acceptodds
Under review as a conference paper at ICLR 2027

Guided Observation-Space Projections for Constrained Flow Matching

Abstract

Generative models of physical fields are often required at inference time to satisfy observations or physical constraints that were not enforced during training. Projection-based samplers meet them by correcting the predicted final sample onto the constraint set at every step, but a gradient-guidance step taken beforehand is largely lost in that correction and in the interpolation that follows. We introduce guided observation-space projection (GOSP), which keeps a guided displacement across repeated projection by updating the source sample with the current state, and characterise the guidance that survives a linear projection. Because the projection runs at every step, we solve it in observation space without forming the constraint Jacobian, with a matrix-free and a structured kernel-Gram solver. On Navier–Stokes benchmarks GOSP improves reconstruction from sparse and moving observations while keeping the constraint residuals of a projection sampler.

Primary area:
generative models
Keywords:
Generative Models, Flow Matching, AI for Science

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_SivTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GOSP is an inference time method for conditioning pretrained flow models on observations and physical constraints which combines gradient guidance with endpoint projection and updates the stored source variable so that guidance is retained when the next state is reconstructed by interpolation. The analysis characterizes the guidance component that survives linear projection and gives a local extension to nonlinear constraints. The paper also develops matrix-free and structured kernel-Gram implementations to avoid explicitly constructing the constraint jacobian. Experiments are done navier stokes and shallow-water models.

Critical strengths
  • i like how the paper makes a clear case for why PCFM style re-interpolation weakens guidance. The fix is simple, and the linear analysis helps explain the mechanism. I found the velocity and streamline ablations convincing. For instance removing source charging substantially reduces the reconstrucion benefit.
  • the structured projection is also a useful pratical contribution. The kernel Gram formulation avoids large Jacobians, and Tab. 16 provides a numerical check against explicit assembly. Plus high-resolution results show why this matters: the structured solver works where dense assembly runs out of memory and is substantially faster than matrix-free CG at comparable residuals.
Critical weaknesses
  1. I am not yet convinced the experiments show a broad advantage for source charging. I appreciate the held-out and noisy observation evaluations, but neither includes the independently tuned Euler form combination or the no-charge variant, and the DPS plus-projection comparison is limited. Tab. 9 supports improvements on the main velocity and streamline references, but the other tasks offer limited evidence: GOSP matches PCFM on the integral sets, and removing charging barely changes kinetic-energy reconstruction. Appendix B also notes that the Navier–Stokes data only vary the forcing phase for a given initial vorticity. The gains seem useful, but their broader practical significance is still unclear.

  2. The tuning and evaluation split for noisy observations needs clarification. In Appendix E.1, the trajectory indices k mod 10 and (17 + 41k) mod 100 repeat every 100 references. Tuning on k = 100,...,103 therefore reuses the trajectories of references 0,...,3, although the sensor layouts and noise differ. If evaluation uses k = 0,...,19, there is overlap at the trajectory level, but the evaluation indices are not listed. This matters because tuning uses the true fields and several reported differences are small. The reference-free guidance rule also leaves unclear how to choose projection onset and damping schedules for a new constraint.

Questions for authors
  1. Which k values are used for the 20 noisy-observation evaluation references? Given the periodic indexing in Appendix E1, do any share an initial-condition/forcing pair with the tuning references? If so, could you repeat the comparison with disjoint trajectories, rather than just different sensor layouts and noise?

  2. Could you compare the no-charge variant, the independently tuned euler-form combination, and DPS with a suitable late/final projection budget across the existing held-out settings? Using the same validation protocol and reporting reconstruction error, constraint residuals, and failures would help clarify whether the charging gains extend beyond the main-reference ablations.

  3. How were the projection onset, damping, and root-search schedules chosen? Were any evaluation trajectories used for tuning? I would like to see how the method performs when both guidance and projection settings are tuned on independent validation trajectories, then transferred without access to the evaluated reference fields.

1 comment

We thank the reviewer for the careful reading.

Overlap in the noisy-sensor test. you are right. We evaluated on references 0 to 19 and tuned on references 100 to 103. Because the indexing repeats every 100, references 100 to 103 use the same trajectories as references 0 to 3, with different sensors and noise. We checked this in two ways.

First, we dropped references 0 to 3 and recomputed the medians on the 16 references that share nothing with tuning. CRPS (lower is better), with the spread–skill ratio in brackets (1 is ideal):

sensors, noiseDPSGOSPGOSP without chargePCFM
16, σ=0.10.021 (0.74)0.026 (1.15)0.049 (1.95)0.050 (1.94)
16, σ=0.30.045 (0.64)0.043 (1.26)0.191 (1.49)0.194 (1.48)
4, σ=0.10.039 (0.48)0.036 (0.93)0.175 (1.51)0.180 (1.50)
4, σ=0.30.080 (0.34)0.087 (0.67)0.270 (1.26)0.279 (1.24)

The results barely change. "Without charge" here means our GOSP without its update of the stored source sample. Its CRPS is then two to five times higher and equal to PCFM's.

Second, we retuned the guidance step of DPS and GOSP on references 20 to 23, whose trajectories appear nowhere in the evaluation. Seven of the eight choices did not change. The one that changed, GOSP with 16 sensors and σ=0.1, gives CRPS 0.029 instead of 0.026 at the new step, so DPS keeps the lower CRPS in that configuration. In every configuration the spread of DPS's samples stays smaller than their actual error.

Overlap in the held-out settings. Three of our 15 held-out settings reuse the trajectory on which we chose the guidance step and the projection schedules, with other random seeds. Below we report only the other 12: four new trajectories, each with a different initial vorticity and forcing phase, and three seeds each. All settings were fixed on the original trajectory and applied unchanged. This is the transfer test the reviewer asks for.

All variants on the 12 independent settings. We ran every variant the reviewer listed: GOSP without the charge, the Euler form (guidance and projection added to the flow's ordinary step) with its own tuned step size, and DPS followed by three projection steps at the end, as many as GOSP uses on kinetic energy. We add PCFM with 200 steps, which makes as many network calls as GOSP. Values are medians. "Residual" measures how far a sample is from meeting the constraints. "Found" counts samples in which every streamline crossing lies within half a grid cell of its target, and diverged samples count as failures.

velocity errorvelocity residualstreamline errorstreamline residualstreamlines in range / found
PCFM6.1e-41.4e-61.2e-40.3162 / 61
PCFM, 200 steps1.7e-41.4e-69.6e-50.001570 / 70
DPS1.4e-40.152.2e-21493 / 0
DPS + final projection1.2e-41.2e-67.0e-28695 / 48
Euler form9.3e-51.3e-61.1e-40.4899 / 96
GOSP without charge4.9e-41.4e-68.8e-50.2768 / 67
GOSP4.8e-51.3e-65.0e-56.9e-490 / 90

GOSP has the lowest error in all 12 settings on both tasks, against every other variant. A final projection makes DPS meet the velocity sensors but does not reduce its error. The Euler form keeps more samples in range and finds every crossing more often, in all 12 settings. About one GOSP sample in ten diverges, while every GOSP sample that stays in range finds all its crossings. We will state this trade-off in the paper.

Kinetic energy and conservation laws. We agree the evidence there is weaker, and we expect it to be. Conservation laws fix only a few global numbers, so there is little for guidance to add. On kinetic energy the projection starts only near the end of sampling, and before that GOSP takes the same steps as DPS. On the 12 settings, removing the charge changes nothing (error 3.2e-2 against 3.4e-2, each lower in 6 settings). What GOSP adds is that it meets the constraint: residual 6.2e-6, against 0.89 for DPS and 0.13 for DPS with a final projection, all at the same error.

How the schedules were chosen. When the projection starts on kinetic energy, and how strongly it acts over time on streamlines, came from PCFM runs on the original trajectory, before GOSP existed. No evaluation trajectory was used. For a new constraint we have no automatic rule. Our procedure was to project at every step and, if the first projections raised the residual, to start later.

Data variety. The 12 settings above change both the initial vorticity and the forcing phase. We will state that each system uses a single flow model.

Reviewer_sz1vOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on the use of pre-trained flow-matching models for solving physical inverse problems. The paper proposes a new sampling procedure that combines projection- and gradient-based corrections. In particular, the authors introduce a mechanism that updates the source sample to preserve the effect of gradient guidance across successive projection steps. They also propose an efficient implementation of the projection step in the observation space. They validate their method using multiple metrics and demonstrate improvements over existing baselines on Navier–Stokes and shallow-water system benchmarks.

Critical strengths

The paper is well written.

The proposed algorithm is simple and easy to use and seems effective based on the experimental results.

Critical weaknesses

The novelty is incremental, as it combines existing methods, and the gains could be modest in some cases.

I would suggest addition of some visual results on the reconstructions by different methods to the main paper.

The paper does not provide a probabilistic view on why both guidance by gradient and projection are required or what they jointly sample.

There are many related works on similar inference-time flow matching methods for inverse problems (e.g., for imaging) that are missing from the relevant work: PnP-Flow, Flower, FlowLPS, ....

Questions for authors
  1. The algorithm updated the input noise also during the inference. I am curious how sampling completely fresh noise at each step affects the performance.
1 comment

We thank the reviewer for the comments.

Fresh noise at every step. We ran this experiment. Drawing fresh noise at every step is what ECI does. Adding guidance to it gives GOSP with the stored source sample thrown away and replaced at every step. We tuned the guidance step the same way as for GOSP and drew 100 samples on the main reference.

in rangefoundMSEresidual
velocity sensors
ECI (fresh noise, no guidance)100–1.1e-31.4e-6
GOSP + fresh noise, ζ=30\zeta=30100–5.2e-51.3e-6
GOSP, ζ=3\zeta=3100–4.4e-51.3e-6
streamline crossings
ECI49492.9e-40.020
GOSP + fresh noise, ζ=1\zeta=164625.9e-50.071
GOSP, ζ=0.3\zeta=0.395954.7e-53.7e-5

Here, Found is the number of samples in which every streamline crossing lies within half a grid cell of its target.

Fresh noise throws away the part of the guidance that GOSP stores in the source sample. What remains is the part that reaches the projected prediction through the network, and the projection removes much of that too. To make up for the loss, the fresh-noise variant needs a guidance step ten times larger on velocity, at the top of the range we tried, before it comes close to GOSP. On streamlines it keeps only 64 of 100 samples in range, against 95 for GOSP, and its constraint residual is about a thousand times larger. Fresh noise does help on shallow water, where ECI has the lowest error. On the sparse sensor tasks, keeping the stored source sample is both more accurate and more stable.

Probabilistic view. We agree this is the main open question, and the paper does not claim that GOSP samples a particular distribution. What we can say is the following. When the projection damping is set to the noise variance, the projection is the standard data-assimilation update with an identity prior covariance. It corrects the field only where the observations look, with no physical spread to the rest of the field. The guidance step sends the misfit back through the network, so the model itself decides how the unobserved part of the field should change. The two steps act on different parts of the field, and each one alone fails in a predictable way:

  • Projection alone matches the data but not the unobserved field. On velocity its error is 1.1e-3, and with noisy sensors its ensembles are too wide (spread–skill ratio 1.2 to 1.9).
  • Guidance alone does not meet the data. Its velocity residual is 0.11, and with noisy sensors its ensembles are too narrow (spread–skill ratio 0.34 to 0.69).
  • Together they reach a spread–skill ratio of 0.71 to 1.20.

This is an empirical result, not a guarantee. Deriving the prior covariance from the model, so that the combination targets a true posterior, is the natural next step.

Novelty and modest gains. A simple combination like guiding and then projecting, does not work inside PCFM, because PCFM rebuilds each state from the stored source sample and so erases the guidance. Charging the guidance to the source sample is the smallest change that fixes this. At another reviewer's request, we added two controls on 15 held-out settings:

  • Without the charge, GOSP falls back close to PCFM. Its velocity error is 6.0e-4 against 4.5e-5 for GOSP, and GOSP is better in all 15 settings.
  • PCFM given the same number of network calls as GOSP, by running 200 steps, is still worse in all 15 settings: 2.0e-4 on velocity and 6.7e-5 on streamlines, against 4.5e-5 and 4.9e-5 for GOSP.

The gains are small on the conservation-law tasks and on kinetic energy, and our analysis predicts this. Conservation laws fix only a few global quantities, so there is little room for guidance to add information. On kinetic energy the projection only starts near the end of sampling, and before that GOSP takes exactly the same steps as DPS. There, GOSP's benefit is that it satisfies the constraint: its residual is 2.0e-6, against 0.40 for DPS, at the same error.

Visual results. We will add a figure to the main text with reconstructions by PCFM, DPS, the Euler-form combination and GOSP, and with error maps at a time level that is not observed, for velocity sensors and streamline crossings.

Related work. We will add PnP-Flow, Flower, FlowLPS and related flow-matching methods for imaging inverse problems. PnP-Flow alternates a gradient step on the data fit, a return to the flow path with fresh noise, and a denoising step with the flow. Methods of this kind mostly handle linear measurements and draw fresh noise at every step, which is the regime tested above. Our setting differs in two ways. The constraints are nonlinear and coupled, and in one task the measured points move with the state. And the sampler rebuilds each state from a stored source sample, which is what carries the guidance.

Reviewer_t2QiOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors introduce guided observation-space projection (GOSP) for inference-time constrained generation on pre-trained flow-based models. Their approach combines gradient guidance, intermediate projections, and source prior updates. They also provided a matrix-free kernel-Gram solver that solves projections without forming the constraint Jacobian. The authors evaluated their method on several PDE tasks like Navier-Stokes and shallow water, with ablations in source-charging, sensor layouts, and heldout references.

Critical strengths
  1. Their ablations show source update has improvements on sparse observation tasks. In Table 9 of the paper, the velocity ensemble mean MSE increases and the number of streamline samples decreases when the source update is removed. This shows that the source update is useful for preserving the gradient guidance with PCFM.

  2. The kernel-Gram projection is accurate and scales well. It is identical to autodiff up to rounding errors and produces the same samples to numerical precision. On 128^2 fields, the kernel-Gram projection takes 1.04s per sample compared to 159s for matrix-free CG with the same constraint residual while for dense assembly, it can run out of memory.

Critical weaknesses
  1. My main concern is the methodological novelty of GOSP and contribution beyond combining PCFM with gradient guidance introduced by DPS. GOSP actually adopts the gradient guidance from DPS and retains PCFM's Gauss-Newton projections and re-interpolation. Their main change is updating the prior source sample to preserve the guided displacement. The advantage over other ways of combining guidance and projection is less clear. For example, for velocity observations, DPS followed by final projection already obtains an MSE of 9.9e-5 with constraint residual of 3.2e-6. Additionally, for streamline crossings, the Euler-form combination achieves 94/100 samples to be close to the reference compared to 95/100 for GOSP. The main question is whether further final projections, with more steps or smaller tolerance, could close this gap on the residual. These comparisons make it difficult to determine the contributions of GOSP beyond combinations of existing methods in DPS and PCFM.

  2. The benefit and generality of source charging are not fully established. GOSP matches PCFM on the shallow water and vorticity integral tasks, while removing the source charging on kinetic energy gives the same MSE (Table 9). On the task of noisy sensors, DPS has lower ensemble mean error while GOSP achieves better CRPS in only 2 (Table 8). Additionally, their improvements in the new metric of streamline-crossing "close samples" mainly track numerical stability, which makes it hard to isolate reconstruction gains against PCFM or DPS. The ablation without source update is also missing from the noisy sensor and held-out experiment tasks, so it will be good to show these ablations to provide evidence of consistent benefits.

Questions for authors
  1. Comparison with alternative combinations of DPS and PCFM. My understanding of GOSP's main new sampling change is "charging" the gradient-guidance displacement to the stored source sample. For the streamline-crossing task, could the authors report this ablation of DPS followed by final PCFM projection, and the Euler-form combination followed by final projection, on this task? Comparing reconstruction accuracy, final constraint residual, and computational costs could help clarify if source updating provides an important advantage over simpler combinations of existing approaches with more final Gauss-Newton projection steps.

  2. Contribution of the source charging across held-out and noisy settings. Can the authors report GOSP results without source charging on the 15 held-out settings (Table 11) and the free initial state noisy observation experiments (Table 8) using the same or comparable guidance and projection settings?

  3. Computational trade-off and more complex problems. From my understanding, Table 4 demonstrates substantial acceleration of their projection, while Table 6 in fact shows that GOSP takes longer time than PCFM when both use the same kernel-Gram solver. Is this due to additional gradient computation through the model? Could the authors clarify the end-to-end generation quality vs runtime trade-off, particularly at higher grid resolutions? The current experiments in GOSP focus on 2D PDE problems, and an evaluation on more complex and 3D problems would help establish if GOSP can extend to more complex and higher dimensional physical systems.

1 comment

We thank the reviewer. We ran every requested comparison: we use same checkpoint, prior seeds and references; ζ\zeta taken from the paper. Results are medians over 15 settings (5 references × 3 seeds).

First, we would like to point out that GOSP addresses a computational bottleneck in previous methods. For example, the original PCFM implementation explicitly constructs the Jacobian, which limits its scalability.

Q1. DPS + projection and the Euler form. The "velocity form" row of Table 9 is the Euler-form combination, and it already ends with a final projection. We will rename it.

velocity, main ref.MSEresidual
DPS9.9e-50.11
DPS + final proj.9.9e-53.2e-6
Euler form9.8e-51.3e-6
GOSP4.4e-51.3e-6

A final projection removes DPS's residual but not its error, and the Euler form inherits DPS's accuracy. As mentioned, only the charge keeps the part (1−τ)M(∂uvθ)⊤A⊤r(1-\tau)M(\partial_u v_\theta)^\top A^\top r of the guidance alive across later projections.

streamlinesin rangefoundMSEresidual
DPS + final proj. (main)96582.2e-296
Euler form (main)100921.2e-312.8
GOSP (main)95954.7e-53.7e-5
Euler form (held-out)99951.4e-40.64
GOSP (held-out)90904.9e-51.0e-4

On held-out settings the Euler form finds every crossing in more samples (13/15 settings), because about 10% of GOSP samples leave the range. GOSP's advantage is accuracy: lower MSE in 15/15 settings and lower residual in 14/15. On the samples both methods keep in range, GOSP's MSE is still lower in 15/15 settings (4.9e-5 vs 1.2e-4).

Q2. The charge on held-out settings and noisy sensors.

held-outvel. MSEstream. MSEfound
PCFM6.8e-41.5e-460
GOSP, no charge6.0e-41.0e-465
GOSP4.5e-54.9e-590

GOSP beats its no-charge ablation in 15/15 settings on both sets. Without the charge, results fall back to PCFM's.

noisy (CRPS, SSR)DPSGOSPno chargePCFM, ε=σ2\varepsilon=\sigma^2
PP=16, σ\sigma=0.10.021 (0.69)0.026 (1.12)0.051 (1.91)0.051 (1.90)
PP=16, σ\sigma=0.30.045 (0.62)0.044 (1.20)0.194 (1.49)0.199 (1.47)
PP=4, σ\sigma=0.10.038 (0.48)0.035 (0.97)0.180 (1.56)0.187 (1.56)
PP=4, σ\sigma=0.30.076 (0.34)0.078 (0.71)0.276 (1.26)0.282 (1.24)

Without the charge, CRPS rises 2–5× and equals PCFM's in all four configurations. We agree GOSP does not beat DPS on CRPS. GOSP stays within 25% of DPS's CRPS with a better-calibrated spread (SSR 0.71–1.20 vs 0.34–0.69).

Q3. Cost. GOSP costs one extra backward pass per step (Table 6: 56 s vs 27 s for velocity). The kernel Gram speeds up the projection for any sampler. The charge improves samples at the price of that extra pass. We will separate the two contributions in the Introduction. To test whether the extra compute explains the gain, we gave PCFM the same network passes:

held-out, 200 passesvel. MSEstream. MSEfound
PCFM, 200 steps2.0e-46.7e-570
GOSP, 100 steps4.5e-54.9e-590

GOSP is better in 15/15 settings on both sets. Table 4 covers resolution scaling (128², up to 4096 sensors). We have no 3D flow model and make no 3D claim. The kernel Gram uses only FFTs and point sampling, and its limit is the number of sensors per level.

Novelty. Earlier guidance + projection samplers advance the state with the flow's own update; inside PCFM this gives the Euler-form row, which has DPS's accuracy with exact constraints. PCFM re-interpolates from the stored prior draw, which erases guidance, and the no-charge runs confirm this on every held-out and noisy setting.

Comments on the paper

1 comment
AuthorsOpenReview ↗

general response

We thank all three reviewers. here is some points we want to highlight:

1. GOSP removes a computational bottleneck of earlier methods. First, we would like to point out that GOSP addresses a computational bottleneck in previous methods. For example, the original PCFM implementation explicitly constructs the Jacobian, which limits its scalability. Its memory grows with the number of constraints times the number of unknowns, and building it dominates the sampling time. On PCFM's own Navier–Stokes benchmark, a full run of 100 samples takes 1963 s with the released implementation, 90 s with our matrix-free projection and 33 s with our kernel-Gram projection. On shallow water the times are 4080 s, 108 s and 18 s. At 128×128 resolution with 4096 velocity sensors, about 41,000 constraints on 800,000 unknowns, building the Jacobian runs out of memory on an 80 GB GPU. The kernel-Gram projection takes 1.04 s per sample there. Matrix-free conjugate gradients needs 159 s to reach the same accuracy. This part of the work does not depend on guidance: it makes every projection-based sampler, PCFM included, usable at sizes where it was not before.

2. The source update is what makes guidance survive projection. PCFM rebuilds each state from a stored source sample, and this erases a guidance step taken before the projection. GOSP updates the stored source sample together with the state. New experiments, on 12 held-out settings with new initial conditions and forcing that were never used for tuning:

  • Without the source update, GOSP falls back close to PCFM. Velocity error is 4.9e-4 against 4.8e-5, and GOSP is better in all 12 settings. With noisy sensors, removing it raises CRPS two to five times, to the level of PCFM.
  • PCFM given the same number of network calls as GOSP is still worse in all 12 settings.
  • DPS followed by a final projection meets the sensors, but its error does not drop (1.2e-4 on velocity, 7.0e-2 on streamlines).
  • The Euler-form combination, with its own tuned step size, has twice GOSP's velocity error. GOSP's error is lower in all 12 settings on both velocity and streamlines.
  • Drawing fresh noise at every step, as Reviewer sz1v asked, needs a ten times larger guidance step on velocity and loses a third of the samples on streamlines.

3. What the new experiments show against us. We report three points plainly.

  • On streamline crossings the Euler-form combination keeps more samples in range and finds every crossing in more of them, in all 12 settings. About one GOSP sample in ten diverges. GOSP's advantage is lower error and a smaller constraint residual, not a higher success count.
  • With noisy sensors, GOSP and DPS have similar CRPS, and DPS is lower in two of four configurations. What GOSP adds is a spread that matches its error, where DPS's spread is too small.
  • Reviewer SivT found that our noisy-sensor tuning shared trajectories with four evaluation references, and we found that three of our 15 held-out settings reused the tuning trajectory. Removing the overlap, and retuning on separate trajectories, leaves the conclusions unchanged. Seven of eight tuning choices stayed the same, and the one that changed made GOSP slightly worse.

On kinetic energy and the conservation-law tasks the gain is small, as our analysis predicts: those constraints leave little of the field for guidance to improve. There, GOSP's benefit is that it meets the constraint, with a residual of 6.2e-6 against 0.89 for DPS.

4. Planned revisions.

  • Separate the two contributions in the introduction, the faster projection and the source update, and state the cost of the extra backward pass.
  • One main-text table with GOSP, GOSP without the source update, DPS with a final projection, the Euler-form combination and PCFM with matched compute, on the 12 independent settings.
  • State the Euler-form trade-off on streamlines and the noisy-sensor result as above.
  • Fix the noisy-sensor split so that tuning and evaluation share no trajectory, and list the evaluation indices.
  • Add a figure of reconstructions and error maps, the fresh-noise experiment, and the missing related work (PnP-Flow, Flower, FlowLPS).
  • State plainly that GOSP does not sample a known distribution, and that choosing projection schedules for a new constraint has no automatic rule.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.