Supplementary resultsAnonymous submission
Reviewer 1Supplementary experiments & responses

Target reliability &
multi-view refinement

Responses on target reliability and multi-view refinement, with highlighted regions and three-view result comparisons.

Route · Diagnose · RefineEvidence accompanying the author response
Orange circles identify target discrepancies; green frames identify Ours; red frames identify rejected guidance, an ablated setting, or a baseline comparison.
Focus circles and enlarged detail crops guide the comparison. The source renderings are unchanged.
01

Question 1 · visibility and diagnosis

Completed visibility and false-alarm experiments

“Can you show that this does not happen, for example by computing the score only over views that see each region?”
Controlled experimental setting

We completed two complementary experiments. The case-level analysis uses 50 manually screened cases with correct reference edits. Source and target images are rendered from consistent source/edited 3D assets with matched cameras. The native 3D state is fixed to the correct edited latent, separating visibility effects from image-target inconsistency.

A. Case-level disagreement by visibility

We compare both-view-visible and one-view-visible regions. The latter are split by the unseen-view prediction error relative to the observing-view error: a ratio at or below one defines the lower-relative-error group; a larger ratio defines the higher-relative-error group. “Lower” is a relative comparison, not a certificate of absolute correctness.

We summarize disagreement within each case and group before aggregating case-level statistics. The score is normalized by the conflict threshold, D / τC: one denotes the conflict threshold.

Measured case-level disagreement across 50 cases.
Visibility / prediction-error group Mean D / τC
Both views visible 0.71
One view visible · lower relative prediction error 0.69
One view visible · higher relative prediction error 0.93
Case-level disagreement · completed experiment Dots show case-level statistics; diamonds and error bars show across-case means and the reported 95% confidence intervals. The dashed line marks D / τC = 1.

Mean disagreement is similar for both-visible regions (0.71) and lower-relative-error one-view regions (0.69), but increases to 0.93 in the higher-relative-error group. Reduced visibility alone therefore produces little change in the lower-relative-error group, while inaccurate hidden-region predictions are associated with elevated disagreement. A mean below one does not imply that every cell or case is below the conflict threshold.

B. Does occlusion create false alarms?

We also completed a visibility-stratified false-alarm experiment. With the state fixed to a correct reference edit, a correct cell classified as conflicting is a false positive. For each case and visibility group, the false-positive rate is the number of evaluated correct cells with D > τC, divided by the number of evaluated correct cells in that group.

We compare regions visible in one view with those visible in at least two views within each case, then aggregate case-level rates. This prevents larger assets from dominating through a pooled cell count.

Visibility-stratified false alarms · completed experiment Connecting lines pair the same case's two visibility strata. The measured false-positive rates are of comparable magnitude, with a modest difference between the plotted means and no pronounced increase under single-view visibility.

Manuscript: Appendix F.3, Table 10 and Figure 7. These experiments do not establish equivalence to a support-restricted score; pairwise disagreement is undefined with fewer than two supporting views.

02

Question 2 · reliability of the feedback signal

Target inconsistency: filtering and region locking

“How often are these targets inconsistent across views, and can you show failure cases where an inconsistent target leads to a wrong edit?”

Thank you for highlighting this distinction. Cross-view flow disagreement can reflect both an incomplete 3D edit and incompatible 2D guidance. We do not interpret it as a guaranteed detector of geometric error.

Candidate targets are screened before refinement. An average of 2.3 targets is rejected per composition case. This count describes screening activity; it is not a calibrated percentage of inconsistent targets. The examples below show why rejecting incompatible guidance matters.

A. What filtering prevents

The retained sofa example contains a small residual pillow mismatch. The rejected example is deliberately used to show how incompatible pillow-placement guidance can damage the resulting composition.

Q2 / ACCEPTED

Sofa target retained · small pillow-placement mismatch

A target retained by filtering despite a residual discrepancy. Compare both target views with the final 3D result.

Target · view 1
Target · view 2

Pillow region · enlarged

Target 1 · pillow and armrestInspect the placement relative to the adjacent armrest and seat.
Target 2 · pillow and armrestSmall mismatch in the same local relationship.

Three rendered views per asset · click any image to enlarge

Ours✓ Retained target · Ours
Ours · view 1
Ours · view 2
Ours · view 3
What to inspect

The target pair is retained despite a small residual mismatch. The three output views above show the resulting pillow–sofa composition.

Q2 / REJECTED

Sofa target rejected · large pillow-placement change

A target rejected by filtering is deliberately used for this diagnostic run. This result illustrates what can happen when the filter is bypassed.

Target · view 1
Target · view 2

Rejected pillow guidance · enlarged

Target 1 · pillow locationCompare the supporting surface and nearby armrest.
Target 2 · changed placementThe discrepancy is a placement constraint, not just a change in image viewpoint.

Three rendered views per asset · click any image to enlarge

Filter bypassed✕ Rejected target deliberately used
Filter bypassed · view 1
Filter bypassed · view 2
Filter bypassed · view 3
What to inspect

This target was rejected by filtering. The result shown above deliberately bypasses that decision to expose the effect of incompatible guidance; it is a diagnostic counterfactual, not an accepted pipeline output.

B. Why an inconsistent target need not corrupt the asset

Target generation is instructed to focus on the interaction region. In addition, region locking limits where the next round may change the 3D asset. A target may therefore contain an irrelevant discrepancy outside the allowed edit region without transferring that discrepancy to the 3D result.

This protection is conditional on the edit/lock masks being correct and the corresponding region remaining locked in that round. It does not resolve contradictory targets inside the interaction region. The two examples below illustrate this distinction; they do not establish that all target inconsistencies are harmless.

Q2 / LOCK 01

Teddy bear · inconsistent leg stripes outside the interaction

Inspect the target discrepancy outside the editable region, then compare the three rendered views of the result to check preservation.

Target · view 1
Target · view 2
Edit / lock maskVisualization of the editable and locked regions.

Leg stripe discrepancy · enlarged

Target 1 · stripe appearanceGreen and black bands in the first target.
Target 2 · stripe appearanceCompare band colors and boundaries.

Three rendered views per asset · click any image to enlarge

Ours · region locking enabled✓ Ours · locked region
Ours · region locking enabled · view 1
Ours · region locking enabled · view 2
Ours · region locking enabled · view 3
Location of the discrepancy

Focus on the leg stripes rather than the cup. Under the stated locking rule, target differences outside the allowed interaction edit do not directly authorize changes to this leg region.

Q2 / LOCK 02

Rider on a horse · hat difference on the upper back

Inspect the hood-bearing upper-back area in the rear-view comparison below.

Target · view 1
Target · view 2
Edit / lock maskVisualization of the editable and locked regions.

Hat on the upper back · rear-view detail

Rear target · upper backInspect the upper-back hood location in the target.
Rear output · hood on the backInspect the blue hood on the upper back.

Three rendered views per asset · click any image to enlarge

Ours · region locking enabled✓ Ours · locked region
Ours · region locking enabled · view 1
Ours · region locking enabled · view 2
Ours · region locking enabled · view 3
Location of the discrepancy

Use the rear-view pair to inspect the hood difference. The intended rider–horse edit concerns the interaction, while the hood-bearing upper-back region is shown separately for preservation.

03

Question 3 · priority ablation

Does new-view priority help?

“How do results change if earlier views take priority instead, or if the order is ignored?”

The new-view priority rule is motivated by the acquisition policy: a new camera is chosen to expose a region that remains unresolved. Its update is routed only to regions with geometric support, and the priority rule gives it local precedence there. This is a design rationale rather than a guarantee of target correctness.

We compared priority enabled and removed across multiple cases. The representative helmet–headphone example below uses the displayed target views and highlights the visible earcup difference between settings.

Q3 / PRIORITY ABLATION

Helmet and headphones · effect of new-view priority

A representative case from comparisons with priority enabled and removed. Related-method comparisons are presented separately below.

Earlier target view
Newly acquired target view

Three rendered views per asset · click any image to enlarge

Ours · new-view priority✓ Priority enabled
Ours · new-view priority · view 1
Ours · new-view priority · view 2
Ours · new-view priority · view 3
Ours · no priority✕ Priority removed · ablation
Ours · no priority · view 1
Ours · no priority · view 2
Ours · no priority · view 3

Same output view · right earcup detail

Priority enabled · earcupThe rim and earcup body remain coherent.
Priority removed · earcupThe same region is visibly deformed in this run.
Observed difference

Inspect the circled right earcup, particularly in view 2. Removing priority visibly distorts its shape; the enabled-priority result retains a more coherent earcup structure in this case.

Figure