Pith. sign in

REVIEW 3 major objections 6 minor 19 references

Inverse Bridge Matching Distillation

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A tractable inverse objective distills diffusion bridge models into one-step generators, using only corrupted images.

desk verdict First distillation that handles unconditional DBMs and one-step generation; the theory is clean, but the algorithm only approximately solves the inner problem and the conditional numbers partly come from train sets. read the letter →

arxiv 2502.01362 v2 pith:PTZJ3O6N submitted 2025-02-03 cs.LG cs.CV

Nikita Gushchin , David Li , Daniil Selikhanovych , Evgeny Burnaev , Dmitry Baranchuk , Alexander Korotin This is my paper · ORCID
classification cs.LGcs.CV
keywords diffusionbridgemodelsone-stepdistillationinversematchingimage-to-imagetranslationgenerativemodelingmodelaccelerationscore
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes Inverse Bridge Matching Distillation (IBMD), a way to compress a trained diffusion bridge model—a teacher that transforms corrupted images into clean ones across many stochastic steps—into a generator that outputs the same result in one or a few steps. The key move is to treat distillation as an inverse problem: find a student output coupling whose own bridge-matching solution reproduces the teacher's drift. The authors prove an unconstrained reformulation (Theorems 3.2 and 3.4) that makes this objective differentiable and trainable by gradient descent. In experiments on super-resolution, JPEG restoration, sketch-to-image, normal-to-image, and inpainting, the method accelerates inference by 4x to 100x, often matching or beating the multi-step teacher while training only on corrupted images.

What carries the argument

The load-bearing object is the inverse bridge matching reformulation (Theorem 3.2 for unconditional, Theorem 3.4 for conditional bridge matching): the constrained problem of matching drifts becomes a difference of two expectations, each an MSE regression of a score-like target on samples generated by the student coupling. The inner term, $\min_\phi \mathbb{E}\lambda(t)\|v_\phi - \nabla\log q\|^2$, is the bridge-matching loss of the student's own coupling; its optimum is the conditional variance of $x_0$ given $x_t$, so subtracting it makes the student prefer couplings whose noise is 'explainable' and penalizes mode collapse. In practice, the inner minimization is carried out by an auxiliary bridge network $b_{x_0}^\phi$ updated with $L$ gradient steps per generator step, and the student $G_\theta$ is trained by backpropagating through both the teacher network $b^*_{x_0}$ and the auxiliary network. A multi-step variant alternates generator prediction with posterior sampling of the diffusion bridge, following the pattern of DMD training to avoid train/inference distribution mismatch.

What would settle it

Train IBMD on one fixed teacher with $L=1$, $L=5$, and $L$ large (e.g. 100) inner updates per generator step, measuring final FID and the norm of the auxiliary network's residual drift against the true student bridge drift; strong dependence of performance on $L$ would confirm the inner-minimization assumption is load-bearing, while near-constant performance would suggest it is not.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that distilling a diffusion bridge model is equivalent to solving an inverse bridge matching problem: find a coupling $p_\theta(x_0,x_T)$ generated by a student $x_0 = G_\theta(x_T,z)$ such that the bridge-matching drift of the student equals the teacher drift $v^*$. The paper proves that this constrained problem has the same solutions as the unconstrained objective $$\min_\$\theta$\Big[\mathbb{E}_{x_0\sim p_\$\theta$,\, x_t\sim q}\$\lambda$(t)\|v^*(x_t,t)-\nabla_{x_t}\log q(x_t|x_0)\|^2 - \min_\phi\mathbb{E}_{x_0\sim p_\$\theta$,\, x_t\sim q}\$\lambda$(t)\|v_\$\varphi$(x_t,t)-\nabla_{x_t}\log q(x_t|x_0)\|^2\Big],$$ with a conditional analogue when the teacher is conditioned on $x_T$. Using the $x_0$-prediction reparameterization, the practical loss compares the teacher's prediction network with the student-generated $x_0$, minus the best prediction an auxiliary bridge network can achieve on student samples; the negative term enforces diversity and prevents collapse. The paper argues that this path-measure KL objective is the right one for bridge distillation, in contrast to marginal Fisher-divergence objectives, because for unconditional bridges the forward drift itself depends on the coupling and is not fixed.

Load-bearing premise

The equivalence in Theorem 3.2 assumes the inner bridge-matching minimization is solved exactly for the student's current coupling; the algorithm approximates it with $L$ gradient updates per generator step, and if the auxiliary bridge network lags the moving generator, the student gradient is no longer the true KL gradient.

Editorial extensions

If this is right

  • Unconditional bridge models, which previously had no dedicated distillation method, can now be accelerated; the paper's I2SB experiments cover super-resolution, JPEG restoration, and inpainting.
  • Conditional bridge models can be distilled into one-step generators, a regime the paper says is inaccessible to consistency-based distillation because of the singularity at the starting point.
  • Distillation no longer requires clean target images: only samples from the corrupted distribution $p(x_T)$ are used, because the student coupling is generated internally.
  • If the objective reaches zero, the student's one-step output distribution reproduces the teacher's path measure, not merely its marginal, so downstream behavior such as multi-step rollouts would match the teacher as well.
  • Speedups of 4x to 100x over teacher NFE are reported, with FID sometimes below the teacher's, which the authors attribute to accumulation of teacher sampling error over many steps.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural stress test is to vary the inner-minimization budget $L$: if the method's success depends on the auxiliary bridge network tracking the moving generator distribution, performance should degrade as $L$ shrinks and stabilize only when the inner problem is well solved.
  • The same unconstrained objective could be applied to other bridge constructions, such as Schrödinger-bridge or flow-matching variants, as long as the bridging kernel $q(x_t|x_0,x_T)$ is Gaussian or otherwise reparameterizable.
  • The claim that KL path-measure distillation is needed specifically for unconditional bridges suggests that marginal-score (Fisher) distillation methods, successful for standard diffusion, may fail on unconditional bridge teachers; comparing IBMD against a Fisher-based baseline on the same I2SB teachers would isolate that effect.
  • Because training needs only corrupted samples, the method may transfer to restoration settings where paired clean/corrupted data are scarce but corrupted samples are plentiful, such as real camera noise or historical degradation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes Inverse Bridge Matching Distillation (IBMD), a distillation method for diffusion bridge models (DBMs). The authors formulate distillation of a pre-trained teacher DBM into a one- or few-step generator as an inverse bridge matching problem: find a generator-induced coupling whose bridge matching drift matches the teacher drift. They prove (Theorem 3.2 and Theorem 3.4) that the constrained inverse problem can be rewritten as an unconstrained objective consisting of a teacher reconstruction loss minus an inner bridge-matching loss, under the assumption that the inner minimization over the auxiliary bridge network is solved exactly. They then present an alternating algorithm (Algorithm 1) that interleaves L gradient steps on the auxiliary bridge network with a student generator update, and evaluate the method on I2SB teachers (super-resolution, JPEG restoration, inpainting) and DDBM teachers (edges-to-handbags, normals-to-image, inpainting). The empirical results show that one- and few-step distilled models achieve competitive or better FID than multi-step teacher sampling on several tasks.

Significance. If the theoretical claim is taken together with the practical algorithm, the contribution is significant: IBMD is the first distillation technique applicable to both conditional and unconditional DBMs, it can produce one-step generators, and it does not require target-domain clean data during distillation. The algebraic reformulation in Theorem 3.2 is elegant and its proof in Appendix A is correct under the stated idealization of an exactly solved inner minimization. The paper also ships code, which aids reproducibility. However, the practical algorithm only approximately solves the inner minimization, and the evaluation has weaknesses (train-set FID in several benchmarks, no error bars). The central derivation is sound, but the gap between the idealized objective and the deployed algorithm needs to be addressed for the paper's main claims to be fully supported.

major comments (3)
  1. [Section 3.4, Algorithm 1] The equivalence in Theorem 3.2 holds only if the inner minimization over the bridge network is solved exactly for each generator state. Algorithm 1 approximates this inner minimization with L=5 gradient updates per generator step, with the bridge network initialized from the teacher checkpoint and never verified to be at the optimum. Under these conditions, the gradient of the surrogate student loss is not equal to the gradient of the KL objective in Proposition 3.1 unless the bridge network is exactly at its minimizer, and no bound or empirical measurement of the tracking error is provided. The paper should either give a convergence analysis (e.g., a two-timescale argument) or add experiments that quantify the gap, such as varying L and measuring the effect on final FID, before claiming that the algorithm optimizes the stated KL objective.
  2. [Section 5.2, Table 5] The quantitative comparison for Edges-to-Handbags and DIODE-Outdoor is performed on training sets, as the paper acknowledges. Since the paper also reports that the teacher overfits these training sets, FID computed on the training set can systematically favor models that memorize training images, and the claim that IBMD outperforms the teacher and prior methods on these benchmarks is not substantiated by held-out evaluation. In addition, no error bars or multiple-seed results are reported for any of the experiments, so it is unclear whether the observed differences are statistically significant. The authors should provide test-set FID (or a principled alternative for the small test sets) and report variance across runs.
  3. [Section 3.2, Theorem 3.2] The proof of Theorem 3.2 relies on the fact that the inner minimizer over v' is the conditional expectation, which requires the minimization to be over all measurable functions. In practice, the bridge network is a restricted function class, so the identity min_phi E[lambda ||b_phi - x0||^2] = E[lambda ||E[x0|xt,xT] - x0||^2] does not hold exactly. The theorem and Proposition 3.3 should explicitly state the assumption that the inner function class is sufficiently rich, or the paper should characterize the bias introduced by function approximation. This is conceptually distinct from the optimization gap in Algorithm 1 but should be clarified to avoid overstating the equivalence.
minor comments (6)
  1. [Section 2.1] The word 'distibution' is misspelled in the phrase 'rather than mapping distribution to distribution'.
  2. [Section 3.5] The section title contains the typo 'Mulitistep' and should be 'Multistep'.
  3. [Throughout] Unresolved cross-references such as 'M3.1', 'M3.2', 'M2.3', 'M3.4', 'M3.5', and 'M5' appear instead of proper section or figure numbers and should be fixed.
  4. [Section 4, equation (11)] The displayed formula contains a repeated 'KL(p(xT)||bp(xT))' term with an awkward brace annotation, and the coefficient 1/(2g^2(t)) should be checked for consistency with the SDE convention used in equation (5).
  5. [Table 7] The column 'L/K ratio' should define L as the number of bridge updates per generator update; Algorithm 1 uses L as an inner loop count and K as the outer loop count, so the notation is ambiguous.
  6. [Appendix C] The caption for Figure 8 contains the typo 'inpaiting' and should read 'inpainting'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the tractable inverse Bridge Matching objective is an exact algebraic identity, and the derivation does not reduce to its inputs.

full rationale

The central derivation is self-contained rather than circular. The teacher is an external, fixed model, and the distillation objective directly minimizes the KL between the teacher's reverse-time diffusion and the Bridge Matching process induced by the generator; no target-domain data or fitted output is relabeled as a prediction. Theorem 3.2 is an exact Hilbert-space projection identity: for v(xt,t) = E_{x0|xt}[∇ log q(xt|x0)], the paper proves E||v − v*||² = E||v* − ∇ log q||² − min_{v'} E||v' − ∇ log q||², so the negative inner term is the exact minimum of the Bridge Matching regression, not a fitted parameter smuggled into the objective. The reparameterized versions in Proposition 3.3 and Theorem 3.4 are substitutions of the Gaussian score relation and preserve the identity. Load-bearing external facts, such as the Bridge Matching characterization and the Girsanov-based KL formula, are cited from independent prior work (Shi et al., 2023; Vargas et al., 2021) and are parameter-free mathematical statements rather than author-supplied assumptions. The only self-citation, Gushchin et al., 2024, appears in background about unpaired Schrödinger bridge setups and does not carry the derivation. Algorithm 1 approximates the inner minimization with a finite number of bridge-network updates, which creates a genuine optimization gap; however, this is a convergence and tracking-error concern, not circularity, because the paper's theoretical equivalence does not depend on that approximation and the deployed method is benchmarked against external baselines and frozen teacher checkpoints.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The method relies on standard bridge matching regularity, linear-Gaussian priors, and an exact inner minimization over the auxiliary bridge network. The free parameters are the weighting schedule and multi-step grid, neither of which is reported. No new physical entities are introduced.

free parameters (2)
  • Weighting function lambda(t) = not reported
    Algorithm 1 uses an arbitrary positive weight function; theory is invariant to lambda, but practical convergence depends on it and the paper does not state the schedule used in experiments.
  • Multi-step timestep schedule {t_1,...,t_N} = not reported
    Section 3.5 defines N timesteps for multi-step distillation but does not specify their values or spacing; used in inpainting and NFE=2 experiments.
assumptions (6)
  • domain assumption All Bridge Matching regularity assumptions from Shi et al. (2023, Appendix C) hold, including smoothness and integrability conditions that justify Doob-h, Girsanov, and KL disintegration.
    Invoked in Appendix A and Proposition 3.1 to express KL(BM(Pi_theta)||M*) as an expectation over the mixture of bridges.
  • domain assumption The teacher drift v* is exactly the Bridge Matching solution for some coupling p*(x0|x_T) p(x_T); any teacher approximation error is inherited by distillation.
    Section 3.1 defines M*=BM(Pi*); in practice v* is a trained network, not the true optimal drift.
  • domain assumption The prior process has linear drift f(t)x_t and Gaussian transition kernels, giving the bridge coefficients a_t, b_t, c_t in Section 2.3.
    Needed for the x0-prediction reparameterization in Proposition 3.3 and for reparameterized sampling xt ~ N(a_t x_T + b_t x0, c_t^2 I).
  • domain assumption The student coupling is parameterized as p_theta(x0,x_T)=p_theta(x0|x_T)p(x_T) with a differentiable generator x0=G_theta(x_T,z), z~N(0,I).
    Section 3.1; this is the structural choice that makes gradients available and defines the class of achievable couplings.
  • ad hoc to paper The inner bridge matching minimization over v_phi is attained exactly in the tractable objective.
    Theorem 3.2 and its proof use the argmin and conditional expectation identity; Algorithm 1 only approximates this with L gradient updates, so there is a gap between theory and practice.
  • domain assumption Conditional Bridge Matching factorizes into a family of unconditional problems indexed by x_T, so the same objective holds with conditioning.
    Section 2.2 and proof of Theorem 3.4; this is the standard augmented bridge matching view from De Bortoli et al.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Inverse Bridge Matching Distillation." pith.science (2026). https://pith.science/paper/PTZJ3O6N

@misc{pith2026250201362,
  author       = {Pith},
  title        = {Pith review of: Inverse Bridge Matching Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PTZJ3O6N}},
  note         = {Machine review of arXiv:2502.01362}
}
read the original abstract

Learning diffusion bridge models is easy; making them fast and practical is an art. Diffusion bridge models (DBMs) are a promising extension of diffusion models for applications in image-to-image translation. However, like many modern diffusion and flow models, DBMs suffer from the problem of slow inference. To address it, we propose a novel distillation technique based on the inverse bridge matching formulation and derive the tractable objective to solve it in practice. Unlike previously developed DBM distillation techniques, the proposed method can distill both conditional and unconditional types of DBMs, distill models in a one-step generator, and use only the corrupted images for training. We evaluate our approach for both conditional and unconditional types of bridge matching on a wide set of setups, including super-resolution, JPEG restoration, sketch-to-image, and other tasks, and show that our distillation technique allows us to accelerate the inference of DBMs from 4x to 100x and even provide better generation quality than used teacher model depending on particular setup. We provide the code at https://github.com/ngushchin/IBMD

Figures

Figures reproduced from arXiv: 2502.01362 by the authors.

Figure 1
Figure 1. Outputs of DBMs models distilled by our Inverse Bridge Matching Distillation (IBMD) approach on various image-to￾image translation tasks and datasets (M5). Teachers use NFE≥ 500 steps, while IBMD distilled models use NFE≤ 4. Zhou et al., 2024a; Yue et al., 2024; Shi et al., 2023; De Bor￾toli et al., 2023). This approach allows DBMs to modify only the necessary components of the data, starting from an input sample ra… view at source ↗
Figure 2
Figure 2. Overview of (Conditional) Bridge Matching with xb0 reparameterization. The process begins by sampling a pair (x0, xT ) from the data coupling p(x0, xT ). An intermediate sample xt is then drawn from the diffusion bridge q(xt|x0, xT ) at a random time t ∼ U[0, T]. The model xb0 is trained with an MSE loss to reconstruct x0 from xt. In the conditional setting (dashed red path), xb0 is also conditioned on xT as an addi… view at source ↗
Figure 3
Figure 3. Overview of our method Inverse Bridge Matching Distillation (IBMD). The goal is to distill a trained (Conditional) Bridge Matching model into a generator Gθ(z, xT ), which learns to produce samples using the corrupted data p(xT ). Generator Gθ(z, xT ) defines the coupling pθ(x0, xT ) = pθ(x0|xT )p(xT ) and we aim to learn the generator in such way that Bridge Matching with pθ(x0, xT ) produces the same (Conditional)… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Uncurated samples for IBMD-I2 SB distillation of 4x-super-resolution with bicubic kernel on ImageNet 256 × 256 images. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Uncurated samples for IBMD-I2 SB distillation of 4x-super-resolution with pool kernel on ImageNet 256 × 256 images. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Uncurated samples for IBMD-I2 SB distillation of Jpeg restoration with QF=5 on ImageNet 256 × 256 images. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Uncurated samples for IBMD-I2 SB distillation of Jpeg restoration with QF=10 on ImageNet 256 × 256 images. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Uncurated samples for IBMD-I2SB distillation trained for inpaiting with NFE= 4 and inferenced with different inference NFE on ImageNet 256 × 256 images. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: Uncurated samples for IBMD-DDBM distillation trained for inpaiting with NFE= 4 and inferenced with different inference NFE on ImageNet 256 × 256 images. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: Uncurated samples from IBMD-DDBM distillation trained on the DIODE-Outdoor dataset (256 × 256) with NFE= 2 and NFE= 1, inferred using the corresponding NFEs on the training set. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: Uncurated samples from IBMD-DDBM distillation trained on the DIODE-Outdoor dataset (256 × 256) with NFE= 2 and NFE= 1, inferred using the corresponding NFEs on the test set. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_11.png]
Figure 12
Figure 12. Figure 12: Uncurated samples from IBMD-DDBM distillation trained on the Edges → Handbags dataset (64 × 64) with NFE= 2 and NFE= 1, inferred using the corresponding NFEs on the training set. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]
Figure 13
Figure 13. Figure 13: Uncurated samples from IBMD-DDBM distillation trained on the Edges → Handbags dataset (64 × 64) with NFE= 2 and NFE= 1, inferred using the corresponding NFEs on the test set. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 12 canonical work pages

  1. [1]

    and Sharma, R

    Barratt, S. and Sharma, R. A note on the inception score. arXiv preprint arXiv:1801.01973,

    arXiv

  2. [7]

    net/forum?id=PqvMRDCJT9t

    URL https://openreview. net/forum?id=PqvMRDCJT9t. Liu, G.-H., Vahdat, A., Huang, D.-A., Theodorou, E. A., Nie, W., and Anandkumar, A. I 2sb: Image-to-image schr\” odinger bridge. The Fortieth International Conference on Machine Learning, 2023a. Liu, X., Gong, C., et al. Flow straight and fast: Learn- ing to generate and transfer data with rectified flow. ...

    work page 2022

  3. [8]

    and Wakolbinger, A

    Pavon, M. and Wakolbinger, A. On free energy, stochas- tic control, and schr ¨odinger processes. In Modeling, Estimation and Control of Systems with Uncertainty: Proceedings of a Conference held in Sopron, Hungary, September 1990, pp. 334–348. Springer,

    work page 1990

  4. [12]

    De- noising diffusion bridge models

    Zhou, L., Lou, A., Khanna, S., and Ermon, S. De- noising diffusion bridge models. In The Twelfth International Conference on Learning Representations, 2024a. URL https://openreview.net/forum? id=FKksTayvGo. Zhou, M., Zheng, H., Wang, Z., Yin, M., and Huang, H. Score identity distillation: Exponentially fast distilla- tion of pretrained diffusion models fo...

    work page 2023

  5. [14]

    Proof of Proposition 3.3

    Integrating these components, we arrive at the final result: Ext,t λ(t)||v(xt, t) − v∗(xt, t)||2 = Ext,t,x0 λ(t)||∇xt log q(xt|x0)||2 − 2Ext,t λ(t) Ex0|xt,t ∇xt log q(xt|x0)|xt, t , v∗(xt, t) + Ext,t,x0 λ(t)||v∗(xt, t)||2 − min v′ Ext,t,x0 λ(t)∥v′(xt, t) − ∇xt log q(xt|x0)∥2 = Ext,t,x0 λ(t)||∇xt log q(xt|x0)||2 − 2Ext,t,x0 λ(t) ⟨∇xt log q(xt|x0), v∗(xt, t...

    work page 2000

  6. [15]

    Our implementation of the degradation operators aligns with the filters implementation proposed in DDRM (Kawar et al., 2022)

    4× super-resolution. Our implementation of the degradation operators aligns with the filters implementation proposed in DDRM (Kawar et al., 2022). Firstly, we synthesize images at 64 × 64 resolution, then upsample them to 256 × 256 to ensure dimensional consistency between clean and degraded inputs. For evaluation, we follow established benchmarks (Sahari...

    work page 2022

  7. [16]

    JPEG restoration

    by computing the FID on reconstructions from the full ImageNet validation set, with comparisons drawn against the training set statistics. JPEG restoration. Our JPEG degradation implementation, employing two distinct quality factors (QF=5, QF=10), follows (Kawar et al., 2022). FID is evaluated on a 10, 000-image ImageNet validation subset against the full...

    work page 2022

  8. [18]

    Subsequent sections outline the experimental setups, adapted from the DDBM (Zheng et al., 2024)

    by integrating our distillation framework. Subsequent sections outline the experimental setups, adapted from the DDBM (Zheng et al., 2024). Multi-step implementation In this setup, the multi-step training (Section 3.5) adopts the methodology of DMD (Yin et al., 2024a), wherein a timestep t is uniformly sampled from the predefined sequence (t1, . . . , tN ...

    work page 2024

Show all 19 references
  1. [19]

    Two versions were trained under the multi-step regime (Section 3.5), with 2 and 1 NFEs during training

    Edges → Handbags The model was trained utilizing the Edges→Handbags image-to-image translation task (Isola et al., 2017), with the 64 × 64 resolution images. Two versions were trained under the multi-step regime (Section 3.5), with 2 and 1 NFEs during training. Both models wer...

    2017

  2. [128]

    ImageNet I2SB 4 5:1 5e-5 2000 ✗ Sketch to Image Edges → Handbags DDBM 2 5:1 1e-5 300 ✓ Sketch to Image Edges → Handbags DDBM 1 5:1 1e-5 14000 ✓ Normal to Image DIODE-Outdoor DDBM 2 5:1 1e-5 500 ✓ Normal to Image DIODE-Outdoor DDBM 1 5:1 1e-5 3700 ✓ Center-inpainting (128 ×

    2000

  3. [1991]

    Diffusion bridge mixture transports, schr¨odinger bridge problems and generative modeling

    Peluchetti, S. Diffusion bridge mixture transports, schr¨odinger bridge problems and generative modeling. Journal of Machine Learning Research, 24(374):1–51, 2023a. Peluchetti, S. Non-denoising forward-time diffusions. arXiv preprint arXiv:2312.14589, 2023b. Saharia, C., Chan,...

    2022 arXiv

  4. [2015]

    R., Pariset, M., Hsieh, Y .-P., Martinez, M

    Somnath, V . R., Pariset, M., Hsieh, Y .-P., Martinez, M. R., Krause, A., and Bunne, C. Aligned diffusion schr ¨odinger bridges. In Uncertainty in Artificial Intelligence, pp. 1985–1995. PMLR,

    1985

  5. [2018]

    A., and Nie, W

    De Bortoli, V ., Liu, G.-H., Chen, T., Theodorou, E. A., and Nie, W. Augmented bridge matching. arXiv preprint arXiv:2311.06978,

    arXiv

  6. [2020]

    Flow generator matching

    Huang, Z., Geng, Z., Luo, W., and Qi, G.-j. Flow generator matching. arXiv preprint arXiv:2410.19310,

    arXiv

  7. [2021]

    Z., Daniele, A

    Vasiljevic, I., Kolkin, N., Zhang, S., Luo, R., Wang, H., Dai, F. Z., Daniele, A. F., Mostajabi, M., Basart, S., Walter, M. R., et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463,

    1908 arXiv

  8. [2022]

    J., Nie, W., Vahdat, A., Lee, S.- g., Santos, J

    Kong, Z., Shih, K. J., Nie, W., Vahdat, A., Lee, S.- g., Santos, J. F., Jukic, A., Valle, R., and Catanzaro, B. A2sb: Audio-to-audio schrodinger bridges. arXiv preprint arXiv:2501.11311,

    arXiv

  9. [2023]

    Imagenet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee,

    2009

  10. [2024]

    and CDBM (He et al., 2024). During training, the model is trained only on the masked regions, while during generation, the unmasked areas are deterministically retained from the initial corrupted image xT to preserve structural fidelity of unmasked part of images. We trained t...

    2024

  11. [2025]

    Bbdm: Image-to- image translation with brownian bridge diffusion mod- els

    Li, B., Xue, K., Liu, B., and Lai, Y .-K. Bbdm: Image-to- image translation with brownian bridge diffusion mod- els. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pp. 1952–1961,

    1952

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.