Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

UniDB: A Unified Diffusion Bridge Framework via Stochastic Optimal Control

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read The paper shows that diffusion bridges built on Doob's h-transform are a limit case of one stochastic optimal control problem, and that using a finite endpoint penalty instead restores detail that the limit blurs.

desk verdict The SOC-to-Doob unification is real and clean, but the finite-gamma training objective is built on a bridge variance that does not match the sampling SDE, so the main empirical claim needs more work. read the letter →

arxiv 2502.05749 v5 pith:5UIKRY56 submitted 2025-02-09 cs.CV cs.AIcs.SYeess.SY

classification cs.CVcs.AIcs.SYeess.SY MSC 60H1049K4593E20
keywords diffusionbridgestochasticoptimalcontrolDoob'sh-transformimagerestorationscore-basedgenerativemodelsGeneralizedOrnstein-Uhlenbeckterminalpenaltycoefficient
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

The paper tries to establish that diffusion bridges — stochastic paths that carry one image distribution into another for restoration tasks — are best understood, and best built, as solutions of a stochastic optimal control (SOC) problem rather than through the standard Doob h-transform recipe. Its central claim is that Doob's h-transform is exactly the $\gamma\to\infty$ limit of that control problem: the limit in which missing the target endpoint costs infinitely much, so the controller is forced to hit the point exactly while paying no attention to trajectory cost. The paper then demonstrates that this infinite penalty is suboptimal — the finite-$\gamma$ controller achieves strictly lower total cost — and that it is the source of the blurred, over-smoothed details seen in h-transform bridges, while a well-chosen finite $\gamma$ preserves high-frequency detail. It derives a closed-form optimal controller, identifies DDBMs and GOUB as the $\gamma\to\infty$ special cases, and shows on super-resolution, deraining, and inpainting that the finite-$\gamma$ version improves perceptual quality with only a few lines of code changed. A sympathetic reader would care because the result turns an empirical artifact of diffusion bridges into a tunable hyperparameter backed by a proof.

What carries the argument

The central object is a linear-quadratic stochastic optimal control problem with terminal penalty: a quadratic running cost $\tfrac{1}{2}\|u_{t,\gamma}\|^2$ on the control, plus $\tfrac{\gamma}{2}\|x^u_T - x_T\|^2$ penalizing deviation from the target endpoint. Because the dynamics are linear and the cost quadratic, the certainty equivalence principle lets the noise be dropped to a deterministic ODE without changing the optimal control, and that ODE is solved in closed form. The quantity that carries the argument is the regulator $d_{t,\gamma} = \gamma^{-1} + e^{2\bar{f}_T}\bar{g}^2_{t:T}$: it appears in the denominator of the optimal controller and continuously interpolates between the uncontrolled diffusion as $\gamma\to 0$ and Doob's h-transform as $\gamma\to\infty$. The same formula reproduces DDBMs (VE), DDBMs (VP), and GOUB as hyperparameter choices (Proposition 4.4), so unification and the finite-$\gamma$ fix come from one expression. The second piece of machinery is the forward-transition variance $\sigma'^2_t = \bar{\sigma}^2_t\bar{\sigma}^2_{t:T}/\bar{\sigma}^2_T$, imported from GOUB through stochastic interpolant theory to reinsert noise into the deterministic trajectory; it makes the conditional score-matching training loss $\mathbb{E}\left[\frac{1}{2\sigma^2_{t-1,\theta}}\|\mu_{t-1,\theta}-\mu_{t-1,\gamma}\|_1\right]$ tractable for finite $\gamma$, at the cost of assuming the bridge still reaches the endpoint exactly.

What would settle it

Fix a clean image $x_0$ and its degraded partner $x_T$, simulate the finite-$\gamma$ forward SDE of UniDB-GOU (Eq. 20) many times at a paper-scale $\gamma$ (around $10^7$), and measure the empirical variance of the terminal states. If that variance is comparable to the GOUB terminal variance $\lambda^2(1-e^{-2\bar{\theta}_T})$ instead of negligible, the assumption $\sigma'^2_T = 0$ behind the training loss (Eq. 19) is violated; one could then retrain with a numerically estimated true conditional transition and compare restoration quality to test whether the mismatch actually hurts.

Watch

Extended reading notes

Core claim

UniDB solves the linear-quadratic SOC problem $\min \mathbb{E}\left[\int_0^T \tfrac{1}{2}\|u_{t,\gamma}\|^2\,dt + \tfrac{\gamma}{2}\|x^u_T - x_T\|^2\right]$ with the linear SDE constraint $dx_t = (f_t x_t + h_t m + g_t u_{t,\gamma})\,dt + g_t\,dw_t$, $x^u_0 = x_0$, and obtains, via the certainty equivalence principle, the closed-form optimal controller (Theorem 4.1) $u^*_{t,\gamma} = g_t e^{\bar{f}_{t:T}}\left(x_T - e^{\bar{f}_{t:T}}x_t - m e^{\bar{f}_T}\bar{h}_{t:T}\right)/\left(\gamma^{-1} + e^{2\bar{f}_T}\bar{g}^2_{t:T}\right)$. The paper's discovery is Theorem 4.2: as $\gamma\to\infty$ this controller becomes $g_t \nabla_{x_t}\log p(x_T|x_t)$, exactly the h-transform term used by DDBMs and GOUB, so every existing Doob bridge is a special case of UniDB with an infinitely stiff terminal penalty. Proposition 4.3 then proves $J(u^*_{t,\gamma},\gamma) \le J(u^*_{t,\infty},\infty)$: forcing exact endpoint matching costs more than the finite-$\gamma$ controller, which trades a tiny, closed-form terminal miss (quantified in Proposition 4.5) for smoother trajectories. Instantiating the framework on the GOU process, UniDB-GOU replaces the GOUB coefficient $e^{-\bar{\theta}_t}\bar{\sigma}^2_{t:T}/\bar{\sigma}^2_T$ with $e^{-\bar{\theta}_t}(\gamma^{-1}+\bar{\sigma}^2_{t:T})/(\gamma^{-1}+\bar{\sigma}^2_T)$ and the h-term $g_t e^{-2\bar{\theta}_{t:T}}/\bar{\sigma}^2_{t:T}$ with $g_t e^{-2\bar{\theta}_{t:T}}/(\gamma^{-1}+\bar{\sigma}^2_{t:T})$, and reports better LPIPS and FID than GOUB on DIV2K super-resolution, Rain100H deraining, and CelebA-HQ inpainting.

Load-bearing premise

The load-bearing premise, introduced in Section 4.3 and Appendix A.4, is that the finite-$\gamma$ forward process can be trained as if it ended exactly at the target image, with zero terminal noise, even though the finite-$\gamma$ controller only steers near the target and so produces non-degenerate terminal scatter; if that gap is real, the trained score does not match the SDE used for sampling.

Editorial extensions

If this is right

  • If the central claim is right, DDBMs (VE), DDBMs (VP), and GOUB are not separate models but one SOC solution at $\gamma\to\infty$, so their known blur is a diagnostic of an over-stiff terminal constraint rather than an intrinsic limit of bridging.
  • A finite $\gamma$ becomes a principled dial: lower $\gamma$ gives smoother, more natural trajectories and better perceptual metrics (LPIPS, FID), while higher $\gamma$ tracks the target pixel more faithfully, letting practitioners tune per task and per metric.
  • The upgrade is a one-term substitution (adding $\gamma^{-1}$ to the variance in the coefficient of $x_0$ and in the h-function), so existing h-transform training and sampling loops can adopt the correction with minimal code changes.
  • Proposition 4.5 gives the residual endpoint mismatch $\|x^u_T - x_T\|^2$ in closed form for the GOU case, so $\gamma$ can be set analytically to meet a desired terminal-error tolerance rather than swept empirically.

Reading between the lines

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

  • The $\gamma^{-1}$ term acts like a regularizer on the h-function; an untested conjecture is that the same substitution transfers to any Doob-bridge application — inverse problems, image translation, conditional generation — wherever high-frequency fidelity matters.
  • The paper never connects $\gamma$ to the entropic temperature of Schr\"odinger bridge problems, whose soft endpoint constraints are analogous; reading $\gamma$ as a regularization temperature is a natural next step that could import the closed-form controller into optimal-transport methods.
  • Because Proposition 4.5 links the endpoint miss to the actual distance $\|x_T - x_0\|^2$ between degraded and clean images, a task-adaptive $\gamma$ (smaller for mild degradations, larger for severe ones) is a concrete extension the paper's own formula suggests but its experiments do not test.
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 / 4 minor

Summary. The paper proposes UniDB, a stochastic optimal control (SOC) framework for diffusion bridges. It casts the bridge construction as a linear-quadratic control problem with a terminal penalty coefficient γ, derives a closed-form optimal controller, and proves that as γ→∞ the controlled SDE recovers the Doob h-transform bridges underlying DDBMs and GOUB. For finite γ, the authors propose a training objective based on an interpolant whose mean is the deterministic optimal path and whose variance is taken from the γ→∞ bridge. Experiments on image super-resolution, deraining, and inpainting report perceptual improvements (LPIPS, FID) over GOUB, with the optimal γ varying by task.

Significance. The γ→∞ limit is rigorously derived and provides a clean unification of existing diffusion bridge models under SOC, which is a valuable conceptual contribution. The framework is general and the code release supports reproducibility. However, the finite-γ generalization, which is the paper's main novelty, currently relies on an unverified approximation in the training transition. Until that gap is closed, the empirical improvements are not explained by the theory, and the paper's central claim of a principled finite-γ improvement is not fully supported.

major comments (3)
  1. [Section 4.3, Eq. (18) and Eq. (20)] The training transition p(xt|x0,xT) in Eq. (18) uses the variance σ̄'^2_t = σ̄^2_t σ̄^2_{t:T}/σ̄^2_T, which is the γ→∞ GOUB bridge variance and vanishes at t=T. However, the finite-γ forward SDE (20) has a transition variance V_t = ∫_0^t exp(-2∫_s^t a_u du) g_s^2 ds with a_t = θ_t + g_t^2 e^{-2θ̄_{t:T}}/(γ^{-1}+σ̄^2_{t:T}), which is strictly positive at t=T for any finite γ. Appendix A.4 postulates σ̄'^2_t without deriving it for the controlled SDE. Consequently, the score network trained on the interpolant samples in Algorithm 1 is not the score of the sampling SDE in Algorithm 2, and the reverse drift (15) is inconsistent with the training objective (19). The γ→∞ limit (Theorem 4.2) is unaffected, but the central finite-γ improvement claim rests on this unverified approximation.
  2. [Section 4.5, Eq. (21) and Algorithm 2] For finite γ, Eq. (21) gives μ̄_{T,γ} = e^{-θ̄_T}(1+γσ̄^2_{T:T})/(1+γσ̄^2_T) x0 + (1 - e^{-θ̄_T}(1+γσ̄^2_{T:T})/(1+γσ̄^2_T)) x_T, which is not equal to x_T unless γ→∞. The controlled process therefore does not terminate at x_T, yet Algorithm 2 initializes the reverse process at x_T and the training interpolant forces σ̄'_T = 0, effectively conditioning on an endpoint that the forward process never reaches deterministically. The paper does not quantify the resulting boundary error beyond choosing e^{θ̄_T}=0.005 and large γ; a bound or an alternative construction is needed to justify the procedure.
  3. [Section 4.2, Proposition 4.3] Proposition 4.3 establishes only that the SOC cost J is lower for finite γ than for γ=∞; it does not establish that finite γ produces better image quality. The experimental evidence in Table 2 shows that the optimal γ is task-dependent and that γ=∞ yields better PSNR/SSIM in most cases, with finite γ improving only perceptual metrics. The statement that Doob's h-transform is 'suboptimal' should be qualified to the perceptual metrics and to the specific SOC cost, rather than presented as a general limitation.
minor comments (4)
  1. [Section 4.3] The phrase 'The derailed derivation is provided in Appendix A.4' should read 'The detailed derivation is provided in Appendix A.4'.
  2. [Eq. (19) vs. Algorithm 1] The score normalization differs between Eq. (19), which has g_t^2/σ̄'_t εθ, and Algorithm 1, which uses g_t^2/σ̄'^2_t εθ; please clarify the intended parameterization and ensure consistency.
  3. [Algorithm 1] The line 'Take a pair of images x0 = x0 and xT = xT' is informal; state that x0 is sampled from the data distribution and xT from the paired conditioning image.
  4. [Notation in Eq. (19) and Algorithm 1] The symbol a_{t,γ} is used in Eq. (19) for e^{f̄_t} d_{t,γ} but in Algorithm 1 for the coefficient e^{-θ̄_t}(γ^{-1}+σ̄^2_{t:T})/(γ^{-1}+σ̄^2_T); unify the notation to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the gamma-to-infinity reduction to Doob h-transforms is a genuine mathematical derivation, and the finite-gamma variance choice is an explicit ansatz rather than a disguised input.

full rationale

The central theoretical claim is Theorem 4.2, which computes the gamma-to-infinity limit of the closed-form PMP controller (13) and identifies it with gt * grad_x log p(xT | xt) for the uncontrolled linear SDE; this is an externally anchored reduction, not an assumption. Proposition 4.4 similarly specializes the SOC parameters to recover the DDBMs and GOUB SDEs at gamma = infinity. The finite-gamma training transition (18) does contain an imposed element: the variance sigma_bar_prime^2_t is set equal to the GOUB bridge variance 'similar to (7)' rather than derived from the controlled SDE, and this creates a genuine train/sample mismatch risk because Algorithm 2 reverses SDE (20) with nonzero terminal variance. However, this is an unverified modeling assumption or approximation, not a circularity: the paper does not define the SOC solution in terms of the target result, and the improvement claim is empirically tested against external baselines with gamma tuned per task (Table 2), which is a hyperparameter choice rather than a prediction forced by the derivation. Self-citations present in the Related Work are not load-bearing for the equivalence theorems. Therefore no claimed derivation reduces by construction to its own inputs.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The only new free parameter introduced by the paper is gamma, tuned per task. The main ad hoc assumption is the reuse of the infinite-gamma bridge variance for the finite-gamma forward transition. No new physical or conceptual entities are postulated.

free parameters (1)
  • gamma (terminal penalty coefficient) = Task-dependent values from 5e5 to 1e8, e.g., 1e7 for FFHQ super-resolution
    Tuned per dataset and task in Table 2 to maximize reported metrics; the central claim of an optimal balance between control cost and terminal penalty depends on this choice.
assumptions (3)
  • standard math Certainty equivalence: for linear dynamics with quadratic costs, the optimal control of the stochastic problem equals the optimal control of the deterministic problem.
    Invoked in Section 4.1 to replace the SDE in Eq. (11) by the deterministic ODE in Eq. (12) before deriving Theorem 4.1.
  • ad hoc to paper For finite gamma, the forward transition p(xt | x0, xT) has the same variance as the gamma-to-infinity bridge, sigma'^2_t = sigma^2_t sigma^2_{t:T} / sigma^2_T.
    Adopted in Eq. (18) and Appendix A.4 from the GOUB bridge without proof for finite gamma; the controlled SDE (20) does not force the terminal state to xT, so the zero-variance-at-T property is not obviously valid.
  • domain assumption The score network parameterization and l1 training loss follow GOUB, with the score written as -epsilon_theta(x_t, x_T, t) / sigma'_t.
    Used in Section 4.3 and Appendix A.5 to derive the training objective in Eq. (19); inherited from prior image restoration bridge work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniDB: A Unified Diffusion Bridge Framework via Stochastic Optimal Control." pith.science (2026). https://pith.science/paper/5UIKRY56

@misc{pith2026250205749,
  author       = {Pith},
  title        = {Pith review of: UniDB: A Unified Diffusion Bridge Framework via Stochastic Optimal Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5UIKRY56}},
  note         = {Machine review of arXiv:2502.05749}
}
abstract

Recent advances in diffusion bridge models leverage Doob's $h$-transform to establish fixed endpoints between distributions, demonstrating promising results in image translation and restoration tasks. However, these approaches frequently produce blurred or excessively smoothed image details and lack a comprehensive theoretical foundation to explain these shortcomings. To address these limitations, we propose UniDB, a unified framework for diffusion bridges based on Stochastic Optimal Control (SOC). UniDB formulates the problem through an SOC-based optimization and derives a closed-form solution for the optimal controller, thereby unifying and generalizing existing diffusion bridge models. We demonstrate that existing diffusion bridges employing Doob's $h$-transform constitute a special case of our framework, emerging when the terminal penalty coefficient in the SOC cost function tends to infinity. By incorporating a tunable terminal penalty coefficient, UniDB achieves an optimal balance between control costs and terminal penalties, substantially improving detail preservation and output quality. Notably, UniDB seamlessly integrates with existing diffusion bridge models, requiring only minimal code modifications. Extensive experiments across diverse image restoration tasks validate the superiority and adaptability of the proposed framework. Our code is available at https://github.com/UniDB-SOC/UniDB/.

Figures

Figures reproduced from arXiv: 2502.05749 by the authors.

Figure 1
Figure 1. Recent advances in diffusion bridge models leverage Doob’s h-transform to establish fixed endpoints between distributions, which introduces an h function into the forward process of a standard stochastic differential equation (SDE) to forcibly match the two endpoints. However, as shown in the figure, this method can lead to local blurring and distortion in the generated images. UniDB formulates the forward process a… view at source ↗
Figure 2
Figure 2. Qualitative comparison of visual results between GOUB (SDE) and UniDB (SDE) on DIV2K with zoomed-in image local regions (UniDB based on GOU process). Building upon equations (20) and (21), we further present a proposition to characterize how the penalty coefficient γ affects the controlled terminal distribution as follows: Proposition 4.5. Denote the initial state distribution x0, the terminal distribution x u T by … view at source ↗
Figure 4
Figure 4. The distances between target and controlled terminal distributions for different datasets (CelebA-HQ, Rain100H, and DIV2K) with different penalty coefficients γ. The red shaded area and blue dotted line highlight our choice of γ. 5.1. Experiments Setup According to Proposition 4.5, we first quantitatively analyze the l2-norm distances between the two terminal distributions depicted in [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figures from the paper (7 more)
Figure 3
Figure 3. Figure 3: Qualitative comparison of visual results between GOUB (SDE) and UniDB (SDE) on the Rain100H dataset on Image Deraining (Left) and CelebA-HQ dataset on Image Inpainting (Right) with zoomed-in image local regions (UniDB based on GOU process). between high-quality and low…
Figure 5
Figure 5. Figure 5: Additional visual results on deraining with Rain100H datasets. LQ UniDB LQ UniDB LQ UniDB LQ UniDB LQ UniDB [PITH_FULL_IMAGE:figures/full_fig_p025_5.png]
Figure 6
Figure 6. Figure 6: Additional visual results on 4×super-resolution with FFHQ datasets. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: Additional visual results on 4×super-resolution with CelebA-HQ datasets. GT GOUB UniDB GT GOUB UniDB GT GOUB UniDB [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]
Figure 8
Figure 8. Figure 8: Additional visual results on thin mask inpainting with CelebA-HQ datasets to show our excellence. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]
Figure 9
Figure 9. Figure 9: Additional visual results on thin mask inpainting with CelebA-HQ datasets. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]
Figure 10
Figure 10. Figure 10: Additional visual results on 4×super-resolution with DIV2K datasets. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Towards Immersive Human-X Interaction: A Real-Time Framework for Physically Plausible Motion Synthesis

    cs.CV 2025-08 unverdicted novelty 5.0 of 10

    Human-X jointly predicts actions and reactions in real time to produce physically plausible human-machine interaction motion.

  2. Diffusion Bridge or Flow Matching? A Unifying Framework and Comparative Analysis

    cs.CV 2025-09 reject novelty 4.0 of 10

    A theoretical and empirical comparison claiming diffusion bridges have lower stochastic-optimal-control cost and greater robustness than flow matching when training data are scarce.

Reference graph

Works this paper leans on

31 extracted references · 6 canonical work pages · cited by 2 Pith papers

  1. [1]

    and Timofte, R

    Agustsson, E. and Timofte, R. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pp. 126–135,

  2. [2]

    ftxt + htm + g2 t xT − e ¯ft:T xt − me ¯fT ¯ht:T e− ¯ft:T (γ−1 + e2 ¯fT ¯g2 t:T ) − g2 t ∇xt log p(xt | xT ) # − gtϵt ≈ xt −

    (53) 16 Therefore, γ 2 ∥xu T − xT ∥2 2 − lim γ→∞ γ 2 ∥xu T − xT ∥2 2 ≤ 1 2 Z T 0 ∥u∗ t,∞∥2 2 − ∥u∗ t,γ∥2 2 dt (54) ⇔ 1 2 Z T 0 ∥u∗ t,γ∥2 2dt + γ 2 ∥xu T − xT ∥2 2 ≤ 1 2 Z T 0 ∥u∗ t,∞∥2 2dt + lim γ→∞ γ 2 ∥xu T − xT ∥2 2, (55) ⇔ J (u∗ t,γ, γ) ≤ J(u∗ t,∞, ∞), (56) which concludes the proof of Proposition 4.3. A.4. Derivation of the transition probability (18...

  3. [8]

    Gao, H., Zhang, Y ., Yang, J., and Dang, D

    URL https://arxiv.org/abs/2304.11751. Gao, H., Zhang, Y ., Yang, J., and Dang, D. Mixed hierarchy network for image restoration. Pattern Recognition, 161: 111313,

  4. [13]

    Kingma, D. P. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,

  5. [16]

    K., Zhao, Z., Sj¨olund, J., and Sch¨on, T

    Luo, Z., Gustafsson, F. K., Zhao, Z., Sj¨olund, J., and Sch¨on, T. B. Image restoration with mean-reverting stochastic differential equations. arXiv preprint arXiv:2301.11699,

  6. [19]

    Rb-modulation: Training- free personalization of diffusion models using stochastic optimal control

    Rout, L., Chen, Y ., Ruiz, N., Kumar, A., Caramanis, C., Shakkottai, S., and Chu, W.-S. Rb-modulation: Training- free personalization of diffusion models using stochastic optimal control. arXiv preprint arXiv:2405.17401, 2024a. Rout, L., Chen, Y ., Ruiz, N., Kumar, A., Caramanis, C., Shakkottai, S., and Chu, W.-S. Rb-modulation: Training- free personaliza...

  7. [20]

    Gradient-free classifier guid- ance for diffusion model sampling

    Shenoy, R., Pan, Z., Balakrishnan, K., Cheng, Q., Jeon, Y ., Yang, H., and Kim, J. Gradient-free classifier guid- ance for diffusion model sampling. arXiv preprint arXiv:2411.15393,

  8. [22]

    P., Kumar, A., Er- mon, S., and Poole, B

    Song, Y ., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Er- mon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456,

Show all 31 references
  1. [23]

    Afforddp: Generalizable diffu- sion policy with transferable affordance

    Wu, S., Zhu, Y ., Huang, Y ., Zhu, K., Gu, J., Yu, J., Shi, Y ., and Wang, J. Afforddp: Generalizable diffu- sion policy with transferable affordance. arXiv preprint arXiv:2412.03142,

  2. [24]

    Policy representation via diffusion probability model for reinforcement learning

    Yang, L., Huang, Z., Lei, F., Zhong, Y ., Yang, Y ., Fang, C., Wen, S., Zhou, B., and Lin, Z. Policy representation via diffusion probability model for reinforcement learning. arXiv preprint arXiv:2305.13122,

  3. [25]

    Image restoration through generalized ornstein-uhlenbeck bridge

    Yue, C., Peng, Z., Ma, J., Du, S., Wei, P., and Zhang, D. Image restoration through generalized ornstein-uhlenbeck bridge. arXiv preprint arXiv:2312.10299,

  4. [26]

    3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations

    Ze, Y ., Zhang, G., Zhang, K., Hu, C., Wang, M., and Xu, H. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. In ICRA 2024 Workshop on 3D Visual Representations for Robot Ma- nipulation,

  5. [28]

    Denoising dif- fusion bridge models

    Zhou, L., Lou, A., Khanna, S., and Ermon, S. Denoising dif- fusion bridge models. arXiv preprint arXiv:2309.16948,

  6. [29]

    Page 13 – A.1 Proof of Theorem 4.1

    12 Appendix Contents • Appendix A: Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Page 13 – A.1 Proof of Theorem 4.1 . . . . . . . . . . . . . . . . . . . ....

  7. [31]

    According to Appendix A.7, we’ve learned that xu t = e−¯θt 1 + γ ¯σ2 t:T 1 + γ ¯σ2 T x0 + 1 − e−¯θt 1 + γ ¯σ2 t:T 1 + γ ¯σ2 T xT + ¯σ2 t ¯σ2 t:T ¯σ2 T ϵ

    (23) Proof. According to Appendix A.7, we’ve learned that xu t = e−¯θt 1 + γ ¯σ2 t:T 1 + γ ¯σ2 T x0 + 1 − e−¯θt 1 + γ ¯σ2 t:T 1 + γ ¯σ2 T xT + ¯σ2 t ¯σ2 t:T ¯σ2 T ϵ. (82) Take t = T , then xu T = e−¯θt 1 + γ ¯σ2 T x0 + 1 − e−¯θt 1 + γ ¯σ2 T ! xT . (83) 22 Therefore, since ¯σ2 ...

  8. [1972]

    Li, X., Ren, Y ., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., and Chen, Z

    doi: 10.1109/TAC.1972.1100008. Li, X., Ren, Y ., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., and Chen, Z. Diffusion models for image restora- tion and enhancement–a comprehensive survey. arXiv preprint arXiv:2308.09388,

  9. [1982]

    An optimal con- trol perspective on diffusion-based generative modeling

    Berner, J., Richter, L., and Ullrich, K. An optimal con- trol perspective on diffusion-based generative modeling. arXiv preprint arXiv:2211.01364,

  10. [1988]

    S., Boffi, N

    Albergo, M. S., Boffi, N. M., and Vanden-Eijnden, E. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797,

  11. [2003]

    Stochastic optimal control for diffusion bridges in function spaces

    Park, B., Choi, J., Lim, S., and Lee, J. Stochastic optimal control for diffusion bridges in function spaces. arXiv preprint arXiv:2405.20630,

  12. [2004]

    and Nakkiran, P

    Bradley, A. and Nakkiran, P. Classifier-free guidance is a predictor-corrector. arXiv preprint arXiv:2408.09000,

  13. [2008]

    Progressive growing of gans for improved quality, stability, and variation

    10 Karras, T. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196,

  14. [2014]

    W., Khan, S., and Khan, F

    Potlapalli, V ., Zamir, S. W., Khan, S., and Khan, F. S. Promptir: Prompting for all-in-one blind image restora- tion. arXiv preprint arXiv:2306.13090,

  15. [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,

  16. [2017]

    and Salimans, T

    Ho, J. and Salimans, T. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598,

  17. [2018]

    Diffusion bridge implicit models

    Zheng, K., He, G., Chen, J., Bao, F., and Zhu, J. Diffusion bridge implicit models. arXiv preprint arXiv:2405.15885,

  18. [2019]

    Kawar, B., Elad, M., Ermon, S., and Song, J

    URL https://arxiv.org/abs/1812.04948. Kawar, B., Elad, M., Ermon, S., and Song, J. Denoising diffusion restoration models. Advances in Neural Infor- mation Processing Systems, 35:23593–23606,

  19. [2020]

    Levine, W

    URL https://arxiv.org/abs/1907.11922. Levine, W. Optimal control theory: An introduction. IEEE Transactions on Automatic Control, 17(3):423–423,

  20. [2021]

    Diffusion-based reinforcement learning via q-weighted variational policy optimization

    Ding, S., Hu, K., Zhang, Z., Ren, K., Zhang, W., Yu, J., Wang, J., and Shi, Y . Diffusion-based reinforcement learning via q-weighted variational policy optimization. In The Thirty-eighth Annual Conference on Neural Infor- mation Processing Systems, 2024a. Ding, X., Wang, Y .,...

  21. [2022]

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S

    URL https://arxiv.org/abs/2111.07243. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30,

  22. [2023]

    T., Klasky, M

    Chung, H., Kim, J., Mccann, M. T., Klasky, M. L., and Ye, J. C. Diffusion posterior sampling for general noisy in- verse problems. arXiv preprint arXiv:2209.14687,

  23. [2024]

    org/abs/2402.16075

    URL https://arxiv. org/abs/2402.16075. Chen, T., Gu, J., Dinh, L., Theodorou, E. A., Susskind, J., and Zhai, S. Generative modeling with phase stochastic bridges. arXiv preprint arXiv:2310.07805,

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.