acceptodds
Under review as a conference paper at ICLR 2027

Characteristic-Steered Flow Matching for Oracle-Defined Sets

Abstract

Generating samples that satisfy complex physical or structural constraints remains a fundamental challenge when valid regions cannot be expressed in closed form. For instance, finding valid candidates, such as a stable crystal structure or a collision-free robot trajectory, fundamentally relies on evaluating proposals and rejecting failures against a binary verifier. This problem can be expressed as learning distributions over oracle-defined sets, where admissibility is accessible only through pointwise verification. We study how to adapt continuous generative flows in this regime, where certified reference samples seed a base transport, but further constraint satisfaction must be learned from oracle evaluations that provide no gradient feedback. We introduce Characteristic-Steered Flow Matching (CSFM), a method that converts discrete membership evaluations into directional supervision along generative paths. By pairing each inadmissible candidate with a certified feasible anchor, CSFM transports the required terminal correction backward along the generative trajectory via reverse-time linearization, supervising the velocity field to realize first-order endpoint corrections without differentiating the oracle. Across synthetic geometric sets, physics-constrained continuous control, and long-horizon trajectory planning, CSFM achieves favorable validity–coverage and downstream trade-offs relative to existing flow-correction baselines, preserving coverage and distributional fidelity as ambient dimension scales.

Primary area:
generative models
Keywords:
Flow Matching, Constrained Generation, Membership Oracles, Generative Maps, Characteristic Steering

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_egjgOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces Characteristic-Steered Flow Matching for constraints accessible through a binary membership oracle. It pairs invalid samples with certified feasible anchors, then transports the endpoint corrections backward through a reference flow to construct velocity-training targets. An optional runtime repair handles remaining violations. Experiments on synthetic sets, constrained control, and trajectory planning examine whether this improves validity while preserving coverage and useful generative structure.

Critical strengths

The paper aims to solve the discrete membership problem into a constraint supervision along generation path which solves a timely question. The stengths are:

  1. Pairing rejected samples with feasible anchors gives a concrete correction, and the reverse-flow construction provides a clear first-order justification for distributing it along the trajectory. This is a neat way to use an oracle without requiring constraint gradients.
  2. The paper shows good comparison over existing method: the dimension-scaling results show cases where CSFM maintains coverage while FM-RE collapses. The single-round planning results are also encouraging: CSFM achieves high safety and success at substantially lower latency than HardFlow.
Critical weaknesses

I think a further ablation will help: Table 15 ablates the steering loss, correspondence choice, and correction strength, but does not compare the reversed flow transported r against simply using r with the same anchors and training setup. Continued flow-matching training already gives similar validity in several ablation cases. It therefore remains unclear how much of the improvement specifically requires the paper’s central transport mechanism, rather than a gentler anchor-based correction.

The planning results leave me somewhat unsure when the extra adaptation is worth it compared with inference-time search and repair. In Table 1, Base + CEM already matches CSFM + CEM on avoidance safety and success and achieves higher mean safety and success on the maze task at similar latency. CSFM does show a clear benefit in the single-round setting, so I am not suggesting adaptation is unnecessary. However, a search-budget sweep would help show where this benefit holds and how much inference effort the extra training actually saves.

Questions for authors
  1. Could you compare CSFM against an otherwise identical correction baseline i.e. using A_r be identity matrix, keeping the anchors, correction strength, and training budget fixed? This would directly test whether characteristic transport is needed.
  2. Could you provide a candidate/search-budget sweep beyond the existing single-round and CEM settings, comparing CSFM and the base model under matched inference costs?
  3. Relatedly, it would be useful to discuss Physics-Constrained Flow Matching (PCFM), which applies constraint corrections during inference without retraining the generator. This raises a broader question about learning corrections versus applying them during sampling.
4 comments

We thank the reviewer for the positive assessment and for precise, actionable remarks: the identity control and the search-budget sweep are now in the paper. To improve our presentation, the revised PDF defines both Suite C deployments (Sec. 4.3) and provides their per-replan oracle budgets (App. C.2, Table R7), expands metric abbreviations where they are introduced, and states which correspondence rule each suite uses (Sec. 3.4).

W1 / Q1. CSFM vs. an otherwise identical correction with A_t = I.

All three reviewers asked for ablations that isolate the transport. Yours, the same correction with A_t = I and identical anchors, alpha, and budget, is now in the paper. Validity, seeds matched across arms (all cells: Table R2):

CellBaseIdentity (A_t = I)CSFM
Box d=320.6740.8910.962
Box d=640.1130.2050.273
Ball d=160.9240.9390.955
Solid Torus d=160.9670.9760.973
Solid Torus d=320.9860.9880.990

Transport adds validity on 10 of 13 Suite A cells, most where the base fails most. On three Solid Torus cells and on both Suite B control tasks, the reviewer's "gentler anchor-based correction" indeed suffices: the two tie, and identity is often more uniform. On continued FM: where the base fails, as on Box d=32, CSFM gains clearly, and it is never below continued FM in validity or coverage on any analytic cell (General Response, point 2). The ablations requested by the other reviewers are in point 1 of the General Response.

W2 / Q2. When is adaptation worth it vs. inference-time search, at matched inference cost?

The sweep below makes explicit a benefit we had under-reported: inference time. With a good base, CEM search is very strong, but slow. To reach CSFM's single-round outcome, the base needs CEM, at about 3 times the time per replan in avoidance and twice in maze. We fully concede the maze row of Table 1: with CEM, the base is higher on both safety and success at similar latency. Adaptation therefore pays when search is limited or latency-bound; this is what we expect: removing edge cases from the flow will reduce expressivity, and our goal is to do that minimally while reducing errors. Oracle calls and time per replan for every point (Table R7, Fig. R-B):

Avoidance

Proposals per replanOracle callsCSFM safe (%)CSFM success (%)Base safe (%)Base success (%)Time CSFM (ms)Time base (ms)
12591799262825
4100949270223026
16 (single round)400999874283127
641,60010010058293030
CEM (3 x 1,024)4181001001001008181

Maze

Proposals per replanOracle callsCSFM safe (%)CSFM success (%)Base safe (%)Base success (%)Time CSFM (ms)Time base (ms)
1254197797241224
410078981197264233
16 (single round)400991002998271242
641,60099986699286274
CEM (3 x 256)400979899100514511

(same evaluation seeds for all methods, 50 episodes per seed.)

Q3. PCFM: learning corrections vs. applying them during sampling.

Thank you for the pointer; we have added PCFM (Utkarsh et al., NeurIPS 2025) to the related work. We do not ablate against it because it requires a differentiable constraint: its Gauss-Newton projection onto equality constraints needs the residual and its Jacobian, whereas the whole point of CSFM is a setting with only a binary membership oracle over sets of positive volume. The two are complementary: PCFM corrects each sample at sampling time; CSFM learns part of the correction into the velocity field and keeps a membership-only repair for deployment.

We hope these additions strengthen the reviewer's support for the paper, and we would be glad to add any further comparisons.

Thank you again for your review and your suggestions. We would like to know whether the identity control and budget sweep address your questions, and we are happy to answer any remaining ones.

I thank the authors for extra experiments and results. The identity ablation is what I was looking for and the budget sweep also answers my question about CEM. The explanation of PCFM makes sense. Given the different constraint information it requires, I’m fine with discussing it as related work without a direct comparison. However, though the method shows advantage with limited inference time, overall evidence still seems mixed to me to rate as a strong accept and hence I will maintain my weak accept evaluation.

We agree that the evidence is mixed: CSFM helps most where the base fails or inference time is limited, and the final version can state this scope more clearly. Thank you for your support.

Reviewer_ofuMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript is about how to make flow-based generative models generate samples inside a feasible set that is only defined by a binary oracle, i.e., the oracle only tells whether a sample is valid or not and gives no gradient. This work proposes Characteristic-Steered Flow Matching (CSFM), which turns the binary rejection into directional supervision for the velocity field, and uses a chord-bisection projection at inference time to guarantee the output is accepted.

The interesting and novel part is the characteristic steering. Each rejected endpoint x−x^- is paired with a certified anchor x⋆x^\star, and the displacement r=x⋆−x−r = x^\star - x^- is transported backward along the reference characteristic by the inverse Jacobian rt=Jˉ1→t(x−) rr_t = \bar{J}_{1\to t}(x^-)\, r with one reverse JVP, so that adding rtr_t at every time tt moves the endpoint by exactly rr to first order. The velocity field is then regressed to vˉ(xt,t)+αrt\bar{v}(x_t, t) + \alpha r_t together with the standard FM loss.

Critical strengths
  1. The work is well motivated, constraints that are only accessible through a binary oracle are common in materials, molecules, and planning, and turning the rejection signal into directional supervision is a reasonable angle.
  2. The method is clean: anchor pairing + inverse-Jacobian transport + velocity regression.
  3. The experiments show good results, especially the one-shot planning in Suite C.
Critical weaknesses
  1. The gain over simple alternatives is not clear. Table 15 shows that at d≤16d \le 16, λ=0\lambda = 0 (just 5,000 more FM updates) already reaches the same validity as CSFM, so the improvement over Base in Fig. 2/3 mainly comes from extra training. In Suite B the return never exceeds Base, and in Suite C, Base + CEM with repair already reaches 100% safe / success, the same as CSFM + CEM. So it seems like the method only clearly helps on Box at d=32d = 32 and in the single-round setting of Suite C?
  2. The oracle budget does not match the motivation: the abstract motivates with expensive oracles (crystal stability, molecules), but training needs about 4×1074\times10^7 queries and all tested oracles are cheap analytic checks.
Questions for authors
  1. It is unclear whether the characteristic transport is really needed. Given the pairs (z,x−+αr)(z, x^- + \alpha r), a simpler way is to regress the corrected endpoint directly as an FM target, and another natural choice is to transport with the adjoint (VJP) instead of the inverse Jacobian (JVP). Neither is compared. Adding these two ablations would help show the necessity of characteristic steering?
  2. Since λ=0\lambda = 0 in Table 15 already matches CSFM at d=16d = 16, should the Base rows in Fig. 2/3 be the model with 5,000 more FM updates?
  3. Why does FM-DD use product coupling while the other methods use Sinkhorn OT coupling? How does it perform with the same coupling?
5 comments

We thank the reviewer for a demanding review; each request became a new experiment (Tables A to D, next comment).

W1 / Q2. Is the gain just extra training?

Our problem asks for validity together with coverage and uniformity, which the paper now states (Sec. 3). CSFM does not always lead in validity, but on every analytic cell neither continued FM, FM-RE nor FM-DD matches it on all three metrics, with 10 times fewer queries than FM-RE.

Yes, the lambda = 0 row of Table 15 is this model. We now show it in Figs. 2/3 and Tables 6 to 8, adding the cells it was missing (Table A), and plot validity during adaptation (Fig. R-A).

  • Continued FM: CSFM is at least as valid and as broad on all 18 analytic cells. The two tie only where the base is already near the validity ceiling, as on Ball and Solid Torus in Table 15; on the harder cells CSFM gains clearly (Table A).
  • Suite B: on most tasks, more admissible actions than the base without projection, and the same return with projection.
  • Suite C: one round of CSFM approaches the base with CEM search, which takes about 3 times longer per replan in avoidance (Table R7).

W2. Oracle budget vs. the expensive-oracle motivation.

We used analytic oracles for two reasons: with an expensive oracle every training run is costly, ruling out the study we needed (multiple seeds, and baselines that need many more oracle calls to be competitive; App. D.1.2); and coverage, uniformity and distortion need Vol(F) and uniform reference samples, which analytic sets provide. Materials and molecular oracles remain future work.

We then reduced the oracle budget for both methods (Table B). At a quarter of its queries CSFM keeps its coverage, keeps nearly all its validity, and stays more uniform than FM-RE at any budget; at the same reduced budget the coverage of FM-RE collapses.

Q1. Is characteristic transport needed?

Both ablations are now in the paper (Table R2). They use CSFM's own nodes, anchors, alpha, budget and checkpoint rule; only the correction differs (Table C).

(a) The corrected endpoint as an FM target falls below the base on every cell tested, suggesting the correction is most useful along the flow's own paths.

(b) VJP uses the adjoint along the same reverse path, Jˉt→1⊤r\bar J_{t\to1}^\top r, the gradient of −12∥x1−x⋆∥2-\tfrac12\|x_1-x^\star\|^2 with respect to xtx_t. At its own length it collapses under CSFM's solver, optimizer and clipping: a few failures have very large adjoints. We extended the reviewer's idea: rescaled at every node to the length of CSFM's transported correction AtrA_t r, the adjoint direction reaches CSFM's validity on most cells, though it is less uniform on most (Tables R2, 15). The rescaling needs the JVP, so we read this as support for the transport: for validity, it mainly supplies the right step length at each node.

Q3. FM-DD coupling.

FM-DD is privileged: it reads the analytic distance to F, while CSFM only queries membership. We had used the authors' settings; as the reviewer notes, its product coupling is inconsistent with our OT-trained bases, so we reran it with Sinkhorn coupling on all its cells; Tables 6 and 8 keep the submitted run as FM-DD (product). Matched coupling helps FM-DD on most cells, but CSFM still covers at least as much of F on every cell, more on almost all, and is more uniform on almost all (Table D); FM-DD is more valid on some cells.

We hope these experiments address the concerns and merit a higher assessment.

Seeds matched across arms; ED x 1000.

A. Continued FM vs. CSFM (validity)

CellContinued FM validityCSFM validity
Box d=80.8450.958
Box d=160.7650.962
Box d=32 (your Table 15)0.6700.962
Box d=640.0720.273
Solid Torus d=80.9100.953
Solid Torus d=640.7020.819
Solid Torus d=16, m=320.8910.956
Ball d=16 (your Table 15)0.9510.955
Solid Torus d=16 (your Table 15)0.9700.973
Solid Torus d=32 (your Table 15)0.9860.990

B. Reduced oracle budget

Oracle queriesArmBall d=16 coverageBall d=16 EDBall d=16 validityBox d=32 coverageBox d=32 EDBox d=32 validity
4.1e8FM-RE (submitted)0.98337.41.0001.0004.380.999
4.1e7FM-RE0.72291.10.9961.00015.20.996
4.1e7CSFM (submitted)0.99710.40.9550.9991.960.962
1.0e7FM-RE0.2711920.9470.6503310.646 (a)
1.0e7CSFM0.99613.90.9520.9993.000.938

(a) one seed collapses (zero validity and coverage).

C. Endpoint target and VJP (validity)

CellBaseEndpoint as FM targetVJPCSFM
Box d=320.6740.5400.1970.962
Box d=640.1130.1000.0520.273
Ball d=160.9240.8680.4160.955
Solid Torus d=160.9670.8420.3480.973
Solid Torus d=320.9860.8960.3100.990

D. FM-DD couplings

CellFM-DD (product) coverageFM-DD (product) EDFM-DD (product) validityFM-DD coverageFM-DD EDFM-DD validityCSFM coverageCSFM EDCSFM validity
Ball d=20.5802450.9500.73255.80.9750.9831.850.945
Box d=80.91599.20.9650.9925.790.9920.9941.030.958
Box d=160.98366.10.9340.98156.30.9350.9971.470.962
Solid Torus d=160.89875.70.9270.86187.10.9341.0001.000.973
Thin slab d=100.7880.830.2680.6970.090.2160.9390.260.367

Thank you again for your review; your questions led us to conduct several new experiments for the paper. As the discussion period ends soon, we would like to ask whether our response addresses your concerns. We are happy to answer any remaining questions.

Reviewer_aq6pOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces CSFM, a way to improve a generative model when validity is determined only by a yes/no checker. It pairs invalid outputs with known valid examples, which creates a straight chord guaranteed to intersect the boundary. By multiple such pairing, one may opportunistically characterizing the boundary crossing to establishing tight training target. One can then work backward through the generation process to construct training corrections. A separate repair step handles remaining invalid outputs. Experiments cover synthetic shapes, constrained robot actions, and trajectory planning. The results show a useful balance between validity, diversity, and task performance, particularly compared with randomized-exploration adaptation.

Critical strengths
  1. The method turns known valid examples into directional training guidance without needing derivatives from the checker is interesting.
  2. With proposal selection and repair, CSFM reaches high safety and success at substantially lower reported latency than existing methods.
Critical weaknesses
  1. The experiments do not isolate the benefit of backward transport. Existing ablations show that steering helps in some settings, but do not establish whether transporting corrections backward is better than simpler correction supervision.

  2. The generation evaluation partly reuses checkpoint-selection states. This leaves uncertainty about performance on fully independent states. Reporting the remaining 448 states separately would directly address the concern.

  3. The theoretical result explains the effect of a small correction along a fixed reference trajectory, but does not establish that the actual training procedure improves validity, preserves diversity, or controls distortion. The validity guarantee comes from a separate repair step and requires an available valid anchor. Consequently, the method’s main performance benefits remain empirical, and the current ablations do not fully establish why backward transport is responsible for them. This may not be a major weakness if the paper is viewed as a pure system/application paper.

Questions for authors
  1. With identical anchors and training settings, how does CSFM compare with applying the endpoint correction directly along the reference path, without backward transport?

  2. In Suite C, how does continuing ordinary flow-matching training for the same number of updates compare with CSFM, under the identical proposal-selection and repair pipeline?

2 comments

We thank the reviewer for a sharp and constructive review. We ran both requested experiments and corrected the evaluation issues we found; new material is in blue in the revised PDF.

W1 / Q1. Endpoint correction without backward transport.

We ran it. With the same failures, anchors, updates, and oracle budget as CSFM, we applied the endpoint correction at every node of the path without transport, and also used the corrected endpoint as a plain FM target (validity, seeds matched across arms; Table R2):

CellBaseCorrected endpoint as FM targetCorrection without transportCSFM
Box d=320.6740.5400.8910.962
Box d=640.1130.1000.2050.273
Ball d=160.9240.8680.9390.955
Solid Torus d=160.9670.8420.9760.973
Solid Torus d=320.9860.8960.9880.990

Without transport, the correction recovers part of the gain; transport adds validity on Box and Ball, most where the base fails most, and they tie on Solid Torus, near the validity ceiling. As a plain FM target, the corrected endpoint falls below the base on every cell. Further baselines requested by the other reviewers are in point 1 of the General Response.

W2. Evaluation partly reuses checkpoint-selection states.

The reviewer is right: checkpoints were selected on 64 of the 512 evaluation states, and our tables reported all 512. We recomputed every such result on the 448 states never used for selection (Table R5). Validity and coverage are unchanged and the other metrics move little; no ranking between methods changes beyond two near-ties. Julia is evaluated on separate held-out scenes and is unaffected.

W3. The theory is first-order; the benefits remain empirical.

We agree: the analysis motivates the training target without guaranteeing what training achieves. Two clarifications: all Suite A metrics are on raw decodes without repair, so those gains come from training; and repair requires an available anchor, which the paper states as Assumption 1.

Q2. Suite C: continued FM for the same updates, same pipeline.

We ran exactly this: the training of CSFM with the correction weight set to zero, deployed through the identical proposal, verification and repair pipeline (same evaluation seeds for all methods, 50 episodes per seed; safety and success in %):

Avoidance

Proposals per replanCSFM safeCSFM successCont. FM safeCont. FM successBase safeBase success
191799057926
16 (single round)999881747428
6410010084765829
CEM (3 x 1,024)100100100100100100

Maze

Proposals per replanCSFM safeCSFM successCont. FM safeCont. FM successBase safeBase success
141972063797
16 (single round)9910053702998
64999877706699
CEM (3 x 256)9798766799100

In a single round of 16 proposals, CSFM clearly leads continued FM in safety and success in both tasks. Continued FM improves the base in avoidance, but only full CEM search closes the gap; in maze, it loses some of the base's goal-reaching. Table R7 and Fig. R-B give the full sweep, with the time per replan at each setting.

The evaluation issue in W2 was a real flaw in our reporting, and we thank the reviewer for finding it. We hope the revision supports a more positive assessment.

Thank you again for your review, and in particular for spotting the overlap between selection and evaluation states. As the discussion period ends soon, we would like to ask whether our response and the two new experiments address your concerns. We are happy to answer any remaining questions.

Comments on the paper

2 comments

We thank the reviewers for their careful and constructive reading. Every review led to a substantial improvement of the paper and helped us sharpen its contributions.

The reviews show the paper was well understood; we still restate the main challenge it addresses, since the new results are best read against it. Raising validity alone is easy: a model can pay for it with coverage or uniformity. The challenge is to fix what the base gets wrong while disturbing as little as possible what it gets right: coverage, uniformity over F, and a regular map (now stated in Sec. 3). On every analytic cell, neither continued FM, FM-RE nor FM-DD matches CSFM in validity, coverage and uniformity together.

Our answers to the concerns shared across reviews follow. Key numbers are in the next comment (Tables G1 to G3), the rest in the new Appendix G. Changes to the paper are in blue. Main text: design goal (Sec. 3); continued FM and transport ablation (Sec. 4.1); scope (Sec. 4); Suite C sweep (Sec. 4.3); PCFM (Sec. 5). Appendix G holds the new tables and figures. Fixes: selection states described as a subset of the evaluation states; FM-DD now Sinkhorn-coupled, with the submitted run kept in the appendix.

  1. Role of the transport (all reviewers). The main point of all three reviews was that the submission did not show that the gain comes from the backward transport. Each reviewer proposed a control, and we ran each as a separate study, from the same base weights and with the anchors, alpha and budget of CSFM:
  • Corrected endpoint as a plain FM target (ofuM): it falls below the base on every cell tested, suggesting the correction should follow the flow's own paths.
  • The same correction without transport, A_t = I (egjg, aq6p): it recovers part of the gain, but transport adds validity on most Suite A cells, most where the base fails most (Table G1). They tie on three Solid Torus cells and on both Suite B control tasks, all near the validity ceiling.
  • In addition, a training-free measurement shows why (Table G2): transport delivers close to the requested step on most environments, while the step of the untransported correction varies widely across environments.
  1. More training (ofuM, aq6p). The submission already contained continued FM, on four cells (lambda = 0, Table 15), but it was not clearly marked as an ablation. We moved it next to the base in the main figures and tables and completed it on every Suite A cell, with the same updates and oracle budget (Table G3; Fig. R-A plots validity over the adaptation updates). It does not explain the gain: CSFM is at least as valid and as broad on every analytic cell and more uniform on almost all.

  2. Other new results.

  • Reduced oracle budget (ofuM): at a quarter of its queries, CSFM keeps its coverage and nearly all its validity, and stays more uniform than FM-RE at any budget.
  • FM-DD with matched coupling (ofuM): Sinkhorn coupling helps FM-DD on most cells, but CSFM still covers at least as much of F on every cell and is more uniform on almost all.
  • Held-out states (aq6p): on the states never used for checkpoint selection, the submitted results are essentially unchanged.
  • Inference budget in Suite C (egjg, aq6p): it is true that with full CEM search the base matches or exceeds CSFM. We added a budget sweep: a single round of CSFM comes close to what the base reaches only with CEM, which in avoidance costs the base about 3 times the time per replan.

3 seeds; ED x 1000.

G1. Transport vs. identity

CellBase validityCont. FM validityCont. FM EDIdentity validityIdentity EDCSFM validityCSFM ED
Box d=320.6740.6702.000.8911.250.9621.96
Box d=640.1130.07243.50.20541.20.27341.3
Ball d=160.9240.95113.20.9399.210.95510.4
Solid Torus d=160.9670.9701.250.9760.980.9731.00
Solid Torus d=320.9860.9861.840.9881.550.9901.70

G2. One correction step, no training

EnvironmentDelivered / requested: transportDelivered / requested: identity
Box d=320.980.58
Box d=641.010.64
Ball d=161.030.29
Solid Torus d=321.131.15
HalfCheetah1.001.06

G3. Validity, continued FM vs. CSFM

CellContinued FMCSFM
Box d=20.9650.984
Box d=80.8450.958
Box d=160.7650.962
Box d=320.6700.962
Box d=640.0720.273
Ball d=20.9370.945
Ball d=80.9110.951
Ball d=160.9510.955
Solid Torus d=20.9470.973
Solid Torus d=80.9100.953
Solid Torus d=160.9700.973
Solid Torus d=320.9860.990
Solid Torus d=640.7020.819
Solid Torus d=16, m=40.9590.966
Solid Torus d=16, m=80.9370.962
Solid Torus d=16, m=160.9240.962
Solid Torus d=16, m=320.8910.956
Thin slab d=10 (a)0.3180.367
Julia d=2 (test scenes)0.5000.500

(a) fluctuates between checkpoints.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.