz Polo World Cup on Snow - 20140202 - Cartier vs Ralph Lauren 9.jpg	Clément Bucco-Lechat	CC BY-SA 3.0
29	360 degrees fogbow.jpg	Brocken Inaglory	CC BY-SA 3.0
30	4-14 Marines in Fallujah.jpg	Lance Corporal Samantha L. Jones, USMC	Public domain
31	4th FW Strike Eagles assist shuttle launch.jpg	Capt. John Peltier, USAF	Public domain
32	A bad sales day.jpg	Wilfredor	CC BY-SA 4.0
33	A Brazilian family in Rio de Janeiro by Jean-Baptiste Debret 1839.jpg	Jean-Baptiste Debret	Public domain
34	A bunch of rebar up close.jpg	W.carter	CC BY-SA 4.0
35	A chinstrap penguin (Pygoscelis antarcticus) on Deception Island in Antarctica.jpg	Christopher Michel	CC BY 2.0
36	A Larva of Papilio memnon 09sai.jpg	池田正樹 (talk)masaki ikeda	CC BY-SA 3.0
37	A member of the ATS (Auxiliary Territorial Service) serving with a 3.7-inch anti-aircraft gun battery, December 1942. TR452.jpg	Malindine E G (Lt), Tanner (Lt), War Office official photographer	Public domain
38	A New View of the Tarantula Nebula.jpg	NASA	Public domain
39	A sculpture at the entrance to the palace of Versailles.jpg	Jean-Baptiste Tuby / After Charles Le Brun	CC BY 2.0
40	A surfer at the wave edit.jpg	Brocken Inaglory edit by user:Kallerna	CC BY-SA 3.0
41	A Taraxacum Ruderalia dandelion clock.jpg	Richard Bartz, Munich Makro Freak	CC BY-SA 2.5
42	A USAF F-16 pilot breaking right on final approach over northern Las Vegas.jpg	Master Sgt. Benjamin Bloker	Public domain
43	A5 Aarebruecke.jpg	Ikiwaner	CC BY-SA 3.0
44	Abeja (Bombus terrestris) en un Ptilotus exaltatus, jardín botánico de Tallin, Estonia, 2012-08-12, DD 01.JPG	Diego Delso	CC BY-SA 3.0
45	Abies koreana (szyszki).JPG	Lestat (Jan Mehlich)	CC BY-SA 3.0
46	Acanthiza pusilla - Austin's Ferry.jpg	JJ Harrison (https://www.jjharrison.com.au/)	CC BY-SA 3.0
47	Acanthodoris lutea laying eggs 1.jpg	Brocken Inaglory	CC BY-SA 4.0
48	Acapulco fishermen.jpg	Tomascastelazo	CC BY-SA 3.0
49	ACE EMD F40PH Fremont - San Jose.jpg	Kabelleger / David Gubler (http://www.bahnbilder.ch)	CC BY-SA 3.0
50	Aceras anthropophora (detail).jpg	Hans Hillewaert	CC BY-SA 4.0
51	Acinonyx jubatus -Southern Namibia-8.jpg	Joachim Huber	CC BY-SA 2.0
52	Acorn woodpecker holding a nut in its beak-0225.jpg	Frank Schulenburg	CC BY-SA 4.0
53	Actinoscyphia aurelia 1.jpg	Image courtesy of Aquapix and Expedition to the Deep Slope 2007, NOAA-OE	Public domain
54	Actinostola.jpg	Expedition to the Deep Slope 2007, NOAA-OE	Public domain
55	Actitis hypoleucos.jpg	Ben Fredericson from Germany	CC BY 2.0
56	AD2009Aug07 Natrix helvetica 01.jpg	Bernie	Public domain
57	Adansonia digitata MS 6682.JPG	Marco Schmidt	CC BY-SA 3.0
58	Adansonia grandidieri04.jpg	Bernard Gagnon	CC BY-SA 3.0
59	Aegires gardineri Papua New Guinea by Nick Hobgood-001.jpg	Nhobgood	CC BY-SA 3.0
60	Aegithalos caudatus front-on 2.jpg	Francis C. Franklin	CC BY-SA 3.0
61	Aepyceros melampus petersi ♂.jpg	Hans Hillewaert	CC BY-SA 4.0
62	Aerial view of Portuguese calçada, Instituto Superior Técnico, Lisbon, Portugal julesvernex2.jpg	Jules Verne Times Two	CC BY-SA 4.0
63	Aeroflot Airbus A330 Kustov.jpg	Sergey Kustov	CC BY-SA 3.0
64	Aerogelflower filtered.jpg	Aerogelflower.jpg: No information. derivative work: JovanCormac	Public domain
65	Aeshna cyanea female 1.jpg	Böhringer böhringer friedrich	CC BY-SA 2.5
66	Aethaloessa calidalis by Kadavoor.jpg	Jeevan Jose, Kerala, India	CC BY-SA 4.0
67	Afgevallen kastanje van een paardenkastanje (Aesculus) 10-10-2021 (d.j.b.) 02.jpg	Dominicus Johannes Bergsma	CC BY-SA 4.0
68	African emerald cuckoo (Chrysococcyx cupreus) male.jpg	Charles J. Sharp	CC BY-SA 4.0
69	African Leopard Near Okevi Waterhole Etosha Namibia.jpg	Axel Tschentscher	CC BY-SA 4.0
70	African Spoonbill-001.jpg	Leo za1	CC BY-SA 3.0
71	Afternoon at Tennfjorden, Raftsundet, Hinnøya, Norway, 2015 September.jpg	Ximonic (Simo Räsänen)	CC BY-SA 3.0
72	Agapanthus 'White Heaven', ingetogen schoonheid van de ontluikende bloemknop. Locatie, Tuinreservaat Jonkervallei 03.jpg	Dominicus Johannes Bergsma	CC BY-SA 4.0
73	Agarplate redbloodcells edit.jpg	Bill Branson – (Edited by Fir0002)(Edited by Drhx)	Public domain
74	Agrigent BW 2012-10-07 12-24-45.JPG	Berthold Werner	CC BY-SA 3.0
75	Agriocnemis pygmaea female, Burdwan, West Bengal, India 11 10 2012.jpg	Joydeep	CC BY-SA 3.0
76	Agrius convolvuli korseby.jpeg	Kristian Peters -- Fabelfroh 15:17, 12 September 2007 (UTC)	CC BY-SA 3.0
77	Air Chief Marshal Sir Arthur Tedder on the Italian coast, 17 December 1943.jpg	British official photographer, minor restoration (sky and a small levels tweak only, no need to credit) by Adam Cuerden	Public domain
78	Air Force Fire Training.jpg	U.S. Air Force photo/Tech. Sgt. Rey Ramon	Public domain
79	Air France A380 F-HPJA.jpg	Joe Ravi	CC BY-SA 3.0
80	Aix galericulata Luc Viatour.jpg	Lviatour	CC BY-SA 3.0
81	Al quartiere latino, bozzetto di Adolf Hohenstein per La Bohème (1896) - Archivio Storico Ricordi ICON000086 - Restoration.jpg	Adolfo Hohenstein / Adam Cuerden	CC BY-SA 4.0
82	Albaicin 2012 San Nicolas Sacromonte.jpg	Jebulon	CC0
83	Albi cathedral - choir and choir screen.jpg	Pom²	CC BY-SA 3.0
84	Alcazaba Antequera.jpg	Jebulon	CC0
85	Alcedo atthis - Riserve naturali e aree contigue della fascia fluviale del Po.jpg	Luca Casale	CC BY-SA 4.0
86	Alemania vs Italia - 2014 CERH European Championship - 11.jpg	Carlos Delgado	CC BY-SA 4.0
87	Aleuria aurantia fungus orange peel (second view).jpg	Holger Krisp	CC BY 3.0
88	Alexander Nevsky Cathedral, Sofia (by Pudelek).JPG	Pudelek (Marcin Szala)	CC BY-SA 3.0
89	AlfredPalmerM3tank1942b.jpg	Alfred T. Palmer	Public domain
90	Alligator mississippiensis 2 babies.jpg	Ianaré Sévi	CC BY-SA 3.0
91	Allium 'Lucy Ball' Pink Flower Head 2236px.jpg	Photo by and (c)2008 Derek Ramsey (Ram-Man). Co-attribution must be given to the Chanticleer Garden.	CC BY-SA 3.0
92	Allium rothii 1.jpg	Gideon Pisanty (Gidip) גדעון פיזנטי	CC BY 3.0
93	Alpen Gelbling, Colias phicomone, Arnica montana 1.JPG	Böhringer Friedrich	CC BY-SA 2.5
94	Alphonse-Desjardins.jpg	Cephas	CC BY-SA 3.0
95	Alpinistes Aiguille du Midi 03.JPG	Benh LIEU SONG	CC BY-SA 3.0
96	Amanita muscaria qtl4.jpg	Quartl	CC BY-SA 3.0
97	Ambersweet oranges.jpg	unknown	Public domain
98	American Bird Grasshopper.jpg	http://www.birdphotos.com	CC BY 3.0
99	American robin (71307).jpg	Rhododendrites	CC BY-SA 4.0
100	An F-A-18C Hornet launches from the flight deck of the conventionally powered aircraft carrier.jpg	U.S. Navy photo by Photographer's Mate 3rd Class Jonathan Chandler	Public domain