acceptodds
Under review as a conference paper at ICLR 2027

Learning Diffusion-Bridge Flow Maps with Continuous-time Consistency

13 views

Abstract

Diffusion Bridge Models (DBMs) provide a diffusion-based framework for image-to-image (I2I) translation. However, high-quality sampling often requires tens of costly neural-network evaluations. Recent consistency-based approaches accelerate sampling by learning direct mappings from intermediate bridge states to the data endpoint , but do not learn the entire bridge flow. In this work, we develop a single consistency model that represents the entire finite-time bridge flow, mapping any intermediate state directly to any earlier state . Learning the full bridge flow poses two main challenges: how to represent the flow, and how to train it stably in continuous time. We address these issues by using the exact solution of the bridge PF-ODE to structure the learned map, and by deriving a bridge-specific sign correction that prevents the continuous-time update from amplifying flow-map inconsistency. On held-out ImageNet-512 super-resolution and deblurring, BFM reduces FID-VAE over reproduced CDBM by and , respectively, at the same 4-NFE budget; at 8 NFEs, it reaches FID-VAE comparable to DBMSolver using 20-30 NFEs.

Primary area:
generative models
Keywords:
Diffusion Bridge Models, Image Translation, Consistency Models, Flow Maps

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_4AkkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper aims to speed up sampling in DBMs for image-to-image tasks. Existing consistency-based bridge methods, such as CDBM, are only learning the jump from an intermediate state straight to the clean image. Instead, this works proposes to train one model that can jump from any time t to any earlier time s along the bridge's probability-flow ODE. The experiments focus on 4x super-resolution and deblurring on ImageNet-512. The paper reports reports generally good results across this tasks.

Critical strengths
  • The problem is well motivated
  • Building the network around the exact bridge solution makes sense and the authors show that it trains much better than a plain average-velocity parameterization.
  • The ablations are useful, covering the key components, that is parameterization, sign correction, full-flow vs. endpoint-only supervision, and flow composition. The finding that full-flow supervision helps even when sampling only ever targets s = 0 is interesting.
  • The paper is clearly written and easy to follow
Critical weaknesses
  • While i like the idea itself, the novelty is slightly overstated, and the closest prior works are missing. For example, the paper says prior bridge-consistency models learn only the endpoint map (Fig. 2, introduction). That is not quite true as "Schrödinger Bridge Consistency Trajectory Models for Speech Enhancement"already learns a CTM-style map for arbitrary t -> s jumps along a Schrödinger-bridge ODE. In the same line, "Generalized Consistency Trajectory Models for Image Manipulation" learns arbitrary-time maps on a flow-matching ODE between two distributions. With paired data, this effectively becomes a deterministic bridge between degraded and clean images. Other bridge-consistency work are also missing, such as "Single-Step Bidirectional Unpaired Image Translation Using Implicit Bridge Consistency Distillation". The parameterization is also close to existing designs like "Trajectory Consistency Distillation: Improved Latent Consistency Distillation by Semi-Linear Consistency Function with Trajectory Mapping".

  • The headline speedup is overstated. In Table 1, DBMSolver gets worse with more steps. Saying that BFM at 8 NFE matches DBMSolver at 20–30 NFE therefore compares against DBMSolver's weaker settings. It also looks to me that most of the improvement seem to come from the BFM pipeline versus the reproduced CDBM, not from learning the full flow.

  • It looks that all results are the authors' own reproduction. This means that none of these numbers can be compared with published work. Ideally the main comparison should focus on settings that match prior work conditions.

  • Following on the above, some of the reproduction appear weak. The CDBM flat across NFE (SR 3.00 / 2.96 / 3.01), whereas published CDBM improves with steps.

  • Unused standard benchmarks. The standard ImageNet-256 restoration suite from has published numbers for amny of these methods (I2SB, DDBM, DBIM, DBMSolver, etc). The paper even uses the I2SB 10k validation subset, but not these tasks.

  • Some baselines are missing, like one step image super-res methods "OFTSR: One-Step Flow for Image Super-Resolution with Tunable Fidelity-Realism Trade-offs", "Consistency Trajectory Matching for One-Step Generative Super-Resolution".

Questions for authors

Please see primarily the weaknesses section.

  • Can the authors please report results on the standard I2SB ImageNet-256 tasks, against the published numbers of these methods?
  • How exactly were CDBM and IBMD reproduced (training budget, learning rate, schedule tuning)? Why is the reproduced CDBM flat across NFE?
2 comments
AuthorsOpenReview ↗

Author response to 4Akk: Part 1/2

We appreciate the reviewer's recognition that BFM is well motivated, that its bridge-structured parameterization is effective, and that the full-flow supervision ablation is interesting. We clarify the novelty and evidence below and report two new benchmarks (ImageNet-256 SR4×, ImageNet-512 JPEG-QF10).

>> Novelty and relation to prior flow models (W1).

We agree that Schrödinger Bridge Consistency Trajectory Models (SBCTMs) already learn arbitrary t→st \rightarrow s Schrödinger-bridge transitions, and have corrected this statement in the introduction and Fig. 2. However, learning an arbitrary-time map is not, by itself, our claimed contribution. BFM develops a continuous-time consistency training formulation for the full conditional DDBM Bridge PF-ODE flow, using its analytic finite-time structure. We derive an analytic finite-time representation that separates the contributions of the current bridge state xt\boldsymbol{x}_t and conditioning endpoint xT\boldsymbol{x}_T from the learned interval-denoising quantity D‾θ\boldsymbol{\overline{D}}_\theta (Eq. 6):

fθ(xt,t,s,xT)=m(t,s)xt+n(t,s)xT+o(t,s)D‾θ(xt,t,s,xT).f_\theta(\boldsymbol{x}_{t}, t, s, \boldsymbol{x}_{T}) = m(t, s) \boldsymbol{x}_{t} + n(t, s) \boldsymbol{x}_{T} + o(t, s) \boldsymbol{\overline{D}}_\theta(\boldsymbol{x}_{t}, t, s, \boldsymbol{x}_{T}).

SBCTM learns Schrödinger-bridge trajectories through teacher-trajectory supervision in discrete-time, whereas BFM learns the conditional Bridge PF-ODE flow directly through a continuous-time objective. GCTM learns arbitrary-time maps on a flow-matching ODE; with paired data this is indeed a deterministic bridge, as the reviewer notes. The difference is that GCTM uses a generic flow-map parameterization with discrete-time training, whereas BFM reuses a pretrained DDBM, builds its exact finite-time coefficients into the map, and trains in continuous time.

TCD develops a semilinear interval representation for ordinary (noise-to-image) diffusion but does not consider conditional diffusion bridges, while IBCD focuses on single-step unpaired translation.

To the best of our knowledge, none of these works combines a continuous-time full-flow consistency objective with an analytic finite-time parameterization of the conditional DDBM Bridge PF-ODE.

We now cite SBCTM and TCD (Sec. 2.4) and IBCD (Sec. 2.3), and state the distinction in Sec. 2.4, and compare all of these methods in App. K.

>> Practical importance of the bridge-specific parameterization (W1).

We further investigate a generic average-velocity (AV) adaptation: our converted-AV baseline analytically converts the pretrained DBM denoiser while preserving its instantaneous Bridge PF-ODE velocity at initialization. We use the same backbone, training budget, and corresponding inference schedules and NFE budgets, with the appropriate parameterization-specific training factors. Despite this improved initialization, at 8 NFEs, converted AV achieves FID-VAE 8.45, compared with 1.99 for BFM. The corresponding FIDs are 8.60 and 2.15, respectively.

These results provide direct experimental evidence that BFM's bridge-structured formulation offers substantial practical advantages over the two tested generic flow-map adaptations.

>> The importance of full-flow learning is directly measured (W2).

The reviewer questions whether BFM's gains arise from full-flow learning or other differences in the training pipeline. Our controlled destination-supervision ablation in the submitted appendix directly isolates this effect: the full-flow and endpoint-directed variants share the same architecture, DBM initialization, training budget, and inference procedure, differing only in destination-time supervision.

Importantly, both models use s=0s=0 endpoint-directed flow-map calls during inference. Nevertheless, full-flow training improves deblurring FID-VAE from 5.84 to 4.06 at 2 NFE, 2.97 to 2.19 at 4 NFE, and 2.16 to 1.99 at 8 NFE. This directly establishes the practical benefit of full-flow supervision independently of the shared training pipeline.

AuthorsOpenReview ↗

Author response to 4Akk: Part 2/2

>> Published results and controlled reproductions (W3).

To clarify, not all comparisons are reproductions: Tab. 3 (E2H/DIODE; Tab. 2 in the original submission) uses the published CDBM, IBMD, DBIM and DBMSolver numbers, and 2-NFE BFM improves on the published 2-NFE results (e.g., on DIODE, BFM-2 achieves FID 1.69 versus 2.93 for published CBT-2 and 3.12 for IBMD-2).

For the ImageNet-512 comparison, we adapt the published CDBM and IBMD training procedures to the same task-specific DBM teachers and held-out evaluation set used by BFM, ensuring controlled comparisons rather than combining incompatible protocols. Our additional “standard” FID measurements (i.e., FID computed against original RGB images) confirm BFM's improvements over CDBM at 4 NFEs: 3.03 to 2.03 on SR4× and 2.97 to 2.49 on deblurring, and at 8 NFEs on deblurring (2.99 to 2.15).

Tab. 1 now reports standard FID alongside FID-VAE.

>> CDBM/IBMD reproduction and NFE behavior (W4/Q2).

Our CDBM reproduction follows the published training recipe with batch size 1024 and learning rate 2.83×10−52.83\times10^{-5}; we select the best-performing evaluated CDBM checkpoint of 20k iterations for both tasks. We observed that the later 25k or 30k checkpoints showed worse FID-VAE, FID, and paired reconstruction metrics (e.g., at 25k, deblurring FID-VAE rises from 2.90 to 4.03 at 2 NFEs). For IBMD, we follow App. B of their paper: 3k student updates, learning rate 3×10−63\times10^{-6}, and five auxiliary updates per student update.

A nearly flat CDBM NFE curve is not itself evidence of an incorrect reproduction: as analyzed in AYF, endpoint-consistency errors can accumulate during repeated denoising and re-noising, so increasing NFEs need not improve quality monotonically. The CDBM, IBMD and BFM training and sampling settings are in App. G.3–G.4.

>> ImageNet-256 benchmarks and experimental protocol (W5/Q1) .

In response to the reviewer's ImageNet-256 request, we additionally evaluate BFM and CDBM on SR4× using both the I²SB 10k validation subset and the full 50k ImageNet validation set, with the same conditional DDBM teacher. I²SB's published ImageNet-256 results are not directly comparable, as they differ in bridge formulation, architecture, and evaluation protocol.

BFM consistently improves FID and LPIPS across both evaluation sets. On the full 50k validation set, BFM reduces standard FID from 9.92/9.75/9.58 to 7.08/6.80/6.71 at 2/4/8 NFEs, respectively. Notably, BFM-2 outperforms CDBM-8 in FID (7.08 versus 9.58).

ImageNet-256 SR4× standard FID↓ (I²SB 10k subset / full 50k validation; same DBM teacher):

NFECDBM (10k)BFM (10k)CDBM (50k)BFM (50k)
212.8610.009.927.08
412.679.789.756.80
812.519.669.586.71

>> Additional ImageNet-512 JPEG restoration benchmark (additional evidence).

We further evaluate BFM on ImageNet-512 JPEG-QF10 restoration using I²SB's corruption operator. All methods share the same task-specific EDM2-XS DBM teacher and the same 10k held-out ImageNet validation images.

ImageNet-512 JPEG-QF10 restoration (same 10k held-out images; sampling in App. G.4):

MethodNFEFID-VAE ↓FID ↓LPIPS ↓PSNR ↑SSIM ↑
DBMSolver102.392.770.245825.110.7765
DBMSolver202.302.510.252924.710.7578
CDBM24.154.310.276524.360.7449
IBMD-224.224.450.270824.630.7641
BFM (Ours)23.674.050.266924.910.7646
CDBM44.324.460.282524.170.7358
BFM (Ours)42.342.540.255524.770.7576
CDBM84.334.470.283124.160.7349
BFM (Ours)82.322.500.256924.660.7539

>> Additional benchmarks (W6).

We now discuss OFTSR and CTMSR in Sec. 2.4 and App. K. Both target one-step super-resolution using different model formulations and evaluation protocols, whereas our controlled comparisons evaluate acceleration of the same conditional DDBM teacher. Their published scores therefore do not provide directly matched comparisons.

>> Solver comparisons (W2).

We agree that DBMSolver-10 is already competitive and have corrected the overly broad 20–30-NFE comparison. On the submitted ImageNet-512 tasks, BFM-8 achieves FID-VAE of 1.82/1.99 on SR4×/deblurring, versus DBMSolver-10's 1.71/2.00. On the new JPEG-QF10 task, BFM-4 achieves FID-VAE 2.34 against 2.39 for DBMSolver-10, while BFM-8 achieves FID 2.50, comparable to DBMSolver-20's 2.51. Moreover, at matched 4/8 NFEs, BFM is ahead of DBMSolver on all three tasks (4-NFE FID-VAE: 1.79/2.19/2.34 vs. 3.14/5.06/5.62; App. I.5).

We have corrected the solver comparison in the abstract, Sec. 4.2, contributions, and conclusion.

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

This paper proposes Bridge Flow Models (BFM) to accelerate few-step sampling for diffusion bridges. Unlike existing consistency-based bridge models that learn only an endpoint map , BFM learns a full conditional bridge flow map. Using the exact finite-time solution of the Bridge PF-ODE, the authors decompose the flow map into known analytic terms and a learned interval-averaged denoising quantity, and train it with a continuous-time consistency objective with a sign correction. Experiments on ImageNet-512 super-resolution/deblurring, Edges2Handbags, and DIODE show that BFM approaches 20–30-NFE teacher DBMs with only 2–8 NFEs.

Critical strengths
  1. The problem formulation is clear and practically relevant. Extending diffusion-bridge acceleration from endpoint mapping to full conditional flow mapping is a natural and valuable direction.
  2. The bridge-structured parameterization is theoretically grounded. By leveraging the exact finite-time solution of the Bridge PF-ODE, it separates known linear evolution from the learned network term, which is better aligned with the bridge structure than generic average-velocity parameterization. 3 The main ImageNet results are strong. At 8 NFEs, BFM achieves FID-VAE comparable to DBMSolver with 20–30 NFEs, and at 2 NFEs it improves over reproduced CDBM/IBMD baselines on super-resolution
Critical weaknesses
  1. Novelty relative to general flow-map works needs clearer positioning. CTM, GCTM, AYF, and MeanFlow already learn flow maps between arbitrary times. The uniqueness of this paper lies in the bridge-specific analytic structure and sign correction, but the paper does not sufficiently discuss why the bridge setting cannot be treated simply as a special case of CTM/GCTM.
  2. Qualitative comparison is clearly insufficient and selective. The quantitative tables compare Hybrid Heun, DBIM-3, DBMSolver, CDBM, and IBMD, but the qualitative figures mainly compare CDBM 2 NFE and Teacher 30 NFE, without same-case qualitative results for IBMD or DBIM-3. Figure 3 selects cases where BFM has an advantage; meanwhile, in Table 1 deblurring at 2 NFEs, CDBM achieves FID-VAE 2.90, better than BFM full-flow’s 4.06. The paper acknowledges this in the text, but the abstract and Figure 1 can easily overstate BFM’s advantage at 2-NFE deblurring. Randomized sampling, multi-baseline qualitative comparisons, and failure cases should be added.
  3. The failure of average-velocity parameterization is extremely severe and may need implementation/tuning controls. The gap between FID-VAE 250.4 and 4.08 is very large. The paper claims both use corresponding output scaling and training configurations, but it should clarify whether sufficient hyperparameter search was performed for the average-velocity baseline; otherwise the comparison may be questioned as unfair.
  4. User study, failure cases and limitations are underexplored. Examples include extreme degradation, noisy conditioning endpoints, and user study.
Questions for authors
  1. Compared with CTM/GCTM/AYF, what is the core irreplaceable contribution of BFM? If one directly applies CTM to the Bridge PF-ODE while ignoring the bridge analytic structure, what key capabilities are lost?
  2. Why do the qualitative figures not include IBMD 2-NFE and DBIM-3 few-step results? Can the authors add randomized multi-baseline qualitative comparisons?
  3. Could the severe failure of average-velocity parameterization be caused by mismatched output scaling, time sampling, or learning rate?
  4. Are there failure cases? User study?
3 comments
AuthorsOpenReview ↗

Author response to aNqX: Part 1/3

We thank the reviewer for recognizing the relevance of our work, the practical importance of learning the full conditional bridge flow, the theoretical motivation of our bridge-structured parameterization, and the strong ImageNet results. We address the concerns regarding novelty, the average-velocity comparison, qualitative evaluation, and limitations below.

>> Novelty and relation to CTM/GCTM/AYF/MeanFlow (W1/Q1).

While CTM, GCTM, AYF, and MeanFlow establish important foundations for learning arbitrary-time flow maps, adapting their formulations to conditional diffusion bridges presents additional challenges that motivate our approach.

The exact finite-time solution of the conditional Bridge PF-ODE provides analytic, time-dependent coefficients for the current bridge state (xt\boldsymbol{x}_t) and conditioning endpoint (xT\boldsymbol{x}_T). However, the generic flow-map parameterizations of CTM/GCTM, AYF, or MeanFlow do not explicitly separate these known contributions, requiring the network to learn the bridge transport alongside the interval-dependent denoising quantity. We observed that this is particularly relevant for few-step sampling, where the model must predict xt→xs\boldsymbol{x}_t \rightarrow \boldsymbol{x}_s transitions across long time intervals (as seen in App. F).

Our proposed bridge-structured finite-time parameterization (Eq. 6) explicitly accounts for the known bridge transport, allowing the network to focus on the remaining interval-denoising quantity rather than the complete finite-time displacement. While a sufficiently expressive generic flow map could represent the same transitions, it would not automatically incorporate this conditional bridge structure.

Furthermore, we develop a continuous-time consistency objective for learning the full conditional Bridge PF-ODE flow. Unlike CTM/GCTM's finite-step trajectory supervision, our objective directly evaluates the continuous-time material derivative through a JVP, avoiding finite-step trajectory targets and their associated discretization errors, as motivated by sCM.

Importantly, applying this objective to our bridge-structured finite-time parameterization reveals a bridge-specific training-update sign reversal when ∂tG(t,s)<0\partial_t G(t,s) < 0, for which we derive a corresponding correction. Its practical importance is demonstrated by our submitted controlled ablation: removing the correction worsens 2-NFE deblurring FID-VAE from 4.64 to 165.42 (App. E.1).

The practical importance of full-flow learning is directly supported by our submitted controlled ablation. Under the same architecture, DBM initialization, training budget, and inference procedure, full-flow supervision improves 4-NFE deblurring FID-VAE from 2.97 to 2.19, even though both models use only endpoint-directed flow-map evaluations during sampling (i.e., s=0s=0 during sampling).

We further evaluate a velocity-preserving AYF-style adaptation (named “converted AV”) to the conditional Bridge PF-ODE, with complete comparisons and implementation details provided in our response to W3/Q3.

Together, these results demonstrate the practical importance of our bridge-structured finite-time parameterization and full-flow continuous-time objective for learning conditional bridge flows.

We clarified these differences in Secs. 1 and 2.4 and summarize them in App. K.

>> Qualitative comparisons and sampling-budget interpretation (W2/Q2).

To directly address this concern, we added same-image comparisons on 16 fixed validation images per ImageNet-512 task, selected independently of model performance (App. L.1). We display 14 images, excluding two whose source images contain a watermark or advertising text. The panels include the requested IBMD-2 (2 NFEs) and DBIM-3 (10 NFEs), alongside CDBM and BFM at matched 2 and 4 NFEs.

At 2-NFE deblurring, CDBM achieves better FID-VAE (2.90 versus BFM's 4.06) and standard FID (3.05 versus 4.47). However, BFM achieves better paired reconstruction metrics at this budget: LPIPS 0.2078 versus 0.2144, PSNR 26.73 versus 25.81, and SSIM 0.834 versus 0.804. Thus, the aggregate results exhibit a distributional-quality versus reconstruction-fidelity trade-off, rather than uniform superiority of either method at 2 NFEs.

Importantly, the abstract's reported improvements over CDBM explicitly concern 4-NFE evaluation, not 2 NFEs. At 4 and 8 NFEs, BFM also achieves substantially better deblurring FID-VAE (2.19 versus 2.85 and 1.99 versus 2.87, respectively), corroborated by standard FID.

We also revised the figures the reviewer points to: Fig. 1's caption no longer states a 15× reduction and labels its examples as selected, and Fig. 3's caption now points to the randomized comparisons in App. L.1.

AuthorsOpenReview ↗

Author response to aNqX: Part 2/3

>> Average-velocity comparison and fairness (W3/Q3).

We clarify that our submitted “direct” AV baseline in App. F follows the average-velocity flow-map parameterization used by AYF and shares BFM's backbone, training data, (t,st, s) time sampling, optimizer, learning rate, batch size, and training budget. Each method uses mathematically justified, parameterization-specific output scalings and weightings; their derivations, preconditioning and continuous-time training objectives are provided in App. F, and the training and sampling protocols in Apps. G.3–G.4.

Thus, the direct AV comparison is conducted under controlled experimental conditions.

To further examine whether initialization accounts for the large performance gap, we introduce “converted-AV”, which analytically converts the pretrained DBM denoiser into the instantaneous conditional Bridge PF-ODE velocity. This preserves the pretrained DBM's instantaneous velocity at initialization while retaining AYF's average-velocity flow-map form (compared in the table below).

Flow-map parameterizations (direct AV / converted AV / BFM):

FeatureDirect AVConverted AVBFM (Proposed)
Flow Mapxt+(s−t)Vθx_t+(s-t)V_\thetaxt+(s−t)[Atxt+BtxT+CtDθ]x_t+(s-t)[A_tx_t+B_tx_T+C_tD_\theta]mtsxt+ntsxT+otsD‾θm_{ts}x_t+n_{ts}x_T+o_{ts}\overline D_\theta
Analytic StructureGeneric average-velocity formInstantaneous Bridge PF-ODE velocityExact finite-time bridge coefficients
Learned QuantityInterval-average velocityss-conditioned denoiser within the converted velocityInterval-denoising quantity
Endpoint PredictionLearned through the velocity outputRequires compensation for the remaining xtx_t and xTx_T termsDirectly recovers D‾θ\overline D_\theta at s=0s=0

We further evaluate all three formulations on the same 10k-image ImageNet-512 deblurring evaluation set, reported in the table below.

ImageNet-512 deblurring: AV vs. BFM (same 10k images):

MethodNFEFID-VAE ↓FID ↓
Direct AV (in appendix)2250.44249.96
Converted AV2389.40389.30
BFM24.064.47
Direct AV4283.48283.42
Converted AV4257.13256.89
BFM42.192.49
Direct AV8308.60309.21
Converted AV88.458.60
BFM81.992.15

These results show that incorporating the instantaneous bridge velocity in converted-AV substantially improves over direct AV at 8 NFEs, reducing FID-VAE from 308.60 to 8.45. Nevertheless, BFM achieves considerably better quality, with FID-VAE 1.99 and FID 2.15. The reported converted-AV configuration is the best of those we trained: variants with the sign correction enabled, at learning rates of 10−610^{-6} and 10−510^{-5}, reach 291.73 and 293.37 at 8 NFEs, against 8.45 (App. F.3).

The trend across NFEs is consistent with the long-interval sensitivity analyzed in App. F. At 2-4 NFEs, converted AV relies on relatively few interior endpoint predictions, whereas at 8 NFEs, additional predictions at smaller source times reduce the amplification of a given velocity-prediction error.

The direct AV control and velocity-preserving converted-AV baseline, together with documented parameterization-specific scaling, training, and sampling protocols, provide strong empirical evidence for the practical importance of BFM's bridge-specific finite-time parameterization. Even when AV is initialized to preserve the pretrained DBM's instantaneous Bridge PF-ODE velocity, BFM achieves substantially better few-step sampling quality.

AuthorsOpenReview ↗

Author response to aNqX: Part 3/3

>> Failure cases, robustness, and evaluation scope (W4/Q4).

We thank the reviewer for suggesting additional failure-case analysis.

We added selected failure cases (App. L.2) from images with the largest LPIPS degradation of BFM relative to CDBM at 2 NFEs, with the displayed examples chosen to illustrate distinct failure modes. In these cases, BFM over-smooths fine high-frequency structure that CDBM preserves (e.g., cloth weave, thin spider legs, foliage). The same-image panels (App. L.1) additionally show failures shared by all methods, which use the same teacher DBM and VAE: small text, fine semantic details such as eyes, and small object structure.

Regarding the suggested user study, our claims concern few-step generation quality and reconstruction fidelity, evaluated quantitatively (through NFEs, FID, CA, PSNR, SSIM, and LPIPS) and supplemented by qualitative comparisons.

A human preference study would provide complementary perceptual evidence, but our paper does not make claims about human perceptual preference.

Extreme degradations and noisy conditioning endpoints would test robustness to changes in the conditioning distribution, beyond our evaluated task protocols of SR4×, Gaussian Deblurring, and (new) JPEG Restoration (QF=10). We do not claim general robustness to unseen conditioning perturbations and state this in the Conclusion (Sec. 5).

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

This paper proposes Bridge Flow Model (BFM), which leverages the exact finite-time solution of the Bridge Probability Flow ODE (PF-ODE) to construct mappings between arbitrary source and destination times. It further employs continuous-time consistency training and a bridge-specific sign correction to improve training stability. The proposed method achieves promising few-step sampling performance on super-resolution, deblurring, and image-to-image translation tasks.

Critical strengths
  1. The proposed bridge-structured parameterization separates the known analytic components of bridge dynamics from the learned interval-denoising quantity, naturally satisfying the identity and endpoint-prediction boundary conditions. The sign correction designed to address the sign reversal introduced by the bridge parameterization is also supported by theoretical analysis and empirical results.
  2. Solid experimental comparisons and ablations. The ImageNet comparisons use the same task-specific teacher DBMs and evaluation images. In Table 1, at the same 4-NFE budget, BFM reduces FID-VAE from 2.96 to 1.79 for super-resolution and from 2.85 to 2.19 for deblurring compared with CDBM. The appendix further provides targeted evaluations of full-flow supervision, sign correction, intermediate-destination accuracy, and flow-composition consistency.
Critical weaknesses
  1. A gap between the theoretical derivation and the practical training objective requires clarification. Appendix D.1 establishes the unbiasedness of the conditional bridge velocity and its JVP-based estimator. However, Eq. (10) further applies nonlinear RMS normalization to each stochastic tangent. An unbiased tangent estimator does not necessarily imply that the expected normalized update preserves the same consistency fixed point. Furthermore, the sign-correction analysis in Appendix E holds the remaining tangent terms fixed and is therefore a restricted local analysis. Its implications for the stability of full neural-network training require further justification.

  2. The parameterization comparison is insufficient to fully establish the advantages over existing flow-map methods. In Table 4, the average-velocity parameterization achieves an FID-VAE of 250.44, indicating nearly complete failure. While this demonstrates the effectiveness of the bridge-structured parameterization under the current configuration, it does not rule out the effects of time scaling, preconditioning, and initialization. In particular, given that the paper uses T = 80, the error amplification argument based solely on the interval length |t − s| is insufficient to establish an inherent conditioning disadvantage of the average-velocity parameterization.

  3. The acceleration benefits should be assessed against more competitive reference configurations. The paper emphasizes that BFM with 8 NFEs achieves quality comparable to DBMSolver with 20–30 NFEs. However, Table 1 shows that DBMSolver already achieves FID-VAE scores of 1.71 and 2.00 at 10 NFEs for super-resolution and deblurring, respectively, compared with 1.82 and 1.99 for BFM at 8 NFEs. Therefore, achieving comparable quality does not necessarily require 20–30 NFEs for DBMSolver. A more comprehensive quality-versus-wall-clock-time comparison, together with training cost analysis, would better substantiate the efficiency claims. Moreover, BFM underperforms CDBM at 2 NFEs on deblurring (4.06 vs. 2.90), suggesting that its few-step sampling advantages are somewhat conditional.

  4. The evaluation of reconstruction fidelity and cross-task generalization remains insufficient. The main ImageNet results primarily report FID-VAE and classifier accuracy, while PSNR, SSIM, and LPIPS are only provided in internal ablations, without matched comparisons against the main baselines. Furthermore, the main E2H and DIODE evaluations are conducted on training splits and involve differences in backbone architectures. Although the authors provide additional comparisons with matched-backbone teachers and held-out test results, these experiments are still insufficient to establish controlled generalization advantages over CDBM and IBMD.

Questions for authors
  1. Appendix D.1 establishes the unbiasedness of the conditional bridge velocity and JVP estimator, while Eq. (10) further applies nonlinear RMS normalization to the stochastic tangent. Could the authors clarify whether this operation introduces bias into the gradient estimation and how it affects the convergence properties of the continuous-time consistency objective?

  2. Appendix E.1 reports that only approximately 2–3% of sampled source-destination pairs satisfy G'(t, s) < 0. However, removing the sign correction dramatically degrades FID-VAE from 4.637 to 165.415. Could the authors further explain why such a small fraction of time pairs leads to severe overall training degradation?

  3. The fairness of the average-velocity parameterization comparison: In Table 4, the average-velocity parameterization achieves an FID-VAE of 250.438, indicating nearly complete failure. Could the authors provide a stronger average-velocity baseline with appropriate time normalization, preconditioning, or training hyperparameter adjustments to verify that the performance gap primarily stems from the proposed bridge-structured parameterization rather than an unsuitable baseline training configuration?

2 comments
AuthorsOpenReview ↗

Author response to RdBr: Part 1/2

We thank the reviewer for recognizing the bridge-structured parameterization, the theoretical motivation and empirical effectiveness of sign correction, and the strong controlled ImageNet results. We address the concerns regarding training-objective analysis, average-velocity comparisons, efficiency, and generalization below.

>> RMS normalization's unbiasedness in the continuous-time training rule (W1/Q1).

App. D.1 establishes unbiasedness of the raw conditional Bridge PF-ODE velocity and JVP-based tangent estimator, not of the subsequently RMS-normalized update. We agree that nonlinear, sample-wise tangent normalization can change the expected update and does not generally preserve its unbiasedness. However, this does not invalidate the underlying continuous-time consistency condition or the unbiasedness of its raw tangent estimator.

Importantly, tangent normalization is an established stabilization technique adopted from sCM and also employed by AYF, rather than a novel theoretical component of BFM. Its purpose is to bound the per-sample tangent target: $ \left|\left| \frac{\boldsymbol {g}}{\|\boldsymbol {g}\|_{\mathrm{RMS}}+c} \right|\right|_\text{RMS} < 1$, when c>0c > 0.

We have clarified that tangent normalization does not change the ideal continuous-time consistency condition: a zero bridge-flow tangent remains zero after normalization. However, the nonlinear normalization can change the expected stochastic training update, so we do not claim that the normalized training rule preserves all stationary points of the original rule (Sec. 3.4; Apps. D.1 and E).

Finally, App. E analyzes the local sign-correction mechanism, rather than claiming global convergence of full neural-network training. Its practical effectiveness is independently supported by the controlled sign-correction ablation.

>> Importance of sign correction despite a small negative-sign region (W1/Q2).

The reported 22-33\\% represents the frequency of sampled time pairs with ∂tG(t,s)<0\partial_t G(t,s) < 0, not their overall influence on training. Importantly, these pairs reverse the local training-update direction, rather than merely introducing additional estimation noise. As derived in App. E, without our proposed sign correction, the update locally increases the tangent magnitude when ∂tG(t,s)<0\partial_t G(t,s) < 0, whereas with our correction, the contracting direction is restored. Because training updates affect shared network parameters, their influence need not be proportional to their sampling frequency. These are also the pairs few-step sampling depends on most. At s=0s=0, ∂tG(t,0)\partial_t G(t,0) is negative for all source times above t≈10.7t\approx10.7 on deblurring (t≈9.9t\approx9.9 on SR4×, t≈10.5t\approx10.5 on JPEG-QF10). This region is rarely sampled in training, but every sampler starts in it: in the 2-NFE deblurring sampler, both network evaluations (t=80t=80 and t=16t=16) lie there.

Our submitted controlled ablation directly demonstrates the practical importance of this correction: removing it increases 2-NFE FID-VAE from 4.637 to 165.415.

>> Reconstruction fidelity and generalization (W4).

We added standard FID for the compared methods in Tab. 1, with complete matched LPIPS, PSNR, and SSIM results in App. I.3.

At 4 NFEs, BFM improves LPIPS from 0.208 to 0.176 and PSNR from 25.70 to 26.88 on ImageNet-512 SR4×, and LPIPS from 0.217 to 0.193 and PSNR from 25.50 to 26.76 on deblurring, relative to CDBM.

For E2H/DIODE, we follow the established evaluation protocols in DDBM, DBIM, DBMSolver, CDBM, and IBMD, as disclosed in our submission. Our principal controlled evidence on held-out images comes from the ImageNet-512/256 experiments, where the competing methods share task-specific teacher DBMs; separate held-out E2H/DIODE evaluations against the DBM teachers are reported in submitted App. I.2.

We further strengthen our evaluation with ImageNet-512 JPEG-QF10 restoration and ImageNet-256 SR4×. On the latter, BFM achieves lower FID and LPIPS than CDBM at 2/4/8 NFEs on the 10k validation subset, while CDBM retains higher PSNR and SSIM.

AuthorsOpenReview ↗

Author response to RdBr: Part 2/2

>> Average-velocity parameterization and baseline fairness (W2/Q3).

We clarify that our argument does not rely solely on the ∣t−s∣|t-s| factor to establish the practical importance of BFM's parameterization. App. F derives the error sensitivity of the tested average-velocity (AV) representation over long intervals, while explicitly recognizing that both AV and BFM can represent the target flow in principle. The central distinction is that BFM analytically incorporates the known finite-time conditional bridge transport, allowing the network to learn the remaining interval-denoising quantity rather than the complete displacement. App. F.1 now states explicitly that this sensitivity characterizes the tested AV form and does not establish a disadvantage under alternative time scalings or preconditioning.

Importantly, our submitted direct-AV comparison already uses the same backbone, training data, (t,s)(t,s) sampling, optimizer, learning rate, batch size, and training budget as BFM, with mathematically justified parameterization-specific output scaling and weighting.

To address the reviewer's concern about initialization and preconditioning, we additionally introduce a stronger “converted-AV'' baseline, which analytically converts the pretrained DBM denoiser into its instantaneous conditional Bridge PF-ODE velocity while retaining the AV flow-map formulation and pretrained denoiser preconditioning. At 8 NFEs, converted AV substantially improves over direct AV (FID-VAE 308.60 →\rightarrow 8.45), yet BFM achieves 1.99. The corresponding standard FID is 8.60 for converted AV versus 2.15 for BFM. The derivation, preconditioning and training details are in App. F.2, and the complete 2/4/8-NFE comparison in App. F.3. Converted AV's network-output gain is (t−s)⋅αt⋅cout(t)/t≤cout(t)(t-s)\cdot \alpha_t \cdot c_\text{out}(t)/t \leq c_\text{out}(t), the same bound as BFM's gain and independent of T=80T=80, yet the gap to BFM remains.

These comparisons provide strong empirical evidence for the practical importance of BFM's bridge-specific finite-time parameterization beyond the ∣t−s∣|t-s| error-scaling argument. Even when the AV formulation is initialized to preserve the pretrained DBM's instantaneous Bridge PF-ODE velocity, BFM achieves substantially better few-step sampling quality.

>> Acceleration benefits and competitive numerical solvers (W3).

We agree that DBMSolver-10 is already competitive, rather than requiring 20-30 NFEs to achieve comparable quality. On SR4×/deblurring, BFM-8 achieves FID-VAE 1.82/1.99, versus 1.71/2.00 for DBMSolver-10 (Tab. 1). At matched budgets, BFM is ahead: DBMSolver obtains FID-VAE 3.14/5.06/5.62 at 4 NFEs and 1.99/2.38/3.08 at 8 NFEs on SR4×/deblurring/JPEG-QF10, versus 1.79/2.19/2.34 and 1.82/1.99/2.32 for BFM (App. I.5).

Regarding 2-NFE deblurring, CDBM's better FID-VAE (2.90 versus 4.06) was already explicitly reported in our submitted Tab. 1. However, BFM achieves better paired reconstruction fidelity at this budget (LPIPS 0.2078 versus 0.2144, PSNR 26.73 versus 25.81, and SSIM 0.834 versus 0.804). On the JPEG-QF10 task, BFM also improves 2-NFE FID-VAE from 4.15 to 3.67 over CDBM (table below).

ImageNet-512 JPEG-QF10 restoration (same 10k held-out images; full results in App. I.4):

MethodNFEFID-VAE ↓FID ↓LPIPS ↓PSNR ↑SSIM ↑
DBMSolver102.392.770.245825.110.7765
DBMSolver202.302.510.252924.710.7578
CDBM24.154.310.276524.360.7449
IBMD-224.224.450.270824.630.7641
BFM (Ours)23.674.050.266924.910.7646
CDBM44.324.460.282524.170.7358
BFM (Ours)42.342.540.255524.770.7576
CDBM84.334.470.283124.160.7349
BFM (Ours)82.322.500.256924.660.7539

We have corrected the original speedup framing in the abstract, Sec. 4.2, the contributions, and the conclusion. Under matched single-H100 measurements, BFM-8 achieves sampling-only throughput of 97.90/85.59/94.67 images/s on SR4×/deblurring/JPEG-QF10, compared with 86.46/72.01/86.87 for DBMSolver-10. App. G.5 reports sampling-plus-VAE throughput. BFM is also cheaper to train: on JPEG-QF10 its reported checkpoint (5k iterations) takes 5.50 H100 GPU-hours, versus 11.97 for the reproduced CDBM to reach its best (20k) checkpoint of a 30k-iteration run.

Reviewer_TZEnOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors consider the Diffusion Bridge Models for the image-to-image translation, then state the problem of simulation require to sample from such model and propose a method for learning an accelerated mapping function (consistency model) which allows for few step sampling. Authors utilize continuous in time consistency models methodology and in contrast to the existing method for the acceleration of the Diffusion Bridge Models authors build a mapping from one arbitrary point of time to other arbitrary point in time, while other consistency and inverse distillation methods build only endpoint mapping. In addition, authors propose efficient mapping parametrization thorough separation of analytic and neural parts and adapt their algorithm for numerically stable training and inference.

Critical strengths
  • The paper is mostly well written and easy to follow.
  • The practical results on ImageNet image inverse problems suggest the practical utility of the continuous in time consistency loss and training of t→st \rightarrow s loss of the method with very competitive results and quality vs speed tradeoff.
  • The final method is rather clean, logical and easy to implement on practice.
  • The authors openly disclose the evaluation on training set in image to image experimental setup, see Section 4.3
Critical weaknesses
  • The methodology is mostly incremental. The main mechanism for consistency based acceleration of diffusion bridges were already introduced in CDBM [1] and the continuous in time loss with slightly different parametrizations was introduced in consistency models literature [2, 3]. While the t→st \rightarrow s mapping wasn't trained for the diffusion bridges in particular, it is well known method for regular diffusion models and at the end the inference procedure almost fully follows CDBM.
  • The possibility of the methodology extension to the unconditional diffusion bridges is not discussed, while for example IBMD [4] explicitly allows for both (IBDM-I2SB and IBMD-DDBM).
    • Furthermore the experimental comparison on ImageNet problems with similar computational budget I2SB and IBMD-I2SB would be nice.
  • Manuscript and clarity:
    • The Section 3 is hard to read:
      • Authors introduce a lot of new notation that is short lived, which makes the reading difficult.
      • The jump from Section 3.3 continuous time consistency property to the loss function Eq 10 is too rapid. The notation is changing between those two subsections, while the basic loss for continuous time consistency property is not described (just stated in Eq 10 as a matter of fact). This makes the introduction of G sign and RMS normalization to the loss Eq 10 rather sharp and hard to understand. I encourage authors to make the loss derivation more explicit with connection to continuous time consistency property in Section 3.3. In addition, I do acknowledge that Appendix D provides explicit and full loss derivation, but it is not references from Section 3.4 and such derivation at least partly should take place in the main text.
    • Sections 4.4 and 4.5 are function more as appendix sections abstracts since they provide very little information about the actual experiment and very brief conclusions. I suggest authors to expand them in the final version or to make them even shorter. In my personal opinion, right now it is something in the middle which seems a bit off. Minor concern.
  • Experimental part is not ideal:
    • I could not find the particular diffusion bridges coefficients, such as base drift and variance coefficients. There are many versions of diffusion bridges with different coefficients, that perform differently, so this is very important technical detail, that must be explicitly stated.
    • ImageNet Image Inverse problems:
      • The Image Inverse problem is solved that has naturally paired ground truth image and corrupted observation. Hence the paired metrics are natural for such task, e.g., MSE, LPIPS, however only distributional level metrics (FID-VAE, CA) are reported in the general comparison Table 1. I encourage authors to provide paired metrics evaluation for Section 4.2 experiment.
      • The FID-VAE is reported. Why in particular FID-VAE, why not just regular FID, as other methods report? I encourage authors to provide FID evaluations.
      • The types of corruption operator for the Section 4.2 experiments are rather limited, while other papers, e.g., [4], provide also other corruption operators such as different kernels for SR, JPEG restoration and inpainting. I encourage authors to provide results for other corruption operators for ImageNet experiment, especially considering the Section 4.2 experiment as the main evaluation protocol. Alternatively authors could utilize some other image restoration setup.
    • General Image to Image tasks
      • The evaluation on the train set makes comparison not strong enough. Although I appreciate the authors explicit disclosure of this fact. The metrics, especially paired MSE and LPIPS, introduce very low error of the method, which may suggest overfitting.
      • That fact shifts the main experimental evidence load to the ImageNet experiment and connects with my concerns regarding Section 4.2 experiment.

References:

[1] He et al. Consistency Diffusion Bridge Models. NeurIPS, 2024.

[2] Song et al. Consistency Models. ICML, 2023.

[3] Lu and Song. Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models. ICLR, 2025.

[4] Gushchin et al. Inverse Bridge Matching Distillation. ICML, 2025.

[5] Liu et al. I²SB: Image-to-Image Schrödinger Bridge. ICML, 2023.

Questions for authors

I encourage authors to address my concerns at least partially and I will adjust the score correspondingly.

5 comments
AuthorsOpenReview ↗

Author response to TZEn: Part 1/3

We thank the reviewer for their careful assessment and constructive feedback, and for recognizing the clarity of our presentation, the clean and practical method formulation, and the competitive quality–speed trade-offs on ImageNet. Following the reviewer's requests, the revision now specifies the bridge coefficients (App. C), reports paired metrics and standard FID (Tab. 1, App. I.3), adds a JPEG restoration benchmark (Tab. 2, App. I.4), makes the loss derivation explicit (Secs. 3.3-3.4), and formulates an extension to I²SB-style unconditional bridges (App. J). We address each concern below.

>> Novelty beyond CDBM/sCM and continuous-time consistency (W1).

We respectfully disagree that BFM is merely an incremental combination of CDBM and existing continuous-time consistency methods.

CDBM learns endpoint consistency, predicting x0\boldsymbol{x}_0 from xt\boldsymbol{x}_t; its multistep sampler constructs intermediate states through repeated endpoint prediction and re-noising. This does not constitute directly learning the Bridge PF-ODE flow between arbitrary intermediate times. In contrast, BFM learns the full conditional Bridge PF-ODE flow through continuous-time consistency supervision, directly mapping xt\boldsymbol{x}_t to xs\boldsymbol{x}_s for arbitrary s<ts \lt t. Crucially, our bridge-structured finite-time parameterization analytically separates the known transport contributions of xt\boldsymbol{x}_t and conditioning endpoint xT\boldsymbol{x}_T from the learned interval-denoising quantity D‾θ\overline{\boldsymbol{D}}_\theta. To the best of our knowledge, BFM is the first to develop continuous-time full-flow consistency supervision with an analytic finite-time flow-map parameterization specifically for conditional diffusion bridges.

Importantly, our contribution is not the continuous-time consistency objective alone. While sCM develops JVP-based continuous-time consistency training for endpoint mappings in standard diffusion, applying this objective to CDBM without changing its endpoint-prediction parameterization would still learn endpoint consistency, rather than the full bridge flow map. Thus, simply combining CDBM's endpoint formulation with sCM's continuous-time training does not yield BFM's arbitrary-destination supervision or bridge-structured finite-time representation.

Moreover, our bridge-structured formulation and analysis reveal a local training-update sign reversal when ∂tG(t,s)<0\partial_t G(t,s)<0, for which we derive a correction (App. E). To the best of our knowledge, such a sign-reversal mechanism and its correction have not been identified in prior consistency or diffusion-bridge works.

The practical importance of this correction is demonstrated by our controlled ablation: without the correction, 2-NFE FID-VAE degrades from 4.64 to 165.42 (App. E.1). Likewise, a standard t→st \rightarrow s flow-map parameterization from regular diffusion (AYF-style average velocity; our converted-AV baseline) is far worse on the same bridge (8.45 vs. BFM's 1.99 at 8 NFEs; App. F.3). Thus, transferring known flow-map training to bridges is not straightforward: both the bridge-structured parameterization and the sign correction are needed.

Regarding inference, CDBM's endpoint-prediction-and-re-noising procedure is a special case (γ=1\gamma=1) of BFM's more general bridge sampling framework. BFM's learned arbitrary-time flow enables γ\gamma-sampling (Sec. 3.5), allowing interpolation between direct t→st \rightarrow s transitions (i.e., γ=0\gamma = 0) and CDBM-style endpoint prediction and re-noising (i.e., γ=1\gamma = 1), as specified in Sec. 3.5 and App. G.4.

Crucially, our submitted controlled ablation isolates full-flow supervision under the same architecture, initialization, training budget, and endpoint-directed inference (γ=1\gamma=1). Full-flow supervision improves 4-NFE deblurring FID-VAE from 2.97 to 2.19. Separately, it reduces the intermediate (20→2)(20 \rightarrow 2) prediction MSE from 0.00863 to 0.00221 (3.9$\times$), as reported in App. H.1-H.2. Thus, the shared sampling procedure does not account for the gains from learning the full conditional bridge flow.

We clarified these distinctions in Secs. 1 and 2.4.

AuthorsOpenReview ↗

Author response to TZEn: Part 2/3

>> Extension to unconditional diffusion bridges (W2).

BFM's main formulation exploits the analytic finite-time structure of the conditional DDBM Bridge PF-ODE, explicitly separating the known transport contributions of xt\boldsymbol{x}_t and xT\boldsymbol{x}_T from the learned interval-denoising quantity. Unlike IBMD, which distills conditional and unconditional bridge models into endpoint generators, BFM learns the full flow map between arbitrary times tt and ss.

Furthermore, we show that the continuous-time full-flow learning principle can also be formulated for I²SB-style unconditional flows. For Brownian-reference I²SB, we derive the marginal PF-ODE velocity, a JVP-based full-flow training objective, and a finite-time flow-map representation, described in App. J. This establishes a concrete theoretical extension pathway, while our present experimental contributions remain focused on conditional DDBMs.

>> Additional ImageNet I²SB and IBMD-I²SB comparisons (W2).

To isolate the effectiveness of an acceleration method, the underlying bridge formulation, teacher, and evaluation protocol should be held fixed. Our ImageNet-512 comparisons therefore use identical task-specific EDM2-based DBM teachers and held-out evaluation images.

Importantly, our submitted Tab. 1 already includes a controlled IBMD-DDBM comparison: at the same 2-NFE sampling budget, BFM improves FID-VAE from 5.26 to 2.72 on SR×4 and from 4.73 to 4.06 on deblurring.

Regarding the I²SB/IBMD-I²SB comparisons, their official ImageNet-256 checkpoints use the ADM-based UNets.

Our attempts to adapt the official ADM-based I²SB checkpoints to continuous-time bridge consistency training encountered substantial optimization instability despite incorporating sCM's proposed stabilization techniques. In contrast, our EDM2-based setup enabled stable training, consistent with sCM's adoption of EDM2-based architectures for large-scale experiments.

Direct comparisons with the published I²SB results would conflate differences in bridge formulations, architectures, and evaluation protocols, rather than isolating the acceleration method.

Training budgets are reported in App. G.3.

Additional ImageNet-256 evaluation. We have additionally evaluated BFM and CDBM on ImageNet-256 SR4× using an EDM2-based DBM teacher, as reported in the table below. Both methods are evaluated on the I²SB 10k ImageNet validation subset.

ImageNet-256 SR4× (10k held-out images; same DBM teacher):

MethodNFEFID ↓LPIPS ↓PSNR ↑SSIM ↑
CDBM212.860.1891524.7020.67151
BFM (Ours)210.000.1849524.2850.66921
CDBM412.670.1881624.7240.67280
BFM (Ours)49.780.1707824.0640.65379
CDBM812.510.1867724.7360.67301
BFM (Ours)89.660.1684023.6070.63558

BFM consistently achieves lower FID and LPIPS than CDBM at 2/4/8 NFEs, while CDBM retains higher PSNR and SSIM.

AuthorsOpenReview ↗

Author response to TZEn: Part 3/3

>> Manuscript clarity and presentation (W3).

Notation and Derivation (Sec. 3).

Sec. 3.3 now derives the raw pseudo-objective from fθ=H+GFθf_\theta = H + G F_\theta (full derivation in App. D.3), and Sec. 3.4 presents the sign correction and RMS normalization as modifications of it, noting that the normalization need not preserve the raw estimator's unbiasedness.

Experimental discussion (Secs. 4.4-4.5).

We have strengthened Secs. 4.4–4.5 with key quantitative results and their implications, while keeping the complete experimental configurations, comparisons, and analyses in the appendix.

>> Experimental evaluation and reproducibility (W4).

We strengthen the experimental specifications and evaluations as follows.

Bridge coefficients.

We have specified the bridge schedule and coefficients in App. C. For all tasks, we use a normalized VE schedule with αt=(1+t2)−1/2\alpha_t=(1+t^2)^{-1/2}, σt=t(1+t2)−1/2\sigma_t=t(1+t^2)^{-1/2}, f(t)=−t/(1+t2)f(t)=-t/(1+t^2), g(t)2=2t/(1+t2)g(t)^2=2t/(1+t^2), and T=80T=80; the bridge coefficients at,bt,cta_t, b_t, c_t are given there in closed form.

Paired metrics, standard FID, and rationale for FID-VAE. We originally reported FID-VAE as it uses VAE-reconstructed clean images to account for the VAE reconstruction bottleneck; standard FID instead uses the original clean RGB images.

We added standard FID alongside FID-VAE in Tab. 1, with complete matched LPIPS, PSNR, and SSIM comparisons in App. I.3. At 4 NFEs, BFM improves over CDBM on both SR×4 (RGB FID 3.03 →\rightarrow 2.03, LPIPS 0.208 →\rightarrow 0.176, PSNR 25.70 →\rightarrow 26.88) and deblurring (RGB FID 2.97 →\rightarrow 2.49, LPIPS 0.217 →\rightarrow 0.193, PSNR 25.50 →\rightarrow 26.76). At 2-NFE deblurring, although CDBM achieves better FID-VAE and standard FID, BFM improves paired LPIPS, PSNR, and SSIM metrics, revealing a distributional-quality versus reconstruction-fidelity tradeoff.

Additional ImageNet-512 JPEG Restoration Results. We have additionally evaluated ImageNet-512 JPEG-QF10 restoration using the corruption operator from I²SB, providing a third corruption family beyond SR4× and deblurring (table below). Full results are in App. I.4 (summary in Tab. 2).

ImageNet-512 JPEG-QF10 restoration (same 10k held-out images; full results in App. I.4):

MethodNFEFID-VAE ↓FID ↓LPIPS ↓PSNR ↑SSIM ↑
DBMSolver102.392.770.245825.110.7765
DBMSolver202.302.510.252924.710.7578
CDBM24.154.310.276524.360.7449
IBMD-224.224.450.270824.630.7641
BFM (Ours)23.674.050.266924.910.7646
CDBM44.324.460.282524.170.7358
BFM (Ours)42.342.540.255524.770.7576
CDBM84.334.470.283124.160.7349
BFM (Ours)82.322.500.256924.660.7539

E2H/DIODE evaluation protocol and generalization. Importantly, our principal evaluations use the 10,000 ImageNet-512 validation images, where BFM and prior works are compared with common teacher DBMs.

For E2H/DIODE, we emphasize that we follow the established evaluation protocol used by DDBM, DBIM, DBMSolver, CDBM, and IBMD, preserving comparability with their published benchmark results. This protocol was explicitly disclosed in our submission.

Furthermore, our submitted App. I.2 already reports held-out E2H/DIODE evaluations against the corresponding DBM teachers. On E2H (200 test pairs), 2-NFE BFM achieves LPIPS 0.2006 versus 0.2025 and MSE 0.0645 versus 0.0680 for the 30-NFE teacher. On DIODE (446 test pairs), BFM achieves LPIPS 0.4464 versus 0.4410 and MSE 0.2429 versus 0.2378. Thus, BFM maintains closely comparable reconstruction quality to the DBM teacher on unseen inputs, requiring only 2 rather than 30 NFEs. Sec. 4.1 points to these held-out results (App. I.2).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author response to all reviewers: Summary of revisions

We thank the four reviewers for their careful reading and constructive comments. We are encouraged that the reviewers found learning the full conditional bridge flow "a natural and valuable direction" (aNqX), the bridge-structured parameterization "theoretically grounded" (aNqX), the method "clean, logical and easy to implement" with "very competitive results and quality vs speed tradeoff" (TZEn), and the experiments "solid experimental comparisons and ablations" (RdBr) on a "well motivated" problem (4Akk).

We have uploaded a revised PDF with the changes in blue.

  • As TZEn, aNqX, and 4Akk asked how BFM differs from CDBM and general flow-map methods, we show that the direct combination falls short on bridges in two ways that BFM fixes: a velocity-preserving average-velocity baseline reaches FID-VAE 8.45 versus 1.99 for BFM at 8 NFEs (Appendix F.2-F.3), and removing the sign correction degrades 2-NFE FID-VAE from 4.64 to 165.42 (Appendix E.1). We corrected our statement about prior bridge-consistency models (Section 1, Figure 2), and discuss SBCTM, GCTM, TCD, IBCD, OFTSR and CTMSR (Sections 2.3-2.4, Appendix K).
  • As RdBr and aNqX asked whether the average-velocity baseline was fairly configured, this velocity-preserving baseline keeps the teacher's preconditioning and is the best of the configurations we trained (Appendix F.2-F.3).
  • As 4Akk and RdBr noted that DBMSolver is already competitive at 10 NFEs, we removed the 20-30-NFE comparison throughout. At matched 4 and 8 NFEs, BFM is ahead of DBMSolver on all three ImageNet-512 tasks (Appendix I.5). We also added sampling throughput and training time (Appendix G.5).
  • As TZEn and 4Akk asked for more corruption types and ImageNet-256, we added ImageNet-512 JPEG-QF10 restoration and ImageNet-256 4× super-resolution on both the 10k subset and the full 50k validation set (Table 2, Appendix I.4).
  • As TZEn and RdBr asked for standard FID and paired metrics, we added standard FID scores for all methods (Table 1) and LPIPS, PSNR and SSIM scores for all compared methods (Appendix I.3).
  • As aNqX asked for randomized and multi-baseline qualitative results, we added same-image comparisons including IBMD and DBIM-3 (Appendix L.1) and failure cases (Appendix L.2), and relabeled Figures 1 and 3 as selected examples.
  • As RdBr asked about tangent normalization and the sign analysis, we clarified the scope of both (Section 3.4, Appendices D.1 and E).
  • As TZEn requested, we made the loss derivation explicit in the main text (Sections 3.3-3.4), expanded Sections 4.4-4.5, specified the bridge coefficients (Appendix C), and derived the extension to unconditional I²SB-type bridges (Appendix J). As 4Akk asked, we added the CDBM and IBMD reproduction details (Appendices G.3-G.4).

We answer each remaining point in the individual responses. We thank the reviewers again and would be grateful to hear whether these responses resolve their concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.