Evidence · updated 8 October 2026
What is known about range and depth in scientific figures
This page gathers what the literature establishes, what the standards say, and what we have measured, each marked by how strong it is. It covers both halves of the name: the range a screen can show above white, and the depth — the number of levels — that the eye and the data each need. Citations were checked against Crossref and against the papers' own abstracts.
Summary
Established
People see about 3.7 orders of magnitude at once [1]. An 8-bit SDR figure clips much of that; HDR displays can show it.
Established
12-bit PQ is at or below the visibility threshold across 0.001 to 10,000 cd/m² [2]. Nothing beyond 12 bits helps the eye [3].
Measured
A 10-bit PQ step is 1.1 to 1.4 just-noticeable differences (DICOM GSDF); a 12-bit step is about 0.33.
Established
Evidence on more grey levels is mixed. A pilot study found 10-bit better than 8-bit [4]; a three-site study found no accuracy difference, only faster reading [19]. More luminance alone did not help [5]. An HDR prototype display gave small gains on hard cases [6].
Established
What people prefer does not predict what is scientifically useful (R² = 0.31) [7]. Preference alone is not evidence.
Measured
HDR figures display today on a phone and a desktop monitor, as gain-map JPEG, 16-bit PNG, AVIF and HDR video. PDF has no portable HDR yet [8].
Measured
A roll-off that flattens at the top cannot be read back. At 2× headroom it put 31.4 % of a CT's declared range into the last 3 % of brightness. The profile's curve now keeps rising.
Measured
On six real datasets, HDR helped most where the range above white was large (a gravitational wave at 327 × the noise) and added little where it was under 4 × (the M87 black hole).
Open
No study has shown that HDR helps people read scientific figures. Rendering scientific data for HDR displays was explored in 2005, on the prototype displays of the time [23]. We found nothing since for today's consumer HDR screens, and no standard for it.
Vision and bits
How much range and precision the eye can use
- Range. Adaptation spans about 14 orders of magnitude (starlight to sunlight), but the range seen at once is about 3.7 orders [1]. A single view on a display fits within that.
- Precision. The DICOM Grayscale Standard Display Function (Barten's model) puts the smallest visible luminance step at about 0.65 % above 100 cd/m², coarser in the dark (2.4 % at 1 cd/m²) [9].
- Encoding. PQ (SMPTE ST 2084) was designed so that 12-bit steps sit at or below threshold over 0.001–10,000 cd/m² [2]. An evaluation of HDR colour encodings found about 11 bits visually lossless [3].
- Colour. BT.2020 primaries cover about 76 % of visible chromaticities and almost all real surface colours.
| PQ bits | Step, in GSDF JNDs (10–3000 cd/m²) | Near black (0.5–3 cd/m²) | Reading |
|---|---|---|---|
| 8 | 4.5–5.7 | 3.4 | visible banding |
| 10 | 1.1–1.4 | 0.82 | at threshold; dither smooth gradients |
| 12 | 0.27–0.35 | 0.20 | below threshold |
Computed by udeep.jnd (tested against the GSDF's published end points). The GSDF target is
not the most sensitive pattern, so these are lower bounds on visibility.
Standards
What already exists, and what it doesn't cover
| Standard | What it gives | Status |
|---|---|---|
| ITU-R BT.2100, SMPTE ST 2084 | PQ and HLG transfer functions, BT.2020 primaries | in use (TV, streaming) |
| ITU-R BT.2408 | Reference white at 203 cd/m²: the basis for "multiples of SDR white" | in use |
| ITU-R BT.2390 | The roll-off (EETF) used to fit content to a display. It flattens at the top, which suits pictures but cannot be read back to data | in use; the profile allows it for display rendering only |
| ISO 21496-1:2025 | Gain maps: an SDR image plus a map that HDR displays apply up to their headroom (ISO/TC 42, July 2025) [18] | published |
| ITU-T H.273 | Colour signalling code points (as in the PNG cICP chunk) | in use |
| DICOM PS3.14 | A perceptually linear greyscale display function for medicine | in use; no HDR above white |
| CSS Color HDR (W3C) | HDR colours and dynamic-range-limit for the web | editor's draft; the property ships, the colours do not |
| PDF (ISO 32000) | HDR only as "private data" that conforming viewers must ignore; JPEG XL chosen for future HDR [8] | no portable HDR yet |
| Scientific figures | How data becomes brightness, the knee, the SDR fallback, verification | nothing, which is what the draft profile addresses |
Prior research
Nearby work, and how this differs
- HDR data on ordinary (SDR) displays: the opposite direction. Visibility-matching [10] and display-adaptive [11] tone mapping compress a scene's range for a given display. Park & Montag compared nine tone-mapping operators on scientific images; the best depended on the data type [7]. Zhou et al. combined tone mapping with simulated glare to flag high values and found the glare preattentive [12]. All of these keep the output within SDR.
- Colour for science. Colour maps should be perceptually uniform and readable by everyone [13]. Jzazbz is a uniform colour space that extends to HDR [14]; UDeep's HDR colour maps are built in it. Astronomy's asinh stretch shows a wide range in 8 bits [15].
- Comparing figures. Juxtaposition, superposition and explicit encoding [16]. UDeep's fair HDR/SDR toggle is a juxtaposition that holds the SDR part identical.
- Measuring visible differences. ColorVideoVDP predicts visible differences with a model of the display and viewer [17].
- Scientific data on HDR displays: the closest prior work. Ghosh, Trentacoste and Heidrich rendered volume data for HDR displays in 2005, when such displays were research prototypes [23]. HDR VolVis (2006) handled volume data with a high range, for ordinary displays [24]. Vision scientists have built lab HDR luminance displays as instruments [25].
The difference: that early work targeted prototype hardware and one kind of rendering. None of it defines how a scientific figure should carry its range on today's consumer HDR screens, with a declared mapping and an SDR figure of record, and no file profile exists for it. The search so far covered the web, a scholarly full-text index and GitHub (September 2026). It is not yet systematic, so "to our knowledge" is the right wording.
Does it help people?
Observer studies with more range or more bits
| Study | Change | Task | Result |
|---|---|---|---|
| Hemminger et al. 1999 [5] | peak luminance 34 to 2056 cd/m² (film on a lightbox) | 20 observers; simulated masses in mammograms | no significant effect of luminance range |
| Tabár et al. 2010 [4] | 8-bit vs 10-bit greyscale displays | 10 radiologists, 33 screening studies (pilot) | Az 0.791 → 0.831, p = 0.016 |
| Krupinski et al. 2007 [19] | 8-bit vs 11-bit greyscale displays | 18 radiologists at 3 sites, 100 chest images | no difference in Az (0.828 vs 0.825); reading faster at 11-bit |
| Marchessoux et al. 2016 [6] | dual-LCD HDR prototype vs a standard medical display | 8 radiologists, 214 chest images, simulated nodules | figure of merit +0.09 % overall, +1.38 % on difficult nodules; >30 % more true positives in the darkest regions |
| Park & Montag 2007 [7] | nine tone-mapping operators (SDR output) | paired comparisons; preference vs "scientific usefulness" | the two judgements correlate weakly (R² = 0.31) |
Reading: extra precision has helped in one pilot study and made no accuracy difference (but faster reading) in a larger one; extra luminance alone has not; a true HDR display helped a little, mostly in dark regions. None of these studies used scientific figures, so a task-based study is the missing piece. Marchessoux's abstract gives no significance test for its differences.
Medical imaging today
The data has 12–16 bits; most of the display chain shows 8
| Part of the chain | What it holds or shows | Source |
|---|---|---|
| The data | CT, MR and DX: 12–16 bits stored in 16. Windowing (Window Center/Width, VOI LUT) maps it down to display levels | DICOM PS3.3 [20] |
| Diagnostic displays | 10-bit input (1,024 simultaneous greys) through 12–14-bit internal look-up tables, for example Barco Coronis and EIZO RadiForce | vendor specifications |
| Graphics output | 10-bit through OpenGL 30-bit, DirectX 10-bit/FP16, or AMD Pro. 12-bit only through a legacy NVIDIA packed-pixel DVI mode | NVIDIA, Microsoft documentation |
| Viewers | Documented >8-bit output is rare. Of 18 workstations tested, only 3 of the 14 hospital ones passed a hardware/OS 10-bit test, and only one (Sectra) passed in the viewer too; RadiAnt states 8-bit output. For OHIF, DWV, Weasis, Horos/OsiriX and 3D Slicer, no 10-bit support was found | Chrzan & Urbanik 2018 [21]; vendor statements |
| Guidelines | ACR–AAPM–SIIM (2012): no evidence that more than 8 bits changes interpretation; 10-bit needs the OS, software, graphics card and monitor to support it | [22] |
| HDR | No DICOM viewer renders above reference white. One research prototype display (2016) exists | [6] |
So CT windowing is the Extend case in daily practice: the data holds two ranges, and the display shows one at a time. An HDR mode could keep the chosen window unchanged and show what lies above it. Diagnostic use of any new display mode is regulated, and would need clinical evidence first.
What we measured
Model predictions and engineering results
Perceptual results are ColorVideoVDP predictions (in JOD, just-objectionable differences) with the display modelled, gated by null and positive controls. They are predictions, not observer data.
| Finding | Result | Strength |
|---|---|---|
| A fair comparison must keep the SDR part identical | The white-point and stretch choice alone was 2–7× more visible than HDR's contribution | model, several datasets |
| Highlight mode (sparse extremes above white) | A small, visible gain: −0.3 to −0.8 JOD on a 500 cd/m² monitor model; 10–60 % larger on an OLED TV model | model, 9 datasets |
| Extend mode (a second range above white) | One CT slice: soft tissue matched its dedicated window (4.75 vs 4.63), and bone about matched a bone window (8.62 vs 8.65). SDR has to choose one. | model, one slice |
| HDR cannot restore lost data | A clipped astronomical mosaic gave the smallest effect (−0.15 JOD); the brightest solar pixels were saturation and blooming, which must be masked | model + data inspection |
| A smooth curve beat a hand-set second window | The BT.2390-style roll-off scored better than a second window set by hand. This comparison used the earlier, flat-topped curve and has not been repeated with the current one | model, one slice |
| A flat-topped roll-off cannot be read back to data | At 2× headroom, 31.4 % of the CT's declared range (39.1 % of the range above the knee, from 1220 HU up) fell in the last 3 % of brightness. Replaced by a cubic curve whose slope never falls below a set fraction of the average slope, so every value keeps its own brightness (profile 6.3) | computed, tested |
| Precision has two budgets | One is for the eye (the table under Vision and bits). The other is for the data: the rounding step, traced back to data units, should stay under a tenth of the noise. An 8-bit gain-map JPEG meets neither above white, so the profile calls it a rendering, not a record | computed |
| Do the colour maps survive colour-vision deficiency? | Lightness was checked at every point of both maps under simulated protanopia, deuteranopia and tritanopia [28]. The first version of the maps had small faults: ice dipped for red-green deficiencies near 15 % of white, each map fell at one point at 2× headroom, and the brightest colour of both was about 0.1 % short of white. Every fall was under 0.05 % of the lightness range. The second version (fire/2, ice/2, 10 October 2026) has none: lightness rises at every point for normal vision and all three deficiencies, and the top is exactly white. 83–94 % of fire's perceptual length survives a deficiency, and 80–98 % of ice's. A third version (fire/3, ice/3, 12 October 2026) is the second with its values below a ten-millionth of white set to zero, so that the tables can be made again bit for bit on any machine; no colour moves by more than that. The earlier versions stay available so older files can be read | computed, simulated vision |
| How far an 8-bit gain-map JPEG is from carrying the data | On the synthetic CT (noise 6 HU, range −160 to 1850 HU, 4× headroom), one stored level spans up to 19.5 HU. Rounding is 0.94 of the noise at its worst, against a limit of 0.1, and it is over the limit across 92 % of the range. The writer now reports this when it saves | computed, synthetic data |
| Gain maps overshoot at sharp edges | JPEG error in the base image, multiplied by the gain, pushed a colour bar's edge 20–27 % past the headroom. Fixed: compute the gain against the decoded base and store the gain map at top quality. Residual ≤5 %. | measured, tested |
Real data
What six real datasets taught
From building the gallery. Each figure's numbers are printed by the build and kept in the repository.
| Dataset | What happened | Lesson |
|---|---|---|
| GW150914 (LIGO) | A knee set from the noise statistics failed its own check on raw strain: noise crossed it 45 times in 9,984 tiles, where one was expected. After whitening, 92 tiles exceeded the one-in-100 level against 92.2 expected | A statistical knee is only as good as the noise model. Check it on noise-only data, and stop the build if it fails |
| GW250114 (LIGO) | The signal peaks at 327 × the mean noise power | The strongest case: a faint noise floor and a very bright signal in one figure |
| Chest CT (LIDC-IDRI) | The scan is contrast-enhanced: 89 % of the pixels above the knee are blood at 240–500 HU, not bone. Covering bone to 1900 HU left the blood crowded near white; the range was cut to 700 HU | Look at what is actually above the knee. The declared range is a trade-off, and the figure should state it |
| The Sun at 171 Å (SDO/AIA) | The brightest pixels are detector saturation | Mask saturated pixels. They must not be lifted above white |
| M51 (Hubble) | The file has no data-quality layer, so saturated foreground stars could not be masked | Say so on the figure when a mask is not possible |
| M87* (Event Horizon Telescope) | The image spans less than 4 × above its white. An SDR figure scaled to the peak shows the same structure | HDR is not always worth it. A small range fits in SDR |
Devices
What displays HDR today
Tested on an iPhone 16 Pro (Photos and Safari) and a desktop monitor (Edge 153, Windows 11), September 2026, with the device checks. Each test had an SDR control and a positive control.
| What | Result | Consequence |
|---|---|---|
| Gain-map JPEG in iOS Photos | HDR | Shareable now |
| Gain-map JPEG, 16-bit PNG (PQ), AVIF (PQ, HLG), HEVC video (PQ, HLG) on a web page | HDR | Static and moving figures both work |
| Gain-map JPEG inside inline SVG | HDR | Vector axes around an HDR data layer |
Gain-map JPEG inside a standalone .svg file | SDR only | SVG files need an AVIF PQ layer, which works |
| Vector colours above white | not supported | No pure-vector HDR yet |
| 10/12-bit AVIF ramps | banded | The same values in 16-bit PNG looked smooth, so the decoder loses precision |
| Blind banding test (16-bit PNG rounded to 10–16 bits) | all smooth | Ran without a coarse positive control. A repeat with 8/9-bit strips is pending |
| 12-bit HEVC / AV1 video playback | pending | Test built; result not yet in |
| Panel bit depth, any iPhone | not published | No official figure or independent measurement exists; Apple states "true 10-bit" only for the Pro Display XDR |
Comfort and accessibility
What guidance exists for bright content
- Flashing is limited; steady brightness is not. WCAG 2.3.1 and ITU-R BT.1702 limit flashes to three per second. A review of international guidelines found only two that address HDR, and both concern flashing [26]. We found no numerical limit on static brightness or bright area.
- The agreed mitigation is user control. The CSS HDR draft says that "some individuals may have a sensitivity to very bright colors, so user agents should provide a mechanism to limit the maximum luminance at user option" [27].
- Bright areas are assumed to be small. HDR production practice reserves levels above reference white for small highlights.
So the profile forbids flashing above white, requires that switching between HDR and SDR is started by the user, and asks pages to offer a limit. Every page on this site has a "Limit HDR" control. A static HDR figure is not a flash hazard; an animation or video can be.
How to test
Lessons from getting it wrong
- Put a reference in the image. CT figures at 2× white were judged "not HDR" on a screen that clearly showed HDR for a pattern with a 1× white band beside the brighter squares.
- Use controls. An SDR control must not glow, and a positive control must glow. Without them, "everything looks smooth" cannot be told apart from "the test cannot show bands".
- Blind the banding test. Shuffle the strips and keep the key.
- Verify the stimulus. Decode every file before showing it. One common image library silently reads 16-bit colour PNGs as 8-bit.
- View locally. Remote-desktop sessions deliver 8-bit SDR.
- Record conditions: device, OS and browser versions, brightness, and room light.
Open questions
What would settle it
- A task study: can people find a faint feature, or read a value, faster or more accurately with HDR figures than with SDR?
- More devices in the evidence table: Android, Windows laptops, OLED TVs, Macs. Send a report.
- Whether steady brightness above white is uncomfortable, at what level and area. No guideline gives a number.
- The model comparisons above, repeated with the current curve.
- Whether 12-bit video plays, and whether decoders can keep 12 bits for stills.
- Gain-map AVIF, and a CSS media-query SDR fallback inside SVG. Both are untested.
- A systematic literature search before any claim of novelty.
References
- Kunkel, T. & Reinhard, E. (2010). A reassessment of the simultaneous dynamic range of the human visual system. Proc. APGV, 17–24. doi:10.1145/1836248.1836251
- Miller, S., Nezamabadi, M. & Daly, S. (2013). Perceptual signal coding for more efficient usage of bit codes. SMPTE Motion Imaging J. 122(4), 52–59. doi:10.5594/j18290
- Boitard, R., Mantiuk, R. K. & Pouli, T. (2015). Evaluation of color encodings for high dynamic range pixels. Proc. SPIE 9394, 93941K. doi:10.1117/12.2077715
- Tabár, L. K., Niño, A. S., Schreiber, E. et al. (2010). Three digital mammography display configurations: observer performance in a pilot ROC study. Digital Mammography (IWDM 2010), LNCS, 280–287. doi:10.1007/978-3-642-13666-5_38
- Hemminger, B. M., Dillon, A. W., Johnston, R. E. et al. (1999). Effect of display luminance on the feature detection rates of masses in mammograms. Medical Physics 26(11), 2266–2272. doi:10.1118/1.598740
- Marchessoux, C., de Paepe, L., Vanovermeire, O. & Albani, L. (2016). Clinical evaluation of a medical high dynamic range display. Medical Physics 43(7), 4023–4031. doi:10.1118/1.4953187
- Park, S. H. & Montag, E. D. (2007). Evaluating tone mapping algorithms for rendering non-pictorial (scientific) high-dynamic-range images. J. Visual Communication and Image Representation 18(5), 415–428. doi:10.1016/j.jvcir.2007.06.008
- PDF Association, Imaging Model TWG (2025). FAQ on HDR in PDF: Focus on PDF/A and PDF/X. pdfa.org
- NEMA. DICOM PS3.14: Grayscale Standard Display Function; Barten, P. G. J. (1999). Contrast Sensitivity of the Human Eye and Its Effects on Image Quality. SPIE Press.
- Ward Larson, G., Rushmeier, H. & Piatko, C. (1997). A visibility matching tone reproduction operator for high dynamic range scenes. IEEE TVCG 3(4), 291–306. doi:10.1109/2945.646233
- Mantiuk, R., Daly, S. & Kerofsky, L. (2008). Display adaptive tone mapping. ACM TOG 27(3). doi:10.1145/1360612.1360667
- Zhou, L., Rivinius, M., Johnson, C. R. & Weiskopf, D. (2020). Photographic high-dynamic-range scalar visualization. IEEE TVCG 26(6), 2156–2167. doi:10.1109/TVCG.2020.2970522
- Crameri, F., Shephard, G. E. & Heron, P. J. (2020). The misuse of colour in science communication. Nature Communications 11, 5444. doi:10.1038/s41467-020-19160-7
- Safdar, M., Cui, G., Kim, Y. J. & Luo, M. R. (2017). Perceptually uniform color space for image signals including high dynamic range and wide gamut. Optics Express 25(13), 15131. doi:10.1364/OE.25.015131
- Lupton, R., Blanton, M. R., Fekete, G. et al. (2004). Preparing red-green-blue images from CCD data. PASP 116(816), 133–137. doi:10.1086/382245
- Gleicher, M., Albers, D., Walker, R. et al. (2011). Visual comparison for information visualization. Information Visualization 10(4), 289–309. doi:10.1177/1473871611416549
- Mantiuk, R. K., Hanji, P., Ashraf, M., Asano, Y. & Chapiro, A. (2024). ColorVideoVDP: a visual difference predictor for image, video and display distortions. ACM TOG 43(4). doi:10.1145/3658144
- ISO 21496-1:2025, Digital photography: gain map metadata for image conversion, Part 1: dynamic range conversion. iso.org
- Krupinski, E. A., Siddiqui, K., Siegel, E. et al. (2007). Influence of 8-bit vs. 11-bit digital displays on observer performance and visual search: a multi-center evaluation. J. Society for Information Display 15(6), 385–390. doi:10.1889/1.2749324
- DICOM PS3.3, Information Object Definitions: C.8.2 (CT Image), C.11.2 (VOI LUT). dicom.nema.org
- Chrzan, R. & Urbanik, A. (2018). The assessment of diagnostic medical images using 10-bit grayscale: fact or myth? Polish Journal of Radiology 83, e127–e132. doi:10.5114/pjr.2018.75877
- ACR–AAPM–SIIM Technical Standard for Electronic Practice of Medical Imaging (2012 version). PMC3553359
- Ghosh, A., Trentacoste, M. & Heidrich, W. (2005). Volume rendering for high dynamic range displays. Fourth International Workshop on Volume Graphics. doi:10.1109/vg.2005.194102
- Yuan, X., Nguyen, M. X., Chen, B. & Porter, D. H. (2006). HDR VolVis: high dynamic range volume visualization. IEEE TVCG. doi:10.1109/tvcg.2006.72
- Hung, C. P., Callahan-Flintoft, C., Walker, A. J. et al. (2020). A 100,000-to-1 high dynamic range (HDR) luminance display for investigating visual perception. J. Neuroscience Methods. doi:10.1016/j.jneumeth.2020.108684
- Jordan, J. B. & Vanderheiden, G. C. (2024). International guidelines for photosensitive epilepsy: gap analysis and recommendations. ACM Transactions on Accessible Computing 17, 1–35. doi:10.1145/3694790
- W3C. CSS Color HDR Module Level 1 (editor's draft), Accessibility Considerations. drafts.csswg.org
- Machado, G. M., Oliveira, M. M. & Fernandes, L. A. F. (2009). A physiologically-based model for simulation of color vision deficiency. IEEE TVCG 15(6), 1291–1298. doi:10.1109/TVCG.2009.113