- 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
-
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.
-
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
-
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.
-
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
-
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?
-
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?
-
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
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.
| Method | Focal Corr. ↑ | Scale MAE ↓ | Adj. LPIPS ↓ | Aligned LPIPS ↓ |
|---|---|---|---|---|
| Direct pinhole crop | 0.998 | 0.012 | 0.242 | 0.009 |
| LAWM (original weights; corrected evaluation) | 0.223 | 0.497 | 0.149 | 0.168 |
| LAWM (corrected training; corrected evaluation) | 0.385 | 0.365 | 0.171 | 0.146 |
Table S2(b). Additional matched-anchor f-number comparison.
| Method | F-number Corr. ↑ | Rank Corr. ↑ | Order Acc. ↑ | Adj. LPIPS ↓ |
|---|---|---|---|---|
| Direct depth-guided blur | 0.973 | 0.995 | 0.996 | 0.041 |
| LAWM (original) | 0.945 | 0.992 | 0.992 | 0.056 |
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.
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 , not a measured focus distance. For source , target , and supported lower bound , 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.
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.
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.