PaperlayerView exact revision on arXiv

Source-linked AI summary

The Camera Inside the Editor: Reading the Implicit Camera of Image Editors with Painted Calibration Patterns

Sebastian Rückerl

arXiv:2609.37732v1cs.CVcs.LG

TL;DR

Image editors imply a camera when painting into photographs, but which camera they assume is unknown. The paper reads a painted checkerboard with classical geometry to measure that camera. On rendered views, Qwen’s camera estimates are more accurate than GeoCalib in horizon, pitch, and focal length, but not roll.

  • Problem

    Which camera an image editor assumes when painting into a photograph, and how it errs, has not been measured by evaluations that judge edits from the outside.

    Source: §1 · p. 2

  • Method

    The editor paints a floor checkerboard, whose projective structure is read with classical geometry to recover its implicit camera without training.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    0.8° pitch error and 6% focal-length error: Qwen is more accurate than GeoCalib in horizon, pitch, and focal length on renders, but less accurate in roll.

    Source: §5 · p. 5

  • Takeaways & Limitations

    Painted probes make implicit camera geometry measurable, including knowledge editors paint but cannot name.

    Source: §9 · p. 10, §8 · p. 10

    The study covers few editors and indoor scenes; the read-out assumes square pixels, a centered principal point unless poles are painted, and a visible floor.

    Source: §8 · p. 10

Abstract

from arXiv · show

Instruction-based image editors insert objects, restyle scenes and render new viewpoints, but it is unknown which camera they assume when they paint into a photograph. Asked to cover the floor with a checkerboard, an editor paints projective structure from which classical vanishing-point geometry reads pitch, roll, focal length, yaw and, on renders, the principal point, without any training. Unlike a calibrator such as GeoCalib, which estimates the camera of an image, this isolates the camera under which the editor paints. On 120 rendered cameras with exact ground truth, Qwen-Image-Edit-2511 paints tile edges that meet their vanishing points within 0.26 degrees, and its implicit camera matches the true one to 0.8 degrees in pitch and 6% in focal length, more accurately than GeoCalib except in roll. Asked to draw the horizon or mark a vanishing point instead, the editor fails, so this knowledge is revealed by painting and not by the explicit tasks we tried. The implicit camera has two priors: roll is pulled towards level (slope 0.71), and telephoto perspective towards a default of about 30 mm, which roughly matches the camera the models paint without any scene. For Qwen, the priors do not grow when blur removes four fifths of the line evidence. They are stronger on real photographs, and on NYUv2 a shorter wording of the task removes the difference for roll. On photographs from a 24--240 mm zoom lens the painted perspective grows with only 0.62 of the lens's slope, while GeoCalib and MoGe-2 saturate at about 52 and 42 mm. FLUX.1 Kontext and LongCat-Image-Edit are pulled much harder. Finally, from a level camera a camera-control LoRA executes pose commands at only 50--70% of their strength, and a board painted into its output agrees with the camera it produced.

1 Introduction

Instruction-based editors imply a camera when painting into photographs, but which camera they assume is unknown. The paper reads a painted checkerboard with classical geometry to measure that implicit camera; Qwen’s estimates are precise but systematically biased.

1 Introduction

  • Which camera an editor assumes when painting into a photograph, and how that camera errs, has not been measured by prior evaluations of edits.Prior evaluations judge plausibility of object sizes or reconstruct viewpoints from outputs.Source: §1 · p. 2
  • A painted checkerboard reveals the camera used for the editor’s addition, unlike a calibrator that estimates the camera of an image as a whole.The read-out uses classical geometry and requires no training or access to model weights or features.Source: §1 · p. 2
  • Qwen’s implicit camera is more accurate than GeoCalib in horizon, pitch, and focal length on renders, but not in roll.Its painted lines meet their vanishing points within 0.26°.Source: §1 · p. 2
  • The paper contributes painted calibration probes, an exact-ground-truth camera catalog with controls, and characterizations of editor cameras and pose-command execution.The study also examines two camera priors and whether they grow when line evidence is blurred.Source: §1 · p. 2

2 Related Work

Prior work studies image calibration, generative models’ geometric knowledge, painted probes for photometry, and camera control. This paper extends these threads by measuring camera geometry that editors assume while painting.

2 Related Work

  • Single-image calibration uses line geometry and learned estimators, while this approach has the editor supply a target that classical geometry can read without training or known intrinsics.Examples include GeoCalib (Veicht et al., 2024) and Perspective Fields (Jin et al., 2023).Source: §2 · p. 2
  • Work probes generative models for scene properties, but not camera parameters; this paper finds projective precision when an editor paints a known structure.Prior work also reports generated images that violate projective geometry (Okumura et al., 2026).Source: §2 · p. 3
  • Prior probes use painted objects to assess illumination or colour, whereas this work extends the principle to camera geometry.Examples include chrome balls for illumination (Phongthawee et al., 2024) and colour checkers for white balance (Chang et al., 2025).Source: §2 · p. 3
  • Camera-control benchmarks measure the viewpoint an edit produces, whereas this work reads the camera an editor assumes while painting, including in generated views.Examples include SpatialEdit (Xiao et al., 2026) and GenScale (Li et al., 2026).Source: §2 · p. 3

3 Reading the Implicit Camera

The method asks an editor to paint a floor checkerboard, detects its line families, and applies classical projective geometry to read the implicit camera. An oracle checkerboard validates the read-out against exact camera ground truth.

3 Reading the Implicit Camera

  • The standard probe asks the editor to cover the visible floor with square black-and-white tiles, while a shorter floor-replacement prompt and vertical poles provide alternatives.The editor is treated as a black box; its weights and features are never accessed.Source: §3 · p. 3
  • The pipeline registers the edited image to the input, locates the changed high-contrast region, and detects long line segments there.Registration uses ORB features and a RANSAC similarity transform; line detection uses LSD at full resolution.Source: §3 · p. 3
  • RANSAC extracts up to three vanishing-point segment families; the floor pair is selected using segment-count, below-horizon, and direction-separation criteria.The thresholds do not affect the reported conclusions.Source: §3 · p. 3
  • Two floor vanishing points define the horizon and roll; with a centered principal point and square pixels, their orthogonality yields focal length and pitch.Painted poles add a vertical vanishing point for principal-point and focal-length read-out, while back-projecting a floor vanishing point gives yaw relative to the room.Source: §3 · pp. 3–4
  • On 120 rendered cameras, an oracle checkerboard recovers horizon to 0.003 image heights and focal length to 0.7%, validating the geometric read-out.Under the true camera, the painted families can also be back-projected to measure board tilt, inter-family angle, and alignment with room axes.Source: §3 · p. 4

4 Benchmark

The benchmark combines rendered cameras with exact ground truth, controlled image manipulations, and photographs with known camera changes. It compares several editors and baselines using camera-error and prior measurements.

4 Benchmark

  • The rendered catalog contains 120 cameras across three Blender scenes, including focal sweeps, random pitch-and-roll views, and separate level views with lens shift.Focal lengths are 35 mm equivalents.Source: §4 · p. 4
  • Rendered controls vary floor appearance, blur, digital zoom, rotation, photo-like effects, and cartoon restyling; photographs use known synthetic camera changes and optical-zoom sequences.The photo sets include NYUv2 indoor scenes and SR-RAW zoom sequences, mostly outdoors.Source: §4 · p. 4
  • The benchmark compares Qwen, Kontext, and LongCat with GeoCalib, MoGe-2, and classical vanishing points, reporting camera errors and slopes that quantify priors.A slope of 1 means estimates follow the true camera; 0 means they are fixed.Source: §4 · pp. 4–5
  • Figure 2 contrasts true and board-read horizons and tile edges; its 85 mm example shows the telephoto board painted too wide, at 64 mm.The figure also shows the editor drawing a level mid-image line when asked to draw the horizon.Source: Figure 2 · p. 5
  • Figure 3 compares catalog horizon errors and painting versus naming: the board-read horizon follows truth, while the explicitly drawn line stays at mid-height.Panel (a) uses the neutral-floor catalog; panel (b) uses set Z in the classroom and pavilion.Source: Figure 3 · p. 5

5 What the Editor Knows

A painted checkerboard yields accurate camera estimates on rendered scenes, while explicit requests to name camera geometry fail. Its reliability varies with the scene and image domain, and the board can recover geometry without floor texture.

Accuracy

  • On NYUv2, horizon error is 0.034 image heights versus GeoCalib’s 0.040, but focal error is 8% versus 4% and rotated-image roll error is 3.0° versus 0.6°.Classical vanishing points on the same photographs are off by 0.209 image heights.Source: §5 · pp. 5–6
  • On 120 renders, Qwen’s board recovers horizon to 0.021 image heights, pitch to 0.8°, roll to 0.7°, and focal length to 6%; it beats GeoCalib except in roll.The comparison is on identical inputs; the table covers the 120 catalog cameras.Source: §5 · p. 5, Table 1 · p. 6

Projective fidelity

  • Painted tile segments meet their vanishing points with a 0.26° median angular residual, and the board estimates room-relative yaw to a median error of 1.1°.On renders, painted poles and the board give the principal point to 3.8% of image width; the pole probe only works on renders.Source: §5 · p. 6

Not copied from the floor

  • Removing floor structure does not measurably reduce accuracy in the classroom and apartment; the probe reads perspective from remaining scene structure.The floor texture is not required for the painted perspective read-out.Source: §5 · p. 6

Painting versus naming

  • Asked to draw the horizon or mark a vanishing point, Qwen fails: it places the horizon near image center and its vanishing-point mark farther away than image center.The editor’s geometric knowledge is revealed by painting rather than the explicit tasks tested.Source: §5 · pp. 6–7

6 The Priors of the Implicit Camera

The implicit camera follows pitch but compresses roll and telephoto perspective toward defaults. The priors persist under substantial blur, strengthen on real photographs, vary with task wording, and differ in strength across editors.

6.1 Roll is pulled towards level

  • Read-out roll grows only 0.71 times as fast as true roll: a 12° roll is painted as 8.5°, while pitch follows 1:1 and GeoCalib follows roll exactly.The roll compression is symmetric and holds across scenes and floors.Source: §6.1 · p. 7

6.2 Telephoto is pulled towards a default

  • Beyond about 40 mm, painted perspective stays too wide: an 85 mm camera reads as about 58 mm, with a log–log fit slope of 0.81 and fixed point f0 = 33 mm.The pooled fixed point is 31 mm [21, 35]; without a scene, Qwen paints f = 32 mm, roughly matching the default.Source: §6.2 · pp. 7–8

6.4 Real photographs and the wording of the task

  • On 24–240 mm SR-RAW zoom frames, painted focal length grows with slope 0.62 and paints 240 mm as 106 mm; GeoCalib and MoGe-2 read those frames as 52 mm and 42 mm.An exact board in the same frames is read with slope 1.00; digital and optical zoom to the same field of view are read alike.Source: Figure 6 · p. 8, §6.4 · p. 9

6.3 The priors do not grow with line blur

  • For Qwen, roll and focal slopes stay flat through 183 px effective width, despite four fifths of usable floor-line evidence being gone; only at 64 px does roll fall further, to 0.53.At 183 px, paired slope changes are within ±0.04 and not significant; the line evidence the editor itself uses is not measured.Source: §6.3 · p. 8

6.4 Real photographs and the wording of the task

  • On real photographs, roll slope is 0.65 and painted focal length saturates near 45 mm; on NYUv2, shorter wording raises roll slope from 0.63 to 0.78 and doubles focal slope.For NYUv2 roll, the paired change is +0.15 [+0.07, +0.20]; the shorter prompt reaches the render value for roll.Source: §6.4 · pp. 8–9

6.5 Editors differ

  • Kontext and LongCat have stronger priors: their render roll slopes are 0.30 and 0.25, focal slopes 0.43 and 0.34, and focal fixed points 25 and 28 mm.Both are less accurate than Qwen on every measure; only half of LongCat outputs yield a readable board with valid registration.Source: §6.5 · p. 9

7 Writing versus Reading the Camera

The paper uses a multi-angle camera-control LoRA to write new viewpoints, then reads the output camera with GeoCalib and a painted checkerboard. Commands are weaker than their targets, but the board agrees with the camera produced.

Camera-control commands

  • From a level input, “elevated shot” (30° target) tilts the camera to −20°, while “high-angle shot” (60° target) reaches −28°.The commands are executed at reduced strength.Source: §7 · p. 9
  • A 45° “front-right quarter view” turns the camera by only 30°, while the eye-level control does not turn it.Yaw is measured using the board read-out.Source: §7 · p. 10

Reading generated views

  • For every command, board-read pitch differs from GeoCalib’s output-camera estimate by 1.4–3.4° median per command, comparable to 2.0° on unedited renders.The board therefore reads generated views consistently even when the produced camera is not the commanded one.Source: §7 · p. 10

8 Discussion and Limitations

The measured priors point to a level camera at about 30 mm, a camera five models also paint without a scene. The authors cannot test the proposed training-data explanation without access to the training data, and the camera read-out has several stated scope limits.

Discussion

  • Five different models paint a level camera at about 30 mm even when given no scene.The authors suggest the editor infers camera from coarse scene structure and biases it toward this default; testing the training-image explanation would require training-data access.Source: §8 · p. 10
  • Long-focal-length underestimation also appears in the tested image-trained models GeoCalib and MoGe-2, not only in editors.Both calibrators saturate on optically zoomed photographs, while the board read-out recovers long focal lengths in the same frames.Source: §8 · p. 10

Implications

  • Editors used for compositing or data synthesis are likely to inherit a level camera of about 30 mm, making inserted adults too small in 85 mm views.For adults of 1.72 m, the implied camera is 2.2–3.8× too high, more than the board’s telephoto prior explains.Source: §8 · p. 10

Limitations

  • The study covers three open editors and three indoor scenes; OmniGen2 was measured only for the default camera.The authors note that the protocol applies to any editor accepting an image and prompt, but their sample is small for a fast-moving field.Source: §8 · p. 10
  • The camera read-out assumes square pixels, a centred principal point unless poles are painted, and a visible floor.Poles work on renders but not real photographs; only 61% of zoom frames yield focal lengths, and the estimated slope depends on the vanishing-point bound.Source: §8 · p. 10

9 Conclusion

A checkerboard makes an editor’s assumed camera measurable, and the conclusion argues for painted probes as a way to read geometric knowledge that editors cannot name. The prompts vary the board task and include explicit geometry and camera-control requests.

9 Conclusion

  • A painted checkerboard makes the camera an editor assumes measurable, and the conclusion proposes painted probes as a routine way to measure generative models’ camera assumptions.The paper reports that editors paint geometry they cannot name; the measured implicit camera is precise on renders but pulled toward a level default of roughly 30 mm, more strongly on real photographs and depending on wording.Source: §9 · p. 10

A Prompts

  • The main checkerboard prompts ask editors to cover or replace the floor, with a paraphrased chessboard task and a separate request to add vertical poles.The pole prompt specifies six straight magenta poles extending from the floor toward the ceiling or image top.Source: A Prompts · p. 15
  • Explicit prompts ask the editor to draw the horizon or mark the vanishing point, while other variants specify lens, tile size, perspective correction, or pattern replacement.The lens prompt names a full-frame-equivalent focal length; tile-size variants specify 50 cm or 1 m squares.Source: A Prompts · p. 15
  • Default-camera prompts request a checkerboard room from a blank image or text, optionally specifying 24 mm, 85 mm, or a 15-degree Dutch angle.Source: A Prompts · p. 15
  • The multi-angle LoRA is tested with commands spanning eight azimuths, four elevations, and three distances, with a checkerboard added to each output.Source: A Prompts · p. 15

B Read-out Details and Robustness

The read-out is checked against threshold changes and line evidence, while the catalog and photograph controls vary camera, image quality, and viewing conditions. The controls include blur, rotations, digital zoom, and optical-zoom sequences.

B Read-out Details and Robustness

  • Registration failures with fewer than 20 inliers and large deviations are discarded, affecting 2% of Qwen, 11% of Qwen original-floor, 20% of Kontext, and 44% of LongCat variants.On real photographs, the probe is read in the edited image itself because editor drift is negligible.Source: B Read-out Details and Robustness · pp. 15–16
  • With seven floor-pair threshold settings, slopes and pitch error differ from the default by at most 0.03 and 0.09°.Table S1 evaluates 119 images from the neutral-floor catalog; its first row is the default setting.Source: B Read-out Details and Robustness · p. 16, Table S1 · p. 16
  • Among sharp images above 45 mm, more floor-line evidence accompanies less focal compression; roll slopes also track evidence across scenes and blur levels.The focal-compression association is Spearman ρ = 0.43, p = 0.03, n = 26; roll slopes correlate with evidence at ρ = 0.62 for Qwen and 0.91 for GeoCalib over 18 scene–level cells.Source: B Read-out Details and Robustness · p. 16

C Catalog and Controls

  • The catalog renders three indoor scenes with cameras spanning 16–85 mm, freely sampled focal lengths, and level shifted-lens views with known camera ground truth.The camera sets include two positions per scene for the fixed-lens set; the shifted views move the principal point from image height 0.5 to 0.34 or 0.23.Source: C Catalog and Controls · p. 16
  • Catalog controls include a neutral floor, oracle checkerboard, blur ladder, digital zoom, image rotation, photo-look degradation, and cartoon restyling.The blur ladder downsamples and upsamples to effective widths of 512 to 64 px; the photo look adds downsampling, color changes, noise, and JPEG compression.Source: C Catalog and Controls · p. 16, Figure S2 · p. 17, Figure S1 · p. 17
  • Real-photo tests use 40 NYUv2 images with digital zoom and 36 SR-RAW optical-zoom sequences, with rotated frames evaluated relative to each sequence’s unrotated frame.SR-RAW sequences are rotated by −12°, −6°, 0°, +6°, and +12°; their true outdoor roll is unknown.Source: C Catalog and Controls · p. 17

D Additional Results

Additional tests characterize when painted-camera readouts remain reliable, where they diverge from calibrators, and how camera priors affect other edits. The results include sensitivity to wording and blur, as well as limits of tile-size and pole-based measurements.

Blur ladder

  • Qwen’s roll compression persists under blur and varies little through r = 5.6; its roll errors also remain reproducible between sharp and 183 px images.The roll slope changes by −0.03, −0.03, and +0.01 through r = 5.6, with intervals covering zero; sharp-to-183 px error correlations are ρ = 0.60 for roll and 0.67 for focal length.Source: D Additional Results · p. 17, Table S2 · p. 18

Accuracy against GeoCalib

  • On rendered scenes, the probe has lower horizon error in all three scenes and lower pitch error in two, while focal and roll comparisons vary by floor and scene.On the neutral floor, focal error is lower by 0.021, while roll error is higher by 0.27°; on the original floor, horizon and pitch errors are lower, focal error is on par, and roll error is higher.Source: D Additional Results · pp. 18–19

Roll prior per scene

  • Roll slopes on set Z range from 0.60 to 0.84 across scenes, and rotating and cropping renders like photographs yields a slope of 0.85.Adding a NYU-like photo look lowers the slope to 0.78; on NYUv2, digital zoom of renders is read like optical zoom, while a shorter prompt raises the focal slope and lowers horizon error.Source: D Additional Results · pp. 18–19

Painted tiles

  • Painted tiles are square and correctly foreshortened, but their unprompted sizes vary by room and are unreliable as a metric ruler for camera height.In the classroom, “50 cm by 50 cm” and “1 m by 1 m” requests yield tiles of 0.29 m and 0.37 m; the pavilion’s requested 1 m tiles measure 1.01 m.Source: D Additional Results · p. 19

Poles

  • On renders, poles and the board recover the principal point to 3.8% of image width and 5.2% of height, and focal length to 7%; on digitally zoomed NYUv2 photos, poles are nearly parallel.Adding poles improves telephoto focal-length error on renders (0.105 versus 0.199 from the board alone) but not on photographs.Source: D Additional Results · p. 19

Accuracy on NYUv2

  • On 40 unmodified NYUv2 photographs, horizon error is 0.034 image heights for the probe versus 0.040 for GeoCalib, while focal-length error is 8% versus 4%.Both line-based methods share a +0.034/+0.030 image-height offset against the depth-derived horizon; it cancels in comparisons using each photograph’s own ground truth.Source: D Additional Results · pp. 19–20

Combining the probe with GeoCalib

  • Using GeoCalib as fallback for images without a probe pitch lowers median pitch error across 120 cameras to 0.77° and 0.94°, versus 1.60° and 1.47° for GeoCalib alone.Dividing the probe’s roll by its catalog slope reduces rotated-photo roll error from 3.0° to 2.3°, compared with 0.6° for GeoCalib.Source: D Additional Results · p. 20

Inserted people

  • At 85 mm, inserted people imply heights 2.2–3.8 times true height; their size grows with focal length more strongly than the board’s telephoto prior predicts.The implied-height log–log slope is +0.55 (r = 0.88); using the camera read from the board gives +0.42, while an 85 mm lens is read as about 58 mm.Source: D Additional Results · p. 20

E SR-RAW: Optical Zoom and Rotation

The SR-RAW experiments test painted-camera readout across real optical zoom and rotation, then compare additional editors and camera-control outputs. The measurements show limits in readable focal lengths and command strength, alongside close agreement between video-board readouts and produced keyframes.

E SR-RAW: Optical Zoom and Rotation

  • On optical SR-RAW frames, Qwen’s painted perspective is slightly too wide at 24 mm and falls increasingly short beyond 100 mm; GeoCalib is exact at 24 mm but levels off near 50 mm, and MoGe-2 near 42 mm.Qwen returns a focal length for 61% of frames with Preplace and 56% with Pcover, whereas GeoCalib and MoGe-2 always return one.Source: E SR-RAW: Optical Zoom and Rotation · p. 21, Table S4 · p. 21
  • The 10-image-height vanishing-point bound discards correct exact-board readings at long focal lengths, so SR-RAW uses a bound of 10 read-out focal lengths instead.With the scale-free bound, exact boards are read in 8 of 12 frames at 240 mm and 19 of 20 at 150 mm, with slope 1.00 under every bound.Source: E SR-RAW: Optical Zoom and Rotation · p. 21, Table S5 · p. 21
  • For rotated SR-RAW frames, Qwen reads roll changes with slopes of 0.69 for Pcover and 0.78 for Preplace, below GeoCalib’s 1.00.A −12°/+12° change is read as −8.9°/+8.9° with Pcover and −9.5°/+10.6° with Preplace.Source: E SR-RAW: Optical Zoom and Rotation · p. 22

F Further Editors

  • LongCat-Image-Edit shows strong camera priors on the catalog, with roll slope 0.25, focal slope 0.34, and a fixed point of 28 mm; it is the least accurate of the three editors in horizon and pitch.It paints a readable board into 89 of 120 renders, 58 of them registrable.Source: F Further Editors · p. 22

G Writing versus Reading: Details

  • Camera-control LoRA pitch outputs depend on the command: fitted slopes are 0.72 for eye-level, 0.44 for elevated, and 0.59 for high-angle, between absolute-target and input-relative control.“Wide shot” opens the field of view and levels the camera, while “close-up” hardly changes a room.Source: G Writing versus Reading: Details · p. 23
  • Across seven LTX-2.5 videos, board-read trajectories end within 0.7° of true keyframes, with focal length at 0.98–1.01 of truth.Trajectories are monotonic apart from single backward steps of at most 0.10° in two videos; one nearly level ending is measurable in only 22% of frames.Source: Figure S6 · p. 22, G Writing versus Reading: Details · p. 23

H Reproducibility

The paper archives its measurements and generated materials and reports the editing setup and bootstrap procedure.

H Reproducibility

  • Measurements, ground truth, renders, and JPEG output copies are archived, and all numbers, tables, and figures are generated from per-image measurements by scripts.Editing ran on one consumer GPU; confidence intervals use 1000 bootstrap resamples over images, with 2000–4000 for some paired tests.Source: H Reproducibility · p. 23

Related papers

Loading 2609.37732v1…