PaperlayerView exact revision on arXiv

Source-linked AI summary

Do Image Editors Follow Depth-Dependent Blur and Aperture Response? A Rendered-Ground-Truth Pilot Audit

Zhihan Chen, Yuhuan Zhao, Yijie Zhu, Xinyu Yao, Mengcong Ren, Yuchen Sun, Yunqing Chen

arXiv:2610.09344v1cs.CV

TL;DR

General image editors are often asked to simulate a wide aperture, but prior evaluations have not jointly measured blur proportions across known depths and aperture response. This pilot compares two editors with rendered thin-lens references and finds compressed depth proportions and attenuated aperture response, while qualitative wording changes blur strength more than f-number or pixel-size requests.

  • Problem

    Prior evaluations measure blur ordering, sharpness trends, effective-aperture error, or judged realism, but none reports depth proportion and aperture response separately at known depths.

    Source: §1 · p. 1

  • Method

    The audit renders Blender Cycles scenes with true thin-lens depth of field and applies one local blur-width estimator to oracle and editor outputs, with preregistered depth and aperture indices.

    Source: §1 · pp. 1–2, §3 · p. 3

  • Results

    0.99–1.18 editor versus 1.98–2.13 oracle f/1.4-to-f/2.8 width ratios, with editor near/far ratios of 1.18–1.25 for Gemini and 0.97–1.05 for GPT-image versus 1.69–1.82 for the oracle.

    Source: §1 · p. 1

  • Takeaways & Limitations

    On tested scenes, qualitative blur wording is a better strength request than an f-number or pixel size, and centre-focus layouts should not be expected to follow thin-lens depth proportions.

    Source: §5 · p. 7

    The pilot covers two editors and rendered scenes; whether the same behaviour appears in real captures with recorded distances remains untested.

    Source: §5 · pp. 6–7

Abstract

from arXiv · show

General image editors are asked to make a photo look as if it were taken at f/1.4, yet it is rarely checked whether the blur they add follows thin-lens optics. A physical aperture edit spreads blur across depth in thin-lens proportions and changes the blur when the aperture changes; prior evaluations check blur monotonicity, sharpness-trend correlation, effective-aperture error, or vision-language judgments, and none we found reports the two properties separately at known depths. In this pilot audit of two editors (Gemini~3.1 Flash Image and GPT-image-2.5) we compare against a rendered oracle: Blender Cycles scenes with true thin-lens depth of field, one blur-width estimator applied identically to oracle and editor outputs, and preregistered depth and aperture indices. In 24 texture scenes rendered in one three-panel geometry, accepted and measurable panels show f/1.4-to-f/2.8 width ratios, $\bar\sigma_{1.4}/\bar\sigma_{2.8}$, of 0.99--1.18 against 1.98--2.13 for the oracle, and ratios of pooled median Gaussian-equivalent near/far blur widths of about 1.18--1.25 (Gemini) and 0.97--1.05 (GPT-image) against 1.69--1.82. Preregistered black-box interventions show that qualitative wording changes blur strength by roughly 2--10 times, whereas a request for 2 versus 6 px changes it 1.1--1.3 times and no tested wording of the f-number meets the registered ``followed'' criterion. The depth compression appears in the original and reversed centre-focus layouts; with the focus on the near panel, the available-panel depth index reaches the registered threshold, and the aperture response stays attenuated in every layout tested. Scalar metrics adapted from published ones give oracle-like scores to synthetic editors whose proportions are compressed.

1 Introduction

The audit asks whether general image editors reproduce both the depth proportions and aperture response of thin-lens blur, properties not jointly measured at known depths by prior evaluations. It compares two editors with a rendered oracle and finds compressed depth proportions and attenuated aperture response in the tested setting.

1 Introduction

  • 0.99–1.18 for the editors versus 1.98–2.13 for the oracle on f/1.4-to-f/2.8 width ratios; near/far ratios are 1.18–1.25 for Gemini and 0.97–1.05 for GPT-image versus 1.69–1.82 for the oracle.The depth compression also appears in original and reversed centre-focus layouts; near-focus available-panel indices meet the registered threshold, while aperture response remains attenuated.Source: §1 · p. 1
  • Ten preregistered interventions found qualitative wording changed blur strength substantially, while four f-number wordings and numeric blur size did not meet registered criteria.The interventions used six scenes each.Source: §1 · p. 1
  • A physical aperture edit doubles ideal defocus diameter from f/2.8 to f/1.4 at off-focus depths, while relative blur widths depend on scene geometry and focus distance.For fixed scene conditions, blur scales with 1/N and with |1/z−1/zf| across depth.Source: §1 · p. 1
  • Blender Cycles renders true thin-lens depth of field for three textured panels, and the same local blur-width estimator measures oracle and editor outputs against the all-in-focus input.The middle panel is in focus; oracle renders use f/1.4, f/2.8, and f/5.6.Source: §1 · p. 1
  • The audit separates depth proportion, aperture response, and blur strength against a rendered oracle; this report is a pilot covering two editors and rendered scenes.Transfer to real photographs and broader editor coverage is left to the next stage.Source: §1 · p. 2

2 Related Work

Prior bokeh and editing evaluations measure different properties, including ordering, global sharpness trends, effective-aperture error, and judged realism. This audit uses rendered ground truth to report depth proportion and aperture response separately, alongside blur strength.

2 Related Work

  • Existing evaluations test aperture ordering [16], global sharpness-trend correlation [2], effective-aperture error, or vision-language judgments [3, 11], but do not separate near/far blur proportion from overall response.They also do not report blur strength relative to a rendered oracle at known depths.Source: §2 · p. 2
  • Bokeh Diffusion conditions on physical defocus [2], while Shrivastava et al. condition on EXIF aperture [16]; CamEdit edits aperture, focal plane, and shutter speed from text.Other methods target refocusing, circles of confusion [5], lens-style rendering, or paired real-image data [14].Source: §2 · p. 2
  • Unlike blind defocus estimators and depth-from-defocus methods, this audit uses a reference-based estimator because each scene’s all-in-focus render is known.Its endpoints measure depth-dependent Gaussian-equivalent width ratios and oracle-normalized aperture response.Source: §2 · p. 3
  • Rendered-ground-truth audits include editor comparisons with light-probe captures and checkerboard-based camera readouts [13], whose rendered cameras lack depth of field.Vision-language realism benchmarks [11], Blender-ground-truth spatial editing, and human-preference bokeh tests address other evaluation settings.Source: §2 · p. 3

3 Audit Protocol

The audit compares two editors with true thin-lens renders using a shared blur-width estimator and preregistered depth and aperture endpoints. Its protocol specifies the rendered scenes, editor prompts, measurement controls, and decision thresholds.

3 Audit Protocol

  • The protocol comprises rendered scenes with a thin-lens oracle, a panel blur-width instrument, and registered endpoints with decision rules.Source: §3 · p. 3
  • The 24-scene oracle uses three textured panels at 1.5, 2.2, and 3.0 m, with the centre panel in focus and true thin-lens renders at f/1.4, f/2.8, and f/5.6.Each scene has two content layouts to counterbalance texture across depth.Source: §3 · p. 3
  • Each editor receives prompts at three requested apertures, including a literal distance-from-focus blur cue; one output is retained per prompt-aperture cell.The cue is not physically exact because thin-lens blur is linear in 1/z, not distance.Source: §3 · p. 3
  • The shared instrument aligns each output to the all-in-focus reference and estimates panel width as the median measurable Gaussian blur, requiring at least 30 measurable positions.Alignment and window-correlation cut-offs determine which panels and positions are accepted.Source: §3 · p. 3
  • The preregistered indices measure depth proportion and aperture response against oracle behaviour, with replicated defined by an upper interval bound at most 0.5 and compliant by a lower bound at least 0.7.Intervals bootstrap 24 scenes; f/5.6 is excluded because the far-panel oracle width is near the estimator floor.Source: §3 · p. 3

4 Results

The pilot audit tests whether the editors reproduce thin-lens blur proportions and aperture response, first validating its measurement against synthetic controls. Editors show compressed depth proportions and attenuated aperture response in the main geometry, with layout-dependent depth results and prompt-sensitive blur strength.

4.1 Measurement controls

  • 0.98 and 0.97 are the oracle-like control’s depth and aperture indices at native size; uniform monotonic blur scores near zero for depth and near one for aperture.The forced 1.2 near/far ratio returns a depth index of 0.31 natively and 0.29 at Gemini’s output size; half-strength recovery varies with delivery size.Source: §4.1 · p. 4, Figure 2 · p. 5

4.2 Depth proportion

  • 1.18 and 1.05 are the editors’ largest pooled near/far ratios at f/1.4, versus 1.82 for the oracle; depth indices are 0.33–0.39 for Gemini and 0.08 to −0.02 for GPT-image.All four editor depth indices meet the registered replicated cutoff for accepted and measurable panels.Source: §4.2 · p. 4, Table 1 · p. 5

4.3 Aperture response

  • 0.99–1.18 is the editors’ f/1.4-to-f/2.8 width ratio, versus 1.98–2.13 for the oracle; aperture indices range from 0.01 to 0.21.Even assigning missing panels their most favourable widths leaves the largest aperture upper bound at 0.33.Source: §4.3 · p. 4

4.4 Sensitivity to missing panels

  • Gemini’s missing-panel worst-case depth upper bounds rise to 0.53 and 0.54, above the registered cutoff, while remaining below the 0.7 compliance cutoff.Its categorical depth classification is sensitive to missing panels; GPT-image’s depth result and both editors’ aperture results remain below the cutoff.Source: §4.4 · p. 4

4.5 Layout and focal plane

  • Reversed centre-focus layouts retain depth compression, but near-panel focus yields compliant depth indices of 1.18 for Gemini and 1.02 for GPT-image; aperture indices remain low.The result is based on accepted measurable panels, with GPT-image’s depth index resting on 16 centre panels and Gemini’s on 42.Source: §4.5 · p. 4, Table 2 · p. 6

4.6 What changes the blur

  • −0.06–0.30 for GPT-image and 0.08–0.48 for Gemini are the aperture indices across four f-number wordings; 2 versus 6 px requests change widths only 1.1–1.3 times.Text-only distance changes produce intermediate registered outcomes; extreme stated distances raise both GPT-image side-panel widths and Gemini’s left/right ratio.Source: §4.6 · pp. 4–5
  • 1.9–3.0 times is the side-panel width change for “very slight” versus “very strong” blur, and “very shallow” versus “very deep” wording changes widths roughly 3–10 times.Under deep wording both editors return almost-sharp panels, at 0.0–0.5 px.Source: §4.6 · p. 5

4.7 Do scalar metrics see this?

  • 59–67% of aperture pairs are ordered over the whole frame and 57–59% over side panels, while monotonicity and trend correlation give oracle-like scores to uniform-blur and forced-1.2-ratio controls.The global strength error is small for those controls and Gemini at f/2.8, despite Gemini’s compressed depth index of 0.33.Source: §4.7 · p. 6

5 Limitations and Conclusion

In the tested three-panel setting, the editors’ blur proportions and aperture response fall short of the rendered thin-lens oracle. The authors bound these findings by the pilot’s measurement and sampling limits and outline broader tests.

Limitations

  • The audit uses three fixed panel distances and flat panels; layout, focal-plane, and intervention tests cover only 6–12 scenes each.The estimator recovers Gaussian blur accurately above 1.5 px, but weak blur below about 1 px is unreliable and its absolute Gaussian-equivalent scale is uncertain by about 25%.Source: §5 · p. 6

Conclusion

  • Table 4 compares adapted published-style measures—aperture monotonicity, trend correlation, and global blur error—with depth and aperture indices.The conclusion says these scalar measures partly miss the distinct failures made visible by separating proportion, response, and strength.Source: Table 4 · p. 7, §5 · p. 7

Limitations

  • One retained output per cell and scene-bootstrap intervals do not capture generation variability; f/5.6 is descriptive because it reaches the estimator floor.The panels are photographs mounted on boards in a rendered room, and transfer to real captures with recorded distances is untested.Source: §5 · p. 7, Figure 3 · p. 7

Conclusion

  • On the tested scenes, qualitative blur wording is a better way to request stronger or weaker blur than an f-number or pixel size.In the tested centre-focus layouts, blur distribution across depth should not be expected to follow thin-lens proportions.Source: §5 · p. 7
  • The editors’ f/1.4-to-f/2.8 width ratios are 0.99–1.18 versus 1.98–2.13 for the oracle, and near/far width ratios are 0.97–1.25 versus 1.69–1.82.The conclusion reports that the tested f-number and numeric blur-size prompts did not meet the registered criteria.Source: §5 · p. 7, Figure 3 · p. 7, Figure 4 · p. 7

Next stage

  • The next stage will test tripod photographs with recorded distances and aperture brackets, broaden editor coverage with repeated generations, and release a public evaluation suite.It will also compare faithful implementations of published scalar metrics and test whether blinded observers detect the measured differences.Source: §5 · p. 7

A Appendix

The appendix identifies six measurement controls and describes the scope and reporting of Table 5.

Appendix

  • Table 5 lists all six measurement controls.The table reports depth and aperture indices with 95% scene-bootstrap intervals at native and Gemini output sizes.Source: A Appendix · p. 8, Table 5 · p. 9
  • Table 5 also reports the median centre-panel width in pixels at native size.The two reported output sizes are 1536 × 1024 and 1264 × 848, each with JPEG q90.Source: Table 5 · p. 9
  • The indices are undefined when a required width ratio or logarithm is undefined, including for zero widths.Source: Table 5 · p. 9

A.1 Prompts

The appendix specifies the prompts used for the centre-focus edit and its distance-cue variant.

A.1 Prompts

  • P1 asks editors to change an all-in-focus image of panels at 1.5, 2.2, and 3 m into a 50 mm f/N appearance while preserving composition and panel content.Source: §A.1 · p. 9
  • P2 adds that both off-focus panels must blur with distance from the focal plane while the centre panel remains perfectly sharp.The cue gives linear distances of 0.7 m in front and 0.8 m behind the focal plane.Source: §A.1 · p. 9
  • The prompts describe the input as f/22 without specifying sensor size, while a follow-up prompt supplies sensor size and calls the input all-in-focus.The distance cue is linear in distance, whereas the thin-lens relation is linear in 1/z.Source: §A.1 · p. 9

A.2 Registration and execution record

The execution record documents preregistration, retained outputs, run failures, and estimated costs.

A.2 Registration and execution record

  • The pilot was registered before editor calls, with the stimuli, prompts, estimator, endpoints, decision rule, and spending limits frozen.The frozen text is timestamped 2026-10-03 and identified by a SHA-256 hash.Source: §A.2 · p. 9
  • Gemini and GPT-image each contributed 300 retained images, including 288 main cells and 12 preservation controls.Gemini had 48 HTTP 402 failures and 78 duplicated cells; GPT-image’s 300 attempts all succeeded.Source: §A.2 · p. 9
  • Estimated total adapter costs were US$41.55, above the registered US$38 cap.The record notes that these are logged estimates, not invoices.Source: §A.2 · p. 9

A.3 Layout and focal-plane run

The follow-up was registered before editor calls, with two layouts, two apertures, and a fixed prompt. All 120 requested images were returned, but usable-panel counts varied substantially by editor, layout, and panel.

A.3 Layout and focal-plane run

  • All 120 requested images were returned without a failed or duplicate call; panels below 20% coverage were prospectively excluded from the indices.Including those panels in the sensitivity analysis gives the same classifications.Source: §A.3 · p. 9
  • Usable-panel counts varied across panels and layouts; for GPT-image F, centre-panel counts were 6 at f/1.4 and 10 at f/2.8.Counts are out of 24 per cell; the full panel-by-panel counts are reported in the passage.Source: §A.3 · p. 9
  • Both blurred panels were usable at both apertures in 18 Gemini R, 16 Gemini F, 20 GPT-image R, and 1 GPT-image F jobs out of 24.The GPT-image F depth index rests mainly on 16 centre measurements.Source: §A.3 · p. 9

A.4 Intervention blocks

The intervention blocks were preregistered and varied scene geometry, focus, aperture wording, and requested blur strength. The prompts included both optical terminology and qualitative descriptions of shallow or deep depth of field.

A.4 Intervention blocks

  • Each intervention block was preregistered before its first editor call, used six scenes unless stated otherwise, retained one output per cell, and had no failed or duplicate calls.The passage reports adapter-estimated costs and advance decision gates for the blocks.Source: §A.4 · p. 9
  • Interventions varied geometry and focus, removed the floor and wall, and requested Gaussian blur of about 2 or 6 pixels.The v4–v8 prompts used P1s with stated distances; v9a focused on the right panel.Source: §A.4 · p. 10
  • Prompt variants named a 50 mm lens and f/N, requested an aperture change, or asked for very shallow or deep depth of field without an f-number.The variants also included a Chinese sentence specifying the lens, aperture, and middle-panel focus.Source: §A.4 · p. 10

A.5 Geometries of the intervention blocks

The intervention geometries specify panel distances and registered thin-lens left/right width ratios; the reported oracle ratio can differ from the design value.

A.5 Geometries of the intervention blocks

  • Table 6 lists each intervention geometry’s left, centre, and right panel distances and its registered design left/right thin-lens width ratio.Where the oracle’s measured ratio differs from the design value, the text reports the measured ratio.Source: §A.5 · p. 10
  • Blocks v5, v6, v9, v10, and v11 use the original 1.5, 2.2, and 3.0 m geometry unless otherwise specified; v6 changes focus to a side panel.v9a focuses on the right panel.Source: §A.5 · p. 10
  • G1, G2, S1, S2, and v8 use equal 340 px projected panels with focus at 2.2 m; measured oracle ratios are 0.10 for G1, 5.8 for G2, and 2.59–2.62 for v8.The three v8 conditions have measured ratios of 2.59, 2.62, and 2.61.Source: §A.5 · p. 10

A.6 Exploratory exponent summary

The exploratory exponent summary, calibration checks, and sensitivity analyses qualify how the measured blur widths should be interpreted. The appendix also records the texture licensing, preservation floors, and estimator coverage details.

A.6 Exploratory exponent summary

  • Gemini’s near/far ratio equals the oracle’s ratio to a power of 0.21–0.60 in nine centre-focus conditions; the post hoc exponent summary was not registered.For the two side-focus conditions, the exponents are 0.74 and 1.07; GPT-image’s exponent is smaller and varies by configuration.Source: §A.6 · p. 10

A.7 Calibration grid

  • From 1.5–8 px, Gaussian recovery is 1.00–1.01 (native) and 1.00–1.06 (Gemini size); disc recovery is 1.11–1.25 from 1.5–6 px at both sizes.The calibration grid tests the left panel only and does not establish a general two-panel bound.Source: §A.7 · p. 10

A.8 Analytic and slanted-edge cross-checks

  • For c/4 ≥1 px, the measured/analytic ratio has median 1.17; on rendered near-panel blur, estimator-to-edge ratios are 1.57 and 1.70, versus 1.24–1.35 on synthetic discs.The reported absolute Gaussian-equivalent width scale is uncertain by about 25%; same-estimator ratios cancel a common scale factor, but width-, texture-, or kernel-dependent bias is bounded only by controls and calibration.Source: §A.8 · p. 10

A.9 Data and licensing

  • Panel textures in the stimuli and figures are RealBokeh f/22 image crops released under CC BY-NC-SA 4.0 for non-commercial research use.The centre-panel mesh and floor and wall textures were generated by the authors.Source: §A.9 · p. 11

A.10 Preservation floors

  • Unchanged-image median widths were 0.38/0.94/0.53 px for Gemini and 0.36/0.98/0.34 px for GPT-image across left/centre/right panels.The mesh centre-panel floor is near 1 px for both editors; resizing an unchanged all-in-focus render to Gemini’s output size gives a centre width near 0.37 px.Source: §A.10 · p. 11

A.11 Estimator details

  • Panel coverage is the fraction of interior window positions with measurable width; the 20% summary is descriptive, and no coverage exclusion applies to registered endpoints.The lowest-coverage panel has 138 of 24066 positions (0.57%).Source: §A.11 · p. 11

A.12 Missing-panel sensitivity

  • Worst-case missing-width assignments yield depth upper bounds of 0.529/0.540 (Gemini P1/P2) and 0.215/0.130 (GPT-image P1/P2).Corresponding aperture upper bounds are 0.255, 0.333, 0.283, and 0.291 (Gemini P1/P2, GPT-image P1/P2).Source: §A.12 · p. 11

Related papers

Loading 2610.09344v1…