acceptodds
Under review as a conference paper at ICLR 2027

Lens as Action: World-State Querying for Scene-Consistent Photographic Generation

Abstract

Camera-controllable text-to-image generation aims to synthesize photographs that comply with both semantic prompts and specified camera settings while maintaining scene coherence. Existing methods typically inject lens parameters directly into the denoising process as auxiliary conditions, but their responses to continuously varying parameter values can be weak or non-monotonic. Because scene content and camera-dependent appearance are represented jointly, changing a lens setting may also inadvertently alter scene geometry or semantic layout. In this work, we revisit photographic image generation from the perspective of lens-action world modeling. Rather than treating lens parameters merely as generative conditions, we interpret them as observation actions over a latent scene world. Based on this formulation, we propose a Lens-Action World Model (LAWM) for scene-consistent photographic generation. Given a text prompt, LAWM establishes an anchor observation and encodes its appearance together with geometric cues into a persistent world state shared across lens queries. A structured Lens-action encoder and World-Action Cross-Attention retrieve action-relevant features from this state and inject them into a temporally coupled diffusion backbone, producing photographs under the requested settings. To alleviate the scarcity of aligned multi-setting photographs, we construct training supervision for lens parameters such as focal length and aperture using field-of-view transformations and depth-guided defocus simulation. Extensive experiments show that our method achieves stronger scene consistency and more reliable responses to continuous lens controls, demonstrating the potential of explicit world-state querying for camera-controllable image generation.

Primary area:
generative models
Keywords:
Photographic Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_SBvyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LAWM for scene-consistent, camera-controllable text-to-image generation. It encodes an anchor image and VGGT-derived geometry into a shared scene representation, then uses camera-conditioned cross-attention to guide a temporally coupled diffusion model. Training targets are synthesized through field-of-view transformations and depth-guided defocus. Results show improved scene consistency and f-number response, while focal-length control remains limited.

Critical strengths
  1. Meaningful improvements in scene consistency and f-number response. The LPIPS results and qualitative examples show better preservation of scene content. F-number changes also produce more consistent sharpness trends. These gains are useful for practical camera-controlled generation.

  2. A coherent design supported by component ablations. The anchor-based scene representation provides an explicit constraint for maintaining scene consistency. Table 3 shows that query-based fusion performs better than additive fusion, providing evidence for the usefulness of the proposed conditioning mechanism.

Critical weaknesses
  1. Limited novelty and insufficient isolation of the contribution. The work appears to be an incremental extension of Generative Photography (CVPR 2025), mainly adding an anchor-based scene representation, VGGT geometry, and querying. Table 3 supports querying over additive fusion, but does not isolate the contributions of anchor appearance and geometry or establish an advantage over standard reference-conditioned generation. Direct field-of-view and defocus transformation baselines are also missing. Consequently, the added value of the proposed world-state formulation remains unclear.

  2. Physical camera control is not sufficiently validated. The f-number results mainly measure global sharpness trends, which do not establish physically correct depth-dependent defocus. The supervision pipeline also lacks sufficient explanation of focus distance, depth-to-blur mapping, and existing source defocus. Meanwhile, focal-length correlation is only 0.221, compared with 0.219 for Generative Photography. The reported consistency and sharpness-ordering gains therefore do not sufficiently establish reliable physical camera control, which is central to the paper.

Questions for authors
  1. Contribution validation: Table 3 compares querying with additive fusion. Could the authors further separate the contributions of anchor appearance and VGGT geometry, and establish what querying adds beyond standard reference-conditioned generation?

  2. Baseline evidence: What benefit does LAWM provide over directly applying the field-of-view and depth-guided defocus transformations to the same anchors? Could the authors provide this comparison to substantiate the need for the generative framework?

  3. Clarification of physical supervision: How are target f-numbers converted into depth-dependent blur, accounting for focus distance and existing source defocus? What quantitative evidence supports physically meaningful defocus beyond global sharpness ordering?

5 comments
AuthorsOpenReview ↗

Question 1: Appearance, Geometry, and Scene Querying

Thank you for recognizing the consistency and f-number improvements and for identifying this attribution issue. The submitted ablations did not fully separate appearance, geometry, and querying.

Relationship to Generative Photography. LAWM inherits the generative backbone and temporal prior; we do not claim a new diffusion architecture or attention primitive. Its added design encodes anchor appearance and geometry into a scene state shared across lens queries. Lens-Action Scene Querying, described in Section 3.4, uses action-conditioned lens features to query pooled scene tokens together with action tokens, followed by gated feature fusion. This models observations of a static scene, not environment transitions.

Controlled input ablations. We independently trained the variants with the same backbone, split, optimization steps, inference settings, and corrected focal supervision. Target construction and the lens-action encoder were held fixed while the scene inputs varied. Geometry-only excludes RGB from the world-state encoder; text conditioning remains available.

Table S1. Additional appearance/geometry ablations under corrected focal supervision and matched training conditions.

ConfigurationFocal Corr. ↑Scale MAE ↓Focal Aligned LPIPS ↓F-number Corr. ↑F-number Order Acc. ↑
RGB only + Lens-Action Scene Querying0.2950.4900.1910.8400.940
Geometry only + Lens-Action Scene Querying0.1900.6100.3000.7800.910
RGB + depth + Lens-Action Scene Querying0.3560.3910.1590.9340.983
RGB + point-map magnitude + Lens-Action Scene Querying0.3240.4250.1720.9080.976
Reference conditioning (RGB + geometry)0.3000.4600.1800.9100.970
Full LAWM (corrected training)0.3850.3650.1460.9460.991

All input ablations retain Lens-Action Scene Querying. The reference-conditioning row instead supplies the same RGB and geometry through a reference pathway without that mechanism. These additional experiments are distinct from the submitted Table 3.

What the results support. Relative to RGB-only input, full LAWM improves focal correlation from 0.295 to 0.385 and reduces aligned LPIPS from 0.191 to 0.146. Geometry-only has higher aligned LPIPS than RGB-only, supporting the contribution of appearance. The combined depth and point-map-magnitude inputs improve on either geometry cue alone on the reported metrics.

Compared with reference conditioning, full LAWM reduces Scale MAE from 0.460 to 0.365 and aligned LPIPS from 0.180 to 0.146, while improving f-number ordering accuracy from 0.970 to 0.991. This supports the proposed conditioning design relative to the tested baseline; it is not a claim about every reference-conditioning architecture or statistical significance.

For context, submitted Table 3 replaces Lens-Action Scene Querying with additive fusion: Scale MAE changes from 0.454 to 1.274, aligned LPIPS from 0.168 to 0.505, f-number correlation from 0.945 to 0.368, and ordering accuracy from 0.992 to 0.657. Those focal results use the original supervision and evaluator and are not directly compared with the corrected-training rows above.

Revision. We will incorporate these controlled comparisons and specify the reference injection pathway, parameter counts, and training settings. The contribution will be attributed to the tested shared-state conditioning design, not the term world model or a new attention operator.

AuthorsOpenReview ↗

Question 2: Direct Transformation Baselines

We agree that direct processing is a valid inference baseline. The training-time crop/resize and depth-guided blur operations can be applied using the same available anchor and geometry.

Matched-anchor evaluation. Supplementary Table 1 used 100 cached photographs rather than the generated anchors in the main comparison. We evaluated the operators on the same generated anchors, geometry, source/target settings, resolution, and query schedules as LAWM. The conditioning anchor and first generated observation are distinguished. All focal rows below use corrected pinhole evaluation.

Table S2(a). Additional matched-anchor focal-length comparison under corrected pinhole evaluation.

MethodFocal Corr. ↑Scale MAE ↓Adj. LPIPS ↓Aligned LPIPS ↓
Direct pinhole crop0.9980.0120.2420.009
LAWM (original weights; corrected evaluation)0.2230.4970.1490.168
LAWM (corrected training; corrected evaluation)0.3850.3650.1710.146

Table S2(b). Additional matched-anchor f-number comparison.

MethodF-number Corr. ↑Rank Corr. ↑Order Acc. ↑Adj. LPIPS ↓
Direct depth-guided blur0.9730.9950.9960.041
LAWM (original)0.9450.9920.9920.056

The f-number LAWM row uses the original checkpoint, matching the submitted results. It is distinct from the corrected-training model in Table S1, whose f-number correlation and ordering accuracy are 0.946 and 0.991.

Interpretation and limits. Direct cropping has lower Scale MAE and aligned LPIPS than LAWM; direct depth blur also has stronger f-number response scores. We therefore do not claim that LAWM is more accurate than the direct operators for these transformations. Lower adjacent LPIPS does not establish superior consistency by itself, because a weaker focal response also reduces image change.

Table S1 answers whether scene inputs and querying improve a learned model under matched conditions. Table S2 answers whether learned generation improves on direct processing. The former provides evidence for the proposed conditioning design; the latter does not establish an additional practical advantage over deterministic transformations under the current diagnostics. Such an advantage would require independent evidence, for example calibrated paired references or a documented perceptual comparison. We do not infer optical bokeh reconstruction or outpainting from these results.

Revision. We will include the deterministic baselines explicitly and distinguish improvements within learned generation from evidence for preferring generation over direct processing.

AuthorsOpenReview ↗

Question 3: Defocus Supervision and Physical Interpretation

Focus and f-number assumptions. Supplementary Section A uses the central median of normalized depth as a focus proxy dfd_f, not a measured focus distance. For source N0N_0, target NkN_k, and supported lower bound Nmin⁡N_{\min}, the mapping in Section 3.3 is

\kappa_k=\kappa_{\min}+\alpha_k(\kappa_{\max}-\kappa_{\min}).$$ For a non-source query, the approximate defocus target is $$m_k=\operatorname{clip}\left(\frac{|\kappa_k(D-d_f)|}{10},0,1\right),\quad I_k^*=(1-m_k)\odot I_{\mathrm{anc}}+m_k\odot B_{\kappa_k/3}(I_{\mathrm{anc}}),$$ where $B$ is Gaussian filtering. The source query retains the anchor. This adds blur without estimating or inverting existing source blur. Fixed focus and compensated exposure are assumptions; the model is not a calibrated thin-lens renderer. **Additional spatial-response analysis.** We used fixed near-focus and far-from-focus regions defined by $|D-d_f|$, excluding boundary margins and poorly textured areas and checking registration. These indicate focus proximity, not foreground/background labels. For regional Laplacian variance $S_r$, define $$R_r=\frac{S_r(N_{\mathrm{end}})}{\max(S_r(N_0),\epsilon)},\qquad N_{\mathrm{end}}=\min_k N_k.$$ $N_{\mathrm{end}}$ is the smallest query, distinct from the global bound $N_{\min}$; $\epsilon$ is a small positive stabilizer. This additional diagnostic uses the original checkpoint, not the corrected-training model in Table S1. **Table S3. Additional depth-stratified sharpness retention from the source setting to the smallest queried f-number.** | Method | Near-focus retention $R_{\mathrm{focus}}$ | Far-from-focus retention $R_{\mathrm{off}}$ | Selectivity $R_{\mathrm{focus}}-R_{\mathrm{off}}$ | |:--|--:|--:|--:| | Direct depth-guided blur | 0.96 | 0.42 | 0.54 | | Generative Photography | 0.71 | 0.65 | 0.06 | | LAWM (original) | 0.89 | 0.52 | 0.37 | Original LAWM retains more near-focus sharpness and has greater selectivity than Generative Photography. Direct blur remains stronger on this proxy, consistent with its mask using the same depth cues. Retention near one is desirable around focus; lower off-focus retention indicates more blur, not necessarily better quality. This supports spatial selectivity under a VGGT-derived proxy, not independent optical calibration. We will report the region-selection, filtering, registration, and aggregation settings. **Related focal-control concern.** The submitted correlations, 0.221 versus 0.219 for Generative Photography, do not establish a meaningful advantage. Corrected focal results distinguish changing the expected scale to $f_k/f_0$ from retraining with corrected supervision. **Table S4. Submitted and corrected focal results: re-evaluation versus corrected training.** | Method / Setting | Focal Corr. ↑ | Scale MAE ↓ | Focal Aligned LPIPS ↓ | |:--|--:|--:|--:| | Generative Photography (submitted evaluation) | 0.219 | 0.572 | 0.443 | | Generative Photography (corrected evaluation) | 0.220 | 0.621 | 0.443 | | LAWM (submitted evaluation) | 0.221 | 0.454 | 0.168 | | LAWM (original weights; corrected evaluation) | 0.223 | 0.497 | 0.168 | | LAWM (corrected training and evaluation) | 0.385 | 0.365 | 0.146 | Evaluation-only correction fixes the expected scale for unchanged images and registration; aligned LPIPS stays unchanged. Separate corrected training improves mean correlation from 0.223 to 0.385 and reduces mean Scale MAE from 0.497 to 0.365. All baseline outputs use the same corrected scale definition. These means do not establish statistical significance or uniformly reliable focal control. **Revision and scope.** We will clarify the focus assumptions and distinguish measured response trends from optical accuracy. Validated queries remain $f_k/f_0\in[1,2]$ and $N_k/N_0\in[0.25,1]$, varying one parameter per sequence. Reverse and joint control are not established by these experiments.
Reviewer_VpXgOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

To solve the issue that changing a lens setting inadvertently alter scene geometry or semantic layout in existing methods, this paper propose a Lens-Action World Model (LAWM) that constructs an explicit appearance-geometry world state from a reference observation and uses Lens-Action Scene Querying to query this state under different lens settings. The extensive experiments on multiple photographic controls verify that LAWM outperforms existing camera-aware diffusion models in terms of scene consistency, camera controllability, and visual fidelity.

Critical strengths

This paper proposes a simple yet effective method LAWM. The motivation is clear and the experiments is sufficient. Overall, I like the idea that query the lens action from a "persistent world state". Thus I think this paper may brings new insight for the community.

Critical weaknesses

In table 1 and 2, the paper mainly compare two camera-aware baselines, including Camera Settings as Tokens and Generative Photography. However, there are several recent relevant methods, such as CamEdit [1] and DeltaCam [2]. I recommend compare with these recent methods to evaluate the contributions of proposed method.

[1] CamEdit: Continuous Camera Parameter Control for Photorealistic Image Editing (NeurIPS 2025)

[2] DeltaCam: Differential Intrinsic Camera Modeling for Video Generation (2026)

Questions for authors

Please see the weakness.

1 comment
AuthorsOpenReview ↗

Comparison with Recent Methods

We sincerely thank the reviewer for the positive assessment and for suggesting CamEdit and DeltaCam as additional baselines. We agree that these recent methods are highly relevant to camera-controllable generation.

CamEdit (NeurIPS 2025) focuses on continuous photographic image editing, supporting aperture, focal-plane, and shutter-speed control. DeltaCam (2026) investigates differential camera-parameter modeling for video generation, including focal length, aperture, and other photographic attributes. While their task settings differ from ours, both provide valuable references for evaluating camera controllability.

Unlike these methods, LAWM maintains a persistent appearance–geometry world state and performs action-dependent querying to preserve scene consistency across lens settings.

In the revised version, we will include both methods in the related-work discussion and reproduce their approaches as additional baselines in Tables 1 and 2. We will adapt their evaluation to our experimental setting and apply the same test prompts, camera-query configurations, and evaluation metrics wherever applicable, while explicitly identifying any differences in supported controls.

We appreciate the reviewer's constructive suggestion, which will help strengthen the experimental comparison and better establish the contribution of LAWM.

Reviewer_mwpnOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses text-to-image generation under continuous focal-length and f-number control while keeping the subject and layout fixed. A draft generated by a Generative-Photography-style pipeline supplies an anchor image, VGGT extracts its geometry, and a CNN encodes both into a "world state". Each lens setting is encoded as a physical lens map plus a CLIP text difference, and Lens-Action Scene Querying (LASQ) retrieves world features and injects them into a frozen U-Net. Training targets are synthesised from single photographs by centre crop and resize (focal length) and depth-weighted Gaussian blur (f-number). On 100 prompts, the method is reported to outperform SD1.5, SD3.5, Camera Settings as Tokens and Generative Photography.

Critical strengths

The goal is relevant and the failure modes are clearly illustrated: lens changes should not rewrite identity or layout, and simply tokenising an EXIF value does not enforce this. Separating a shared scene encoding from an intrinsic-camera query is a reasonable approach to lens control. The aperture response is clearly monotonic, with f-number correlation of 0.945 and ordering accuracy of 0.992 against 0.809 and 0.900 for Generative Photography.

The ablation is informative within the training setup. Removing the physical lens map drops f-number correlation from 0.945 to 0.165, and replacing LASQ with additive fusion raises focal Scale MAE from 0.454 to 1.274. The code, configurations, baseline scripts and metric definitions are provided, which allowed me to check how targets, inference and metrics are built.

Critical weaknesses

The paper's own deterministic supervision procedure is a missing baseline. At inference LAWM has the anchor and its depth, and its training target is exactly a centre crop for focal length and a depth-weighted blur for f-number. Supplement Table 1 applies this procedure to 100 cached photographic anchors and obtains focal correlation 1.000 vs 0.221, Scale MAE 0.005 vs 0.454, aligned LPIPS 0.004 vs 0.168 and f-number correlation 0.981 vs 0.945. Since the procedure uses only test-time inputs, it should be included as a baseline rather than a calibration reference, and it should be evaluated on the same generated anchors as the main tables.

The focal-length targets and Scale MAE use the wrong scale. For a fixed viewpoint, magnification is f_k/f_0, but the released code crops by the minimum of the angular field-of-view ratios (genphoto_style.py, lines 172–189), and Supplement Eq. 4 defines the expected scale as the mean of inverse angle ratios. For 24→48 mm this gives about 1.893× in training and 1.843× in the metric against the correct 2.0×, and the gap grows for wider lenses (about 1.621× in training and 1.536× in the metric for 10→20 mm). A geometrically correct output therefore receives a non-zero error, and the model is trained toward the wrong magnification.

The comparison and evaluation do not isolate the contribution. All baselines generate every frame from text, while LAWM conditions on a fixed anchor and its depth, so the consistency comparison is confounded by this difference. Focal correlation (0.221 ± 0.589) is essentially the same as that of Generative Photography (0.219 ± 0.742). The metrics (SIFT scale against an analytic curve, global Laplacian variance) check exactly what the crop-and-blur supervision produces, on 100 prompts at 256×384 without real paired data. The action space is also restricted to zooming in and opening the aperture; zooming out and stopping down are never tested.

Questions for authors
  1. What are the results of applying crop and depth-weighted defocus to the same generated anchors used in Tables 1–2, and which properties of LAWM outputs does this pipeline fail to reproduce?
  2. Why are angular field-of-view ratios used instead of the pinhole focal-length ratio in both the targets and Scale MAE, and how do the corrected metrics change the rankings?
  3. How does the model behave when zooming out, stopping down, or changing both parameters together?
5 comments
AuthorsOpenReview ↗

Question 1: Direct Transformations and Matched Conditioning

Thank you for the detailed assessment and implementation checks. We agree that direct transformations are valid baselines: they use the same inference-time anchor and estimated geometry.

Matched-anchor comparison. Supplementary Table 1 used cached photographs rather than the generated anchors in the main evaluation. We therefore evaluated direct processing on the same generated anchors, geometry, source/target settings, and resolution as LAWM. The conditioning anchor and first generated observation are distinguished. All focal rows below use corrected pinhole evaluation, explained in Reply 2.

Table R1(a). Additional matched-anchor focal-length comparison.

MethodParam. Corr. ↑Scale MAE ↓Adj. LPIPS ↓Aligned LPIPS ↓
Direct pinhole crop0.9980.0120.2420.009
LAWM (original weights; corrected evaluation)0.2230.4970.1490.168
LAWM (corrected training and evaluation)0.3850.3650.1710.146

Table R1(b). Additional matched-anchor f-number comparison.

MethodParam. Corr. ↑Rank Corr. ↑Order Acc. ↑Adj. LPIPS ↓
Direct depth-guided blur0.9730.9950.9960.041
LAWM (original)0.9450.9920.9920.056

The f-number LAWM row uses the original checkpoint. The corrected-training checkpoint is separate and appears in Table R3 below.

Interpretation. Direct cropping achieves lower Scale MAE and aligned LPIPS than LAWM; direct blur also has stronger f-number response scores. These results do not establish a property that LAWM reproduces better than direct processing under the reported diagnostics. A lower adjacent LPIPS alone is insufficient, since a weaker focal response can reduce image change. Additional practical benefits of generation would require separate evidence, such as calibrated paired references or a documented perceptual comparison. We do not claim outpainting, optical bokeh reconstruction, or optical correction from these tables.

Separating anchor access from scene querying. The main tables compare complete pipelines and do not isolate LAWM's access to an anchor. We added the following controlled learned variants, with the same backbone, training setup, and corrected focal supervision. Target generation and the lens-action encoder were held fixed while scene inputs and fusion varied.

Table R3. Additional input and reference-conditioning ablations under corrected focal supervision.

ConfigurationFocal Corr. ↑Scale MAE ↓Aligned LPIPS ↓F-number Corr. ↑Order Acc. ↑
RGB only + Lens-Action Scene Querying0.2950.4900.1910.8400.940
Geometry only + Lens-Action Scene Querying0.1900.6100.3000.7800.910
Reference conditioning (RGB + geometry)0.3000.4600.1800.9100.970
Full LAWM (corrected training)0.3850.3650.1460.9460.991

The input ablations retain Lens-Action Scene Querying, the mechanism described in Section 3.4. Geometry-only removes RGB from the world-state encoder, not text conditioning. The reference-conditioning baseline receives RGB and geometry without that mechanism. Full LAWM improves on this baseline and the RGB-only variant on the reported means, supporting geometry and querying within the tested learned architecture. This does not show that learned generation is preferable to the direct operators.

The submitted Table 3 is separate evidence under the original protocol: replacing Lens-Action Scene Querying with additive fusion changes Scale MAE from 0.454 to 1.274, aligned LPIPS from 0.168 to 0.505, f-number correlation from 0.945 to 0.368, and ordering accuracy from 0.992 to 0.657. Its absolute focal values are not compared with the corrected-training rows.

Revision. We will add these comparisons and document the reference injection pathway, parameter counts, and training settings. The contribution claim will distinguish improved conditioning within learned generation from an independently established advantage over direct processing.

AuthorsOpenReview ↗

Question 2: Focal Geometry and Corrected Evaluation

The reviewer is correct. The original implementation used angular field-of-view ratios as an approximate source-relative control region and reused them for target construction and evaluation. This conflated angular extent with linear image magnification. We should have separated these quantities.

Corrected geometry. Under a fixed viewpoint, sensor, principal point, and output framing,

r_{\mathrm{crop}}=\frac{\tan(\phi_d(f_k)/2)}{\tan(\phi_d(f_0)/2)}=\frac{f_0}{f_k},$$ where $s_d$ is the sensor extent along image dimension $d$. The expected image scale is $s_k=f_k/f_0$. Thus, the crop fraction is the inverse of magnification, not the ratio of field-of-view angles. **Re-evaluation of existing outputs.** We recomputed correlation against $f_k/f_0$ and log-scale MAE for every compared method: $$\mathrm{ScaleMAE}=\frac{1}{|\Omega|}\sum_{k\in\Omega} \left|\log\hat{s}_k-\log(f_k/f_0)\right|.$$ Here, $\hat{s}_k$ is the scale estimated by image registration, and $\Omega$ is the set of valid scale comparisons used by the evaluator. **Table R2(a). Submitted metrics and corrected evaluation of the same existing outputs.** | Method | Original Corr. ↑ | Corrected Corr. ↑ | Original MAE ↓ | Corrected MAE ↓ | |---|---:|---:|---:|---:| | Stable Diffusion 1.5 | -0.173 | -0.174 | 0.796 | 0.834 | | Stable Diffusion 3.5 Medium | 0.110 | 0.109 | 0.627 | 0.673 | | Camera Settings as Tokens | -0.273 | -0.275 | 0.488 | 0.535 | | Generative Photography | 0.219 | 0.220 | 0.572 | 0.621 | | LAWM | 0.221 | 0.223 | 0.454 | **0.497** | All methods have higher Scale MAE under the corrected expected magnification. Original LAWM retains the lowest MAE among the learned methods listed, but its focal correlation remains close to Generative Photography. The direct-crop baseline is reported separately in Reply 1 and has substantially lower error. **Correction of training supervision.** Re-evaluation does not change the old model's targets. We therefore separately trained LAWM with the corrected focal construction. **Table R2(b). Evaluation-only correction versus corrected-training LAWM.** | Configuration | Focal Corr. ↑ | Scale MAE ↓ | Aligned LPIPS ↓ | |---|---:|---:|---:| | LAWM (original training and evaluation) | 0.221 | 0.454 | 0.168 | | LAWM (original weights; corrected evaluation) | 0.223 | 0.497 | 0.168 | | LAWM (corrected training and evaluation) | **0.385** | **0.365** | **0.146** | The second row changes only the expected-scale definition for fixed outputs; the third is a separate corrected-training experiment. Aligned LPIPS remains unchanged in evaluation-only correction because the images and registration procedure are fixed. Corrected training improves mean correlation from 0.223 to 0.385 and reduces mean Scale MAE from 0.497 to 0.365. We do not infer statistical significance or uniformly reliable focal control from these means. **Metric scope and revision.** SIFT scale and Laplacian variance diagnose magnification and sharpness responses, not calibrated optical fidelity. The additional depth-stratified analysis uses the original checkpoint and a VGGT-derived focus proxy; it is separate from corrected training. We will revise Section 3.3 and Supplementary Section C.2, distinguish re-evaluation from retraining, and qualify the focal-control claims accordingly.
AuthorsOpenReview ↗

Question 3: Reverse and Joint Lens Control

These capabilities are outside the validated scope of the submitted model. Section 4.1 uses

fk/f0∈[1,2],Nk/N0∈[0.25,1],f_k/f_0\in[1,2],\qquad N_k/N_0\in[0.25,1],

and varies one lens parameter per query sequence. The same model learns both controls, but this does not establish simultaneous control within a query.

Zoom-out. Crop-based supervision supplies narrower fields of view, not the scene regions outside an anchor. The world state represents observed anchor evidence; a diffusion prior does not establish reliable reconstruction of unseen surroundings. We therefore do not claim validated zoom-out generation.

Stopping down. The supervision adds defocus and does not invert source blur. The bounded f-number representation also saturates beyond the source setting, so it does not provide distinct stopping-down conditions. Recovering a sharper observation would require an extended representation and suitable supervision. We do not claim validated stopping-down behavior.

Joint control. Parameter-specific focal-length and f-number features do not guarantee independent composition. Focal length and f-number interact in depth-of-field formation, and the submitted model is trained with one variable changed per sequence. Simultaneous queries require separate modeling and validation.

Reversing an existing sequence would not establish reverse control: that requires generating wider or sharper observations from a newly supplied narrower-view or more-defocused anchor.

Revision. We will state the single-variable, one-sided operating ranges more prominently in the task definition and conclusions. Our claims concern continuous source-relative queries within those ranges, rather than arbitrary bidirectional or compositional lens editing.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Rebuttal Summary and Clarifications

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair and all reviewers for their careful evaluations and constructive feedback. We appreciate the recognition of our motivation and the potential of lens-action querying over a persistent scene representation. We have addressed the major concerns in our individual responses and would like to summarize the key clarifications and additional evidence.

1. Corrected Focal-Length Geometry and Evaluation.

We acknowledge the inconsistency in the original angular field-of-view approximation and have corrected the focal-length formulation using the exact pinhole relationship. We provided a consistent re-evaluation of all compared methods and separately examined the effect of corrected training supervision. The results clarify the improvements and remaining limitations of focal-length controllability.

2. Additional Baselines and Contribution Validation.

We introduced matched-anchor deterministic crop and depth-guided blur baselines, together with controlled ablations of anchor appearance, VGGT geometry, and reference-conditioning mechanisms. These comparisons help isolate the contribution of Lens-Action Scene Querying (LASQ). Importantly, we acknowledge that deterministic transformations achieve stronger scores on their directly corresponding control metrics. Our contribution is therefore more precisely characterized as improving action-conditioned consistency within learned generative modeling, rather than universally outperforming deterministic image processing.

3. Physical Control and Model Limitations.

We clarified the focus-depth approximation, f-number-to-defocus mapping, and treatment of existing source blur. We also provided depth-stratified sharpness analysis to complement global sharpness metrics. Furthermore, we explicitly acknowledge that the current framework supports source-relative zoom-in and aperture-opening queries, while reverse and simultaneous lens control remain outside the validated scope. These clarifications distinguish controllable photographic responses from fully calibrated optical simulation.

4. Comparison with Recent Methods.

Following the reviewers' suggestions, we will extend the related-work discussion and incorporate CamEdit and DeltaCam into our experimental comparisons, adapting their implementations and evaluation protocols where applicable while accounting for differences in supported controls and generation tasks.

Overall, we believe these additional analyses and clarifications provide a more rigorous understanding of LAWM's contributions, particularly the role of shared appearance–geometry representations and action-dependent scene querying. We will incorporate the additional experiments, corrected formulations, and clarified limitations into the revised manuscript, ensuring that the conclusions remain aligned with the available evidence.

We sincerely appreciate the reviewers' valuable feedback and hope that our detailed responses help address the concerns regarding technical soundness, experimental fairness, and contribution validation.

Thank you again for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.