acceptodds
Under review as a conference paper at ICLR 2027

Vectile: Learning Reusable Transport References for Counterfactual Generation

Abstract

A visual reference shows a response under particular conditions; editing often requires that response under another stimulus, installation, or arrangement of interacting objects. We introduce transport references: reusable local responses rather than the trajectories or illumination patterns observed in a reference. VECTILE extracts a response tile, installs it at a target site, and rolls out the installed tiles through TILEFLOW, where the target repeatedly supplies their current inputs and states. Visual synthesis then expresses the resulting consequence in images or videos. We show that this separation supports both optical interactions, where outgoing influence changes what other sites receive, and active-material deformation, where changing shape alters subsequent exposure. Our optical implementation couples local response and typed transport to spatial-goal-guided diffusion; the material implementation supplies camera-specific controls from one shared predicted world trajectory. Across 68 configurations of a held-out active-material world, VECTILE produces 52 correct edits versus 6 for our VIPER-style/Wan implementation, with 16 low-visibility configurations. On a relighting benchmark, VECTILE recovers seven edits missed by our natively adopted LightMover-style/Wan baseline.

Primary area:
generative models
Keywords:
representation learning, counterfactual generation, compositional generative modeling

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_nRHZOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a method for reusing a way of responding from a reference. The proposed method encodes a reference into a response representation and installs it at sites in a target scenes. It then rollouts local responses with a target-dependent updates, implemented as a heat simulation for materials with a learned light transport. The experiments show that proposed method results in 52 correct material edits (out of 68 total) against 6 for a baseline.

Critical strengths

Disentangling a reusable component of response from the scene-dependent component is an interesting framing, and such factorization could allow for a more controlled interventions.

Critical weaknesses
  1. The paper considers narrow, constructed tasks with a highly engineered method. The material benchmark is a synthetic family of four responses defined by hand-specified state equations. The proposed method uses those equations, and only predicts their coefficients. All four responses are seen in training, and the “held-out” set is configurations of a single world. The task therefore largely reduces to identifying one of four known responses and running the ground-truth simulator.
  2. The evaluation is limited and does not support the claims the paper makes. The material comparison is not like-for-like. The baseline gets an RGB clip and text, while the proposed method gets explicit state, geometry, stimulus, transition labels and a simulator. The baseline’s six “successes” are scenes where the correct output is no motion, and the extended baseline fails its own training controls.
Questions for authors

Please see Critical Weaknesses.

3 comments

You have reduced our study of reference-conditioned transport to a four-way classification problem, then used that reduction to dismiss its scientific contribution. This is a ridiculous evaluation.

The history of our field makes your evaluation very difficult to justify. ControlNet (ICCV 2023) introduced spatial control through a pretrained generator; DragonDiffusion (ICLR 2024) constructed explicit editing energies; ObjectMover (CVPR 2025) and LightMover (CVPR 2026) established controllable interventions through specialized conditions and generated or rendered supervision. In learned physics, NCLaw (ICML 2023) retained governing PDEs while learning constitutive behavior, and PhysDreamer (ECCV 2024) inferred material properties within an established elastic model and solver.

These choices have a constructive necessity. Learning arbitrary physical laws, reconstructing geometry, modeling interactions, and achieving faithful visual synthesis simultaneously would make both experimental attribution and computational cost nearly unmanageable. Constructive research isolates the dependency worth understanding, builds a controlled way to manipulate it, and tests the resulting capability. This is how advances in the field have been made possible and verifiable.

If a specified physical model, an existing solver, and deliberately constructed interventions are sufficient to dismiss our contribution, what consistent scientific principle would qualify the contributions of PhysDreamer, LightMover, and the established control methods above?

Their value cannot be erased by reducing them to their inherited components. Neither can ours. The relevant question is what new operation those components have been organized to make possible, and what the experiments establish about it.

We selected two existing problems with complementary transport difficulties. And we went substantially further than demonstrating that this organization produces plausible edits.

We manually examined the LightMover authors' publicly released evaluation outputs, identified their lighting propagation failures, and showed that, in fact, we could do better. LightMover's own evaluation cases 183/185/187/200 exposed relay, branched reflections, directional response, and occluded cascades. These are failures from an established evaluation. We did not invent these tasks to favor us. In fact, this motivated us to distinguish local response, propagation, and visual expression.

Our submitted Fig. 4 already shows a direct visual win. The comparator changes one kitchen receiver but leaves the other almost white; Vectile restores the missing blue response. We also have six additional joint-movement edits recover emitter placement and visible response.

For the material problem, in Fig. 3, we independently install responses in an ordered pairing excluded from target training and validation. The same responses produce different downstream behavior under different layouts, correctly observed from two cameras. In Table 15, we impose a stricter intervention: replace only the front reference while keeping the rear response and target conditions fixed within each matched pair. The rear changes its motion. Across source and installation settings, the very same replacement produces different downstream consequences. All 16 generated views succeed in this additional development factorial.

We deliberately designed these experiments because reproducing a reference motion would prove far too little. The rear sheet has no new reference, and neither observation contains its counterfactual future. Our experiment tests whether a locally reusable response can participate in a new interaction whose consequences are determined by the target.

We then examined the computation itself. With the response model fixed, removing temperature, activation, and conduction reduces numerical world-response success from 12/12 to 4/12; freezing renewed exposure geometry changes 14/14 to 11/14 (Table 4). These interventions distinguish carried response history from target-mediated propagation. They provide empirical information about which dependencies our transport abstraction must preserve that we consider valuable to the field.

We chose established problems, identified a reusable transport operation they could expose, imposed strict additional interventions to distinguish genuine response reuse from reproducing an observed outcome, and demonstrated improvement on prior work's failure modes.

Your review gives almost none of this scientific work its proper place. The very constraints that make our experiments interpretable somehow become your grounds for dismissal, and it is unclear if our demonstrated work actually received your careful evaluation.

Thank you for the rebuttal to me review! Unfortunately, it does not address my concerns.

Weakness 1. The rebuttal defends the use of a specified physical model and solver by analogy to prior work. That is not my issue. My issue is that all four responses are seen in training (App. C.3) and the update equations and solver are given to the method, which predicts only their coefficients (App. B.2). The response does not dispute either fact. The experiments it cites (Fig. 3, Table 15, Table 4) use the same four known responses and the same prescribed equations. They show that the simulator produces different outcomes under different layouts and states, not that a response outside the training set can be learned and reused.

Weakness 2. The response does not address the input mismatch in Table 1, the fact that the baseline's six successes are static controls (App. E.1), or the fact that the extended baseline fails its own training controls (Table 14). The Table 15 result (16/16 vs 0/16) is measured against that same baseline. Table 4 is judged on the numerical rollout, not on generated output, and shows that a heat-driven simulation depends on its state and geometry.

I am keeping my score.

Thank you for your review!

While we want to point out the important distinction between the "four response" in training and the full response the tests actually require to generate, we recognize our setting may not make it obvious enough. It is also related to your concern about testing actual generative output. To address your concern, we added another material experiment to demonstrate Vectile's own learned capability. We make a shell bend-unbend-bend under continuous heat, creating a non-monotonic transition in the middle of its response property. None of the training data demonstrates the non-non-monotonic transition, so this non-monotonic response is guaranteed to be never seen during training. This also makes nearest-neighbor impossible to recover the correct response. At test time, we show Vectile a reference of this bend-unbend-bend material. Vectile, not nearest-neighbor, successfully generated this non-monotonic behavior. We added this result to Appendix J in our new revision, page 34.

On your input mismatch concern, we want to make two points: first, all information that Vectile uses is also semantically given as input to VIPER-Style/Wan. For example, we gave VIPER-Style/Wan the initial heat state as text. The main difference is interface structure. Second, in established literature, it is a common practice to give information to the baseline in its own format. In ObjectMover, LightMover, and Ctrl&Shift, the geometric or spatial conditions are all adapted to each baseline's supported interface that may be less structured than the authors' methods.

The baseline performance may give you concern about the comparison. However, firstly, it does receive extra state and geometry information as our last note points out, and secondly, in our training setting and budget it does indeed perform very badly. VIPER-Style/Wan may perform differently if given a different budget or training setup, but that shouldn't invalidate the particular finding and report in our paper.

We also want to emphasize that our VIPER-style/Wan comparator is not an off-the-shelf, prompt-only video generator. It implements VIPER's dedicated reference-physics conditioning pathway, including 128 learnable physics queries through Qwen3-VL-4B, an MLLM-to-DiT connector, and LoRA adaptation of Wan2.1-VACE-1.3B, with all three alignment, transfer, and generator-adaptation stages completed (2,200 updates). The original VIPER is itself a competitive reference-conditioned physics-generation method, reporting a physical-similarity score of 3.42 compared with 2.14 for VACE in the authors' evaluation. Our comparator is a smaller, independently trained implementation rather than the original authors' checkpoint, and we explicitly document this distinction. It's not an unadapted or deliberately weakened baseline.

Thank you again for reviewing our paper. If our new information addressed your concern, would you kindly raise your score?

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

This paper argues that a visual reference should transfer a local response law rather than an observed outcome, so that the target supplies the inputs and determines the downstream consequences. Vectile extracts a compact response tile from a reference, installs it at target sites, alternates local response evaluation with target-dependent updates (TileFlow), and renders the result with a diffusion model. It is implemented for active-material sheets, where a 32-D code predicts six coefficients of a hand-specified thermal/activation model solved by LibShell and rendered through Wan2.1-VACE, and for optical relighting with a latent typed transport over 256 sites. The main results are 52/68 correct configurations against 6/68 for a VIPER-style baseline on one held-out synthetic world, and seven recovered edits among 69 failures of a LightMover-style baseline.

Critical strengths

The conceptual distinction is clear and the target phenomenon is well chosen. Separating response identity from target-dependent consequences is a clean way to think about reference-guided counterfactual editing, and Figure 3 illustrates it well: the same tiles in a new layout produce a new consequence, seen from a second camera on the same process.

The intervention experiments address the right question. With the response model fixed, zeroing temperature, activation and conduction reduces world-response successes from 12/12 to 4/12, and freezing exposure geometry reduces them from 14/14 to 11/14 (Table 4). The paper is also unusually transparent about scope, including per-configuration outcomes, costs, negative results of the earlier optical checkpoint, the single-world caveat and the statement that the seven optical cases are not a success rate.

Critical weaknesses

The material experiment does not separate the learned response from access to the simulator. The constitutive equations, coefficient bounds and shell solver are fixed in advance, and all four teacher responses come from the same six-parameter family and occur in training, so the encoder only needs to recognise which of the four seen teacher responses produced the reference. Vectile also receives explicit initial state, geometry, stimulus and bindings, and VACE is conditioned on renders of the simulated trajectory, whereas the VIPER-style baseline receives one RGB clip and text. Removing the rendered controls drops Vectile from 5/5 to 0/5. The 52-vs-6 comparison therefore does not isolate the learned response; the same pipeline with oracle coefficients, or with a nearest-neighbour choice among the four training identities, is needed to establish what the learned tile adds. The extended multi-slot baseline that is meant to equalise inputs fails even its unmodified training controls (0/4), so it does not provide a usable input-matched comparison.

The generalisation test and the scoring are too narrow to support the headline numbers. All 68 configurations come from one held-out world at 256×256 and 17 frames, no experiment uses a response law outside the four training identities, and 16 configurations are low-visibility for both methods. Outcomes are scored by the authors or by non-blind automated assessment with only qualitative criteria and no second rater or agreement statistic. Expected rest under cold or no-stimulus controls counts as correct, and the six baseline successes are such static cases, so motion-required and rest cases should be reported separately.

The optical half does not support the proposed mechanism. On the full 193-request set, the original optical checkpoint is worse than the baseline (42 vs 46; Table 19), and on a separate 100-image set it gets none of the reference-replacement, receiver-installation, source-movement or joint edits right (Table 20). The final H2 model is evaluated only on 69 prior baseline failures that were inspected during development, and recovers seven, without a count of how many of the 46 baseline successes it retains. The alternative H1 architecture recovers the same seven cases, and the authors state that H2 was chosen for architectural clarity.

From Section 3 onward the paper introduces many terms specific to the paper (transport reference, response tile, installation, TileFlow, legal condition, world packet, H1/H2) and adds lengthy qualifications to most sentences. The abstract's "52 versus 6" does not mention that all six baseline successes are rest controls, and "recovers seven edits" omits the 69-failure denominator.

Questions for authors
  1. What does VACE conditioned on LibShell renders achieve with the oracle teacher coefficients, and with coefficients chosen by nearest-neighbour matching to the four training identities?
  2. Can references with response coefficients outside the four training identities, and multiple held-out worlds, be evaluated?
  3. For the final H2 checkpoint, what are the correct, incorrect and ambiguous counts over all 193 optical requests, including retained and regressed baseline successes?
3 comments

We find your weakness summaries very confusing. You said the optical results do not support our proposed mechanism, but then you showed that with our fully operational mechanism, in particular the development from original checkpoint to final model, is the reason why we finally outperformed baseline. Isn't that result precisely supporting our mechanism? We made it clear in paper that H2 is our final model, whille Tables 19–20 evaluate our original checkpoint that is not our final model: it already computed four interaction rounds, yet only delivered their fields through a single projection. The required receiver often remained unchanged. We reported this because it showed that computing transported influence alone does not ensure that diffusion expresses its consequence. That is the challenge that makes realizing our reusable transport a difficult problem.

So, your optical conclusion uses the failures that motivated our mechanism as evidence against the mechanism.

In fact, the “qualifications” you criticize in our paper explain precisely this progression and gives the full picture of the whole mechanism we actually proposed.

Our method is Extract → Install → TileFlow → Visual Synthesis (Section 3). H2 explicitly separates local response and geometry-conditioned transport, converts interaction fields into spatial latent/RGB goals, and uses repeated control across Wan layers with denoising feedback. These are our proposed transport-to-generation mechanism.

The recovery difference is human-visible in Fig. 4 we presented. The qualitative difference is a much more useful signal than whole-picture color difference for our relighting task that cares about local semantic response. Our LightMover-style comparator recolors one kitchen receiver but nearly misses the other; H2 restores it. Six joint edits recover emitter placement and visible consequences. Other baseline failures are also part of the report, and note that the some other authors have substantial more computing budget than us, so the denominator is large, that is normal. Asking for more numbers here is not meaningful. We achieved observable success in our experiments on failure modes of lightmover's own publicly released evaluation, and as you noticed, our checkpoints and experiments show this can be attributed to our proposed mechanism. We consider that has reached a reasonable bar for our field.

For the shell material part, we first want to emphasize that having a second task itself is our way to demonstrate the generalization capability of our methods, and it's a more difficult challenge than just testing held-out different worlds.

In learned physics, it is common to incorporate solver as additional prior and study the structured gain, so that we can avoid parts that we already well understood. For example, NCLaw (ICML 2023) retains governing PDEs while learning constitutive behavior; PhysDreamer (ECCV 2024) uses an established material model and simulator to execute inferred material properties under new conditions.

The simulator is only part of our proposed computation and it does not supply all the counterfactual answer.

Please look at what is actually computed in Section 3 and Appendix B.2. The reference code is queried together with the current temperature and activation. The resulting response updates local state and deformation; LibShell computes the constrained geometry; the changed geometry then determines the next heat exposure and incoming influence. That entire response → state → shape → renewed exposure loop is TileFlow. The target future does not exist in the input. Our method has to produce it.

On your demand for input-matched comparisons: Table 1 discloses the additional authored initial conditions and supervision used by Vectile. These differences matter only for the native system comparison, and we never present it as an isolated estimator comparison. But matching information available before the intervention is entirely different from handing a baseline the future trajectory, exposure, and VACE controls computed by our method. The latter would give it the results of TileFlow itself. What exactly would such a comparison still tell us about whether our proposed computation is necessary to realize the editing operation?

And we went further than simply running a known response through a solver. Figure 3 tests an unseen ordered composition under changed layouts. Table 15 replaces only the front reference while the rear response remains fixed, yet its downstream motion changes with the target. Table 4 distinguishes carried state from geometry-dependent propagation through fixed-model interventions.

We think you already partially recognized some value of our experiments in your assessment for which we appreciate, and we hope with these explanations you can better understand the context and the value our results entail.

Thanks for the response! The full definition of TWT_W and the aggregate numbers over all optical requests clear up my earlier question about regressions on the optical side so I have raised my score to 2.

The material experiment is still my main issue. I am not against using a solver. The question is how much of the reported result comes from the learned tile versus the fixed equations the solver and the rendered controls. The ablation I suggested runs your own pipeline with the encoder swapped for the true coefficients or a nearest neighbour pick among the training responses. It does not hand the baseline anything from TileFlow. Without it and with all experiments staying within the four seen responses it is hard to tell what the learned part contributes.

On the optical side the aggregate numbers look like the baseline counts plus the recovered cases. Could you confirm that H2 outputs on the remaining cases were actually re inspected. The paper itself notes that the gains cannot yet be attributed to the transport component specifically.

The points on input mismatch rest only baseline successes and non blind scoring were not addressed.

Thank you for your update!

We added a material experiment to demonstrate Vectile's own learned capability. We make a shell bend-unbend-bend under continuous heat, creating a non-monotonic transition in the middle of its response property. None of the training data demonstrates the non-non-monotonic transition, so this non-monotonic response is guaranteed to be never seen during training. This also makes nearest-neighbor impossible to recover the correct response. At test time, we show Vectile a reference of this bend-unbend-bend material. Vectile, not nearest-neighbor, successfully generated this non-monotonic behavior. We added this result to Appendix J in our new revision, page 34.

For optical case numbers, yes, actually, in our rebuttal to another reviewer @fxuG https://openreview.net/forum?id=zf9EzlRQyT&noteId=xfAeVtvyKP, we gave a detailed re-inspection of those indeterminate cases with more qualitative details which is also done in a blind way. We think their similar failure modes and successes on other cases may be explained by their shared WAN backbone and training data.

"The paper itself notes that the gains cannot yet be attributed to the transport component specifically." Yes, but we honestly do not see why this is a concern about our paper. What our paper studied in section 5 and appendix are internal ablation of Vectile's components and our result suggests it cannot be reliably reduced to an isolated component of our mechanism. In addition, and perhaps a more valuable insight, is that it reveals a fundamental dependency in the task: the downstream consequence actually depends on target and not the transport operator alone.

On your input mismatch concern, we want to make two points: first, all information that Vectile uses is also semantically given as input to VIPER-Style/Wan. For example, we gave VIPER-Style/Wan the initial heat state as text. The main difference is interface structure. Second, in established literature, it is a common practice to give information to the baseline in its own format. In ObjectMover, LightMover, and Ctrl&Shift, the geometric or spatial conditions are all adapted to each baseline's supported interface that may be less structured than the authors' methods.

Would you let us know if you have any remaining concerns?

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

This paper studies whether behavior observed in a reference can be reused when an object’s surroundings or mounting conditions change. VECTILE uses the reference to model how a material or surface responds and uses this model to guide image or video generation under new conditions. The authors evaluate the method on heat-induced sheet deformation and light editing. For material deformation, VECTILE produces 52 correct configurations out of 68 in one held-out world with four known responses, compared with 6 for a locally implemented VIPER-style baseline. Both methods have 16 low-visibility configurations. For light editing, the authors examine 69 cases where their LightMover-style baseline failed and report seven successful edits by VECTILE on this development set.

Critical strengths
  • The paper studies an interesting problem: how to reuse a material’s response when its surroundings or mounting conditions change. For example, one sheet can block heat from reaching another. Changing the first sheet’s response can alter the heat reaching the second sheet and thus its motion, even when the second sheet’s response is held fixed.
  • The material experiments provide evidence that responses can be reused across the tested reference views and mounting conditions. They also report successful edits for a pair of known responses that was excluded from target training (Table 3).
Critical weaknesses
  • The optical transport operator TWT_W needs a more complete definition. Appendix B.3 describes its dimensions, edge indexing, and normalization, but does not explain how geometry determines propagation between sites or how the four field modes enter this computation. Without these details, it is difficult to reconstruct the update in Equation (4).

  • The reported results for the current H2 model do not directly show what happens when only the reference is changed. Six of the seven cases where H2 succeeds and the baseline fails use fixed references during joint movement edits, while the remaining case uses no independent reference. These examples demonstrate successful editing, but H2 also retains other conditioning paths (Appendix B.4). A controlled comparison that changes only the reference or its binding, with the other inputs and generation settings fixed, would help determine whether the generated local and downstream effects change as expected.

  • The authors report seven H2 successes among 69 cases where the baseline fails, but do not report aggregate H2 results across all 193 requests. These results show improvements on individual cases, but leave open whether H2 also fails on cases the baseline handles correctly. A complete summary is needed to assess its overall performance relative to the baseline.

Questions for authors
  • Could you provide the full computation for TWT_W, including how geometry and edge features determine propagation between sites and how the four field modes enter this computation?
  • For the final H2 model, have you tested replacing only the reference or changing its binding while keeping target geometry, light sources, and generation settings fixed? If so, do the local and downstream effects change as expected?
  • Could you provide aggregate results for the current optical H2 model across all 193 requests, including the number of cases where the LightMover-style baseline succeeds but H2 fails? This would help assess H2’s overall performance relative to the baseline.
3 comments

We want to first explain the reason why we chose to test changing targets instead of changing reference in our tests.

Changing a reference while keeping the world fixed would test reference sensitivity. That would not, by itself, establish that a response has been transferred. A generator could pass that test simply by copying a different highlight or color pattern. Our scientific question is more demanding: can a response acquired once continue to act correctly when the source, geometry, visibility, and receivers change, producing consequences never observed in the reference? This requires holding the reference fixed and changing the target. The new consequences are precisely what we need to examine.

In production development, we don't consider a material reusable just because swapping it for another produces a different image under the same lighting. A collection of baked appearance maps could pass that test. The real challenge begins when an approved material moves into a different shot lighting: the key light changes, an occluder blocks an indirect path, or a new receiver enters the scene. Artists expect the same material to respond correctly without rebuilding its appearance for every shot. This is why established workflows such as MaterialX and OpenPBR invest in reusable material descriptions whose behavior can be evaluated under changing scene conditions. If a generative model merely changes its output when given another reference, it has demonstrated something far weaker than the capability these workflows actually need. The meaningful test is to keep the reference fixed, change the world, and demand a correct consequence that the reference never contained. That is the capability we chose to investigate.

That is why our six optical joint edits deliberately preserve the independent reference responses while changing target conditions. Source position, target geometry, and visibility provide the new queries and propagation conditions. They should not be memorized as part of the reference identity. Our formulation also supports the complementary operation. In material Table 15, we change only the front reference while keeping the rear response and target conditions fixed within each matched pair. The rear nevertheless changes motion, and the same reference replacement produces different downstream effects across target configurations.

On the details of our optical transport operator: Our optical transport computes how each site's outgoing response becomes another site's incoming influence under target geometry. On a 256-site grid, each sender–receiver pair contributes a 9-D feature containing relative direction, distance, normal alignment, local size, and ray-tested visibility. Appending both endpoint normals gives 15 features, which a 15-to-32-to-16 MLP maps into a learned 4-by-4 mode-mixing matrix A_ij = reshape4x4(sigmoid(MLP(edge,n_i,n_j)))/4. A shared 16-by-16 matrix B mixes field channels. Each receiver accumulates the outgoing fields of all senders through:

T_W(O)[i,c,m] = sum_(j,c',m') alpha_ij * A_ij[m,m'] * B[c,c'] * O[j,c',m']

where alpha_ij = V_ij*q_i*q_j/max(1,sum_l q_l) accounts for visibility and site validity. Starting from the source field, four rounds alternate local response and target-dependent propagation:

outgoing = P_reference(incoming, context); incoming = source + T_W(outgoing)

The installed reference remains fixed while geometry determines the renewed inputs.

The definition we use is one concrete realization of reference-conditioned transport, without restricting the identity of a reusable reference to a particular optical parameterization. Our material implementation follows the same separation through carried state, shape, and renewed exposure.

We now provide the complete optical aggregate for all 193 requests. Our paper’s original assessment for both methods gives 53 correct, 62 incorrect, and 78 indeterminate for H2, versus 46 correct, 69 incorrect, and 78 indeterminate for LightMover-style. H2 recovers seven baseline failures, with no baseline success regresses.

To give a more resolved comparison on the rest ambiguous cases (note if we want to be strict, all those cases should be fail since they do not show the intensity level of required change. Here we want to see whether we can find cases where h2 fails while baseline succeeds), we did another round of inspection that is more permissive for semantic quality, and we ended up labelling 51 of those 78 cases as successful for both methods: 44 brightness edits show the requested direction of change, while seven color edits pass when secondary color spill or weaker intermediate responses are accepted. This raises the success counts to 104 for H2 versus 97 for the baseline, still with seven recoveries and zero regressions. The remaining 27 brightness cases comprise 22 whose source-to-target change just cannot be reliably distinguished and five whose outputs do not demonstrate the requested dimming.

We also have added an optical experiment result to address your concern in our new revision (Appendix K, page 35). We show in four different scenes changing only reference can indeed produce different downstream effects.

I appreciate the authors’ response. Most of my concerns have been addressed, so I’m raising my score from 2 to 3. My confidence remains limited because this is outside my main area of expertise.

Comments on the paper

1 comment

Thank you for the reviews. The authors posted rebuttals on 12 Nov; please read them and respond before the discussion closes.

@fxuG: The authors now give the full definition of T_W and the full results over all 193 optical requests (H2 53/62/78 vs LightMover-style 46/69/78; 7 fixes, 0 regressions). Do these resolve your first and third concerns? Do you accept Table 15 as the reference-only test you asked for, and does your score change?

@KcGD, @nRHZ: Your shared concern was that the material result does not separate the learned response from the fixed solver and the extra inputs. The authors defend the design by analogy to NCLaw and PhysDreamer, but did not run the true-coefficient or nearest-neighbour ablations. Does the rebuttal change your assessment? Please say which concerns remain and whether any are resolved.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.