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

Sebastian Rückerl

cs.CVSep 29, 2026

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

Authors: Sebastian Rückerl

Organizations: University of Hagen, Germany

Abstract

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.

Figures & tables

Figure 1 : Reading the camera an editor assumes. (a) A real photograph from SR-RAW Zhang et al. (2019) , taken with a 24 mm lens. (b) Asked to replace the floor with a checkerboard, Qwen-Image-Edit-2511 paints one in perspective. (c) The painted tile edges (red, blue) meet in vanishing points, from which we read the camera the editor assumed, with its horizon (magenta), pitch, roll and focal length (here 23 mm). (d, e) On 120 rendered cameras, this camera is pulled towards a default, with roll pulled towards level and long focal lengths towards about 30 mm.

Figure 2 : Qualitative results (Qwen unless noted). Green: true horizon, dashed magenta: horizon read from the painted board, red/blue: detected tile edges. (a, b) The same position at 24 and 85 mm. The telephoto board is painted too wide (64 mm). (c) On a rolled camera, the board painted by Kontext is read as roll $-4.8$ ∘ (Qwen: $-10.9$ ∘ , truth: $-11.4$ ∘ ). (d) On a real photograph the painted board recovers the horizon to 0.03 image heights, which is the median over NYUv2. (e) Asked to draw the horizon, the editor draws a level line (thin magenta) through the middle of the image.

	Horizon	Pitch	Roll	$\|\log f\|$	$f$ slope	$\perp$
Rendered board (oracle)	0.003	0.1	0.1	0.007	0.99	0.3
Qwen probe	0.021	0.8	0.7	0.060	0.81	2.0
original floor	0.022	0.8	0.7	0.054	0.82	2.0
Kontext probe †	0.062	2.1	2.5	0.226	0.43	6.8
LongCat probe §	0.104	3.6	3.1	0.143	0.34	5.8
GeoCalib Veicht et al. (2024)	0.054	1.6	0.2	0.094	0.85	–

Table 1 : Accuracy of the implicit camera on the 120 catalog cameras (F

\cup

Z), medians over images: horizon error (image heights), pitch and roll errors (degrees) and focal error. The

f

slope is the Theil–Sen slope of

\log\hat{f}

\log f

, and

\perp

is the deviation of the angle between the painted families in 3D from 90 ∘ . Inputs have the neutral floor unless noted, and baselines see the same inputs as the probe. † Classroom and pavilion (80 cameras), two variants on set F and one on set Z. § One variant, board with a valid registration on 58 of the 120 cameras. ‡ On the original-floor input, where

f

exists for 57% of the images. Horizon error 0.420 in the classroom and 0.008 and 0.005 in the other scenes.

	Renders (catalog)				NYUv2 photographs		SR-RAW photographs
	Pitch	Roll	Focal length	$f_{0}$	Roll	Focal length	Roll	Focal length
Rendered board (oracle)	1.01	1.00	0.99	–	–	–	–	1.00
Qwen probe, $P_{\text{cover}}$	1.00 [0.95,1.05]	0.71 [0.63,0.78]	0.81 [0.75,0.87]	33 mm	0.65 [0.60,0.69]	0.23 [0.09,0.34]	0.69 [0.59,0.77]	0.52 [0.37,0.73]
Kontext probe, $P_{\text{cover}}$ †	0.85 [0.58,1.17]	0.30 [0.09,0.45]	0.43 [0.27,0.66]	25 mm	0.18 [0.07,0.33]	$-$ 0.02 [ $-$ 0.39,0.32]	–	0.36 [0.18,0.57]
LongCat probe, $P_{\text{cover}}$ §	0.84 [0.66,0.94]	0.25 [0.13,0.36]	0.34 [0.22,0.51]	28 mm	–	–	–	–
GeoCalib Veicht et al. (2024)	0.95 [0.87,1.04]	1.00 [0.98,1.02]	0.85 [0.79,0.90]	20 mm	1.00 [0.98,1.01]	0.59 [0.53,0.68]	1.00 [0.99,1.01]	0.43 [0.37,0.49]

Table 2 : Priors of the implicit camera : Theil–Sen slopes of read-out against truth (1: follows the camera, 0: fixed) with 95% bootstrap intervals. Renders have the neutral floor. Pitch and roll are measured on set Z, focal length (log–log) on F

\cup

Z, and

f_{0}

is the fixed point of the focal fit. NYUv2: 40 images (Kontext: 20), rotated by

\pm

6 ∘ and

\pm

12 ∘ (roll) or centre-cropped to 34–102 mm (focal length). SR-RAW: roll on the widest frames of 36 sequences rotated by

\pm

6 ∘ and

\pm

12 ∘ (change against the unrotated frame, intervals over sequences), focal length on 196 optically zoomed frames (24–245 mm) with vanishing points within 10 read-out focal lengths ( Sec. 3 ). The oracle row projects an exact board into the same frames. † Classroom and pavilion only. § One variant per image, 58 of 120 renders evaluable. ∗ First variant per image, focal length on renders from set Z, roll on NYUv2 from 20 base images (80 rotations), paired on identical images.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

below line	direction	RANSAC	roll slope	$f$ slope	pitch err.
0.9	10 ∘	1.5 ∘	0.71	0.81	0.76 ∘
0.8	10 ∘	1.5 ∘	0.71	0.81	0.76 ∘
0.95	10 ∘	1.5 ∘	0.71	0.81	0.76 ∘
0.9	5 ∘	1.5 ∘	0.71	0.81	0.76 ∘
0.9	15 ∘	1.5 ∘	0.71	0.81	0.76 ∘
0.9	10 ∘	1.0 ∘	0.68	0.82	0.67 ∘

Appendix

Table S1 : Robustness of the read-out to its thresholds (neutral floor, sets F

\cup

Z, 119 evaluable images): fraction of a pair’s segments below its vanishing line, minimum direction difference of the two families, and RANSAC inlier angle. First row: default.

Figure S1 : Catalog examples (Qwen, one variant each). Titles give the true camera. Green: true horizon, dashed magenta: horizon read from the painted board, red/blue: detected floor families, yellow: pole axes. Top: the classroom at 16 mm and at 85 mm from the same position (telephoto painted too wide), and rolled cameras in the classroom and the pavilion, read as $-8.9$ ∘ for $-10.0$ ∘ and 8.9 ∘ for 10.0 ∘ (roll compressed). Bottom: painted poles in three-point perspective at 16 mm and parallel for a shift lens, a cartoon restyling and the rendered oracle board.

Figure S2 : Blur ladder examples. Input at 1024, 183, 93 and 64 px effective width (top) and the board painted by Qwen (bottom), with the read-out as in Fig. S1 . The editor paints a sharp board into a blurred scene without re-sharpening the scene.

$r$	px	evid.	roll	$f$	pitch	hor.	GC roll	GC $f$
1	1024	6.0	0.76	0.82	0.99	0.028	1.00	0.85
2	512	4.7	0.72	0.80	1.00	0.024	1.00	0.86
3.5	293	2.7	0.73	0.82	1.05	0.028	0.99	0.91
5.6	183	1.1	0.77	0.78	1.03	0.044	0.95	0.92
11	93	0.2	0.58	0.74	1.00	0.046	0.86	0.87
16	64	0.0	0.53	0.70	1.08	0.042	0.79	0.73

Appendix

Table S2 : Blur ladder (sets F

\cup

Z, neutral floor, 120 renders per level, one variant): effective width, line evidence (image diagonals), Qwen’s roll, focal and pitch slopes and horizon error, and GeoCalib’s (GC) slopes. Paired against the sharp level, Qwen’s roll slope changes by

-0.03

-0.03

and

+0.01

up to

r=5.6

(all intervals cover zero), and by

-0.16

[

-0.31

+0.03

] and

-0.23

[

-0.36

-0.10

] at

r=11

and 16. The fixed point of the focal fit is 31 mm [21, 35] pooled up to

r=5.6

, and 27 and 28 mm at

r=11

and 16. Kontext (80 renders): roll 0.30, 0.38, 0.05 and focal 0.36, 0.51, 0.32 at

r=1

, 3.5, 5.6.

Figure S3 : The default camera : rooms generated without any scene to follow (one seed each, GeoCalib read-out above each image). All models paint a level camera with a moderate wide-angle focal length. Lens instructions barely change the focal length, and a requested 15 ∘ Dutch angle barely changes the roll.

Model	Condition	$f$ (mm)	Pitch	Roll
Qwen editor	blank grey image	32	$-5.2$ ∘	0.1 ∘
Kontext editor	blank grey image	26	$-17.2$ ∘	$-0.5$ ∘
LongCat editor	blank grey image	23	$-9.6$ ∘	$-0.1$ ∘
OmniGen2 editor	blank grey image	29	$-9.7$ ∘	0.2 ∘
Qwen text-to-im.	neutral	31.5	$-7.0$ ∘	$-0.1$ ∘
	“24mm wide-angle”	25.6	$-8.9$ ∘	0.3 ∘

Appendix

Table S3 : Default camera without scene cues : medians over 10 seeds (GeoCalib on the generated images).

Figure S5 : An SR-RAW sequence painted at six focal lengths (photographs from SR-RAW Zhang et al. (2019) , Qwen with $P_{\text{replace}}$ , variant closest to the median). Titles give the true and the read-out focal length. The painted board widens less than the lens: the 240 mm frame is painted like a 93 mm one.

Nominal focal length (mm)	24	35	50	70	100	150	240
True focal length	24	36	53	74	106	159	243
Exact synthetic board	24 (36/36)	36 (35/36)	53 (36/36)	74 (31/32)	105 (22/24)	159 (19/20)	227 (8/12)
Qwen, $P_{\text{replace}}$	23 (23/36)	31 (25/36)	36 (21/36)	49 (21/32)	56 (12/24)	74 (12/20)	106 (6/12)
Qwen, $P_{\text{cover}}$	22 (22/36)	29 (21/36)	45 (24/36)	44 (17/32)	50 (11/24)	132 (8/20)	71 (6/12)
Kontext, $P_{\text{cover}}$	22 (15/36)	21 (17/36)	32 (23/36)	33 (9/32)	33 (13/24)	56 (9/20)	34 (7/12)
GeoCalib Veicht et al. (2024)	24 (36/36)	33 (36/36)	42 (36/36)	47 (32/32)	51 (24/24)	51 (20/20)	52 (12/12)

Appendix

Table S4 : Read-out focal length per focal length on the optical SR-RAW frames (median in mm, 35 mm equivalent). Parentheses give the frames that yield a focal length and all frames. First row: true focal length (median). Painted and exact boards with vanishing points within 10 read-out focal lengths.

Data	Probe	$\leq$ 10 image heights	$\leq$ 10 read-out focal lengths	no bound
Catalog renders	Qwen, $P_{\text{cover}}$	0.81 [0.75,0.87] (105)	0.80 [0.74,0.87] (108)	0.79 [0.72,0.85] (119)
	Kontext, $P_{\text{cover}}$ †	0.43 [0.27,0.66] (44)	0.38 [0.20,0.62] (45)	0.25 [0.04,0.52] (53)
NYUv2 centre crops	Qwen, $P_{\text{cover}}$	0.23 [0.09,0.34] (104)	0.23 [0.12,0.37] (111)	0.30 [0.17,0.43] (151)
	Qwen, $P_{\text{replace}}$ ∗	0.55 [0.44,0.66] (97)	0.55 [0.44,0.65] (107)	0.54 [0.43,0.65] (132)
SR-RAW optical zoom	Qwen, $P_{\text{replace}}$	0.56 [0.45,0.65] (95)	0.62 [0.52,0.73] (120)	0.69 [0.56,0.82] (167)
	Qwen, $P_{\text{cover}}$	0.41 [0.28,0.54] (92)	0.52 [0.37,0.73] (109)	0.49 [0.32,0.68] (167)

Appendix

Table S5 : Bound on the vanishing points : focal slope (log–log Theil–Sen with 95% interval, over images with a focal length) when both vanishing points must lie within 10 image heights of the centre (default for catalog and NYUv2), within 10 read-out focal lengths (default for SR-RAW), or anywhere. † Classroom and pavilion only. ∗ First variant per image.

Command (card)	$n$	produced	believed	$\Delta$	$f$ GC	$f$ board
eye-level (0 ∘ )	25	$-9.9$	$-12.0$	1.4	0.99	1.02
elevated (30 ∘ )	26	$-26.6$	$-26.7$	2.4	0.96	0.97
high-angle (60 ∘ )	25	$-36.9$	$-37.1$	3.4	0.89	0.98
low-angle ( $-30$ ∘ )	15	$+2.7$	$+3.5$	2.0	0.98	0.99
close-up ( $\times$ 0.6)	14	$-12.7$	$-14.7$	1.5	1.01	1.02
wide ( $\times$ 1.8)	16	$-1.7$	$-0.2$	1.5	0.86	0.82

Appendix

Table S6 : Multi-angle LoRA : median output pitch ( ∘ ) produced (GeoCalib) and believed (painted board), their median absolute difference

\Delta

, and focal length output/input by GeoCalib (GC) and by the board. Inputs have a median pitch of

-15

∘ (

-11

∘ to

-22

∘ ). For eye-level, elevated, high-angle and front-right, 12 inputs at 0 ∘ and

-30

∘ are included. Variants with a failed registration are discarded. Reference on the original renders:

\Delta=2.0

∘ .

Explore similar work

Sep 1, 2026cs.CV

CameraEditor: Camera-Controlled Image Editing via Video-Prior Sequential Modeling

Beyond semantic content, camera parameters play a pivotal role in dictating the geometric perspective and appearance of any given image. While recent image editing models excel at semantic and stylistic manipulation, they struggle with explicit camera parameter control. When handling large perspective shifts, instruction-driven models face a dilemma: they either suffer from structural tearing or generate conservative outputs that ignore geometric instructions. To address this, we introduce CameraEditor, a framework that reformulates camera-controlled editing from a spatial problem into a temporal sequence prediction task. By leveraging the temporal coherence of video diffusion models, our approach integrates an explicit geometric perception module with a dynamic reference routing mechanism. This allows us to construct geometrically rigorous visual reference pairs via dynamic panorama cropping, overcoming the ambiguity of text-based instructions. Furthermore, CameraEditor strategically inserts intermediate transition frames to decompose large perspective shifts, providing a robust temporal buffer that preserves content identity and spatial coherence. We construct a training dataset of 5,760 instances. As an independent contribution, we introduce CamEditor-Bench, a model-agnostic evaluation suite of 462 test cases. Extensive experiments demonstrate that CameraEditor achieves state-of-the-art camera control precision and source identity preservation, outperforming existing methods.

Xin Shen, Chengyou Jia, Keshuo Xing +6

Xi’an Jiaotong University, China · Agency for Science, Technology and Research (A*STAR), Singapore

Camera Control Video Diffusion Models Spatial Coherence

Ask PDF

May 6, 2026cs.CV

Open-Source Image Editing Models Are Zero-Shot Vision Learners

Recent studies have shown that large generative models can solve vision tasks they were not explicitly trained for. However, existing evidence relies on closed-source models~(Veo~3, Nano Banana Pro) or requires task-specific instruction tuning, leaving open whether publicly available image-editing models possess zero-shot vision abilities out of the box. We conduct a systematic evaluation of three open-source image-editing models -- Qwen-Image-Edit, FireRed-Image-Edit, and LongCat-Image-Edit -- on dense visual prediction tasks \emph{without any fine-tuning}. We benchmark monocular depth estimation on NYUv2 and DIODE, surface normal estimation on NYUv2, and semantic segmentation on Cityscapes, covering both geometric and semantic scene understanding. Results show that open-source image-editing models exhibit non-trivial zero-shot visual understanding. On NYUv2 surface normals, FireRed-Image-Edit achieves a mean angular error of

17.69^\circ

, surpassing the fine-tuned Marigold (

20.86^\circ

) and matching the instruction-tuned Vision Banana (

17.78^\circ

) without any task-specific training. On NYUv2 depth estimation, LongCat-Image-Edit obtains

δ_1{=}0.822

with affine alignment, and Qwen-Image-Edit leads on DIODE Indoor (

δ_1{=}0.868

). On Cityscapes semantic segmentation, Qwen-Image-Edit reaches 25.7 mIoU at the 19-class level and 49.5 mIoU at a coarser 7-category level. By comparing three independently trained editors, we test whether zero-shot vision ability is an emergent property of image-editing pretraining rather than a model-specific artifact. Code, evaluation scripts, and all results are publicly released to serve as a reproducible baseline for future work.

Wei Liu, Jiaxin Lin, Rui Chen

Tencent Inc.

Instruction Tuning

Ask PDF

Sep 27, 2026cs.CV

InfoEdit: Probing Global Layout Reasoning in Infographic Editing

Multimodal foundation models edit natural photographs at production quality, yet the same models struggle with structured visual content such as infographics. Unlike photographs, infographics encode information through logical relations; editing one element often requires surrounding elements to be adapted. We refer to this global layout reasoning capability as reflow. Existing image-editing benchmarks neither provide a dedicated setting for structured visual content nor evaluate the reflow capability. We introduce InfoEdit, a novel benchmark of 1,000 infographics across eight logical-relation families, paired with 4,000 editing instructions across four editing tasks, and a reflow-aware evaluation protocol. Across eight frontier editors, only GPT-Image-2 clears 60% average success rate; most models fall below 7%, and no editor exceeds 36% on the Swap-Block task even with perfect target localization. We further show that code-level editing can match the strongest pixel-level editor, revealing complementary strengths across tasks. InfoEdit identifies reflow as a central challenge in structured visual content editing and provides a diagnostic benchmark to facilitate future progress.

Cheng Yang, Chufan Shi, Huijuan Wang +6

University of California San Diego · University of Southern California · University of Illinois Urbana-Champaign +1

Multi-Image Editing Benchmarks Layouts

Ask PDF