Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Outsourced diffusion sampling: Efficient posterior inference in latent spaces of generative models

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

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

Pith's one-line read The paper claims that intractable posterior sampling over data can be replaced by diffusion sampling over a pretrained generator's noise variable, with the generator itself carrying samples back to data space.

desk verdict The core pullback idea is correct and well-packaged; the finite-time TB bias and a few evaluation choices keep this from being a clean accept. read the letter →

arxiv 2502.06999 v2 pith:TJTCOVKD submitted 2025-02-10 cs.LG

Siddarth Venkatraman , Mohsin Hasan , Minsu Kim , Luca Scimeca , Marcin Sendera , Yoshua Bengio , Glen Berseth , Nikolay Malkin This is my paper · ORCID
classification cs.LG
keywords posteriorinferencediffusionsamplersnoisespacetrajectorybalanceamortizedgenerativemodelsblack-boxconstraintslatent-spacesampling
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's thesis is that posterior inference over data — sampling $\mathbf{x}$ with density proportional to a generative prior $p(\mathbf{x})$ times a constraint $r(\mathbf{x},\mathbf{y})$ — can be moved into the noise space of the generator. For any model written as $\mathbf{x}=f_\theta(\mathbf{z})$, drawing $\mathbf{z}$ from the pullback density proportional to $p(\mathbf{z})r(f_\theta(\mathbf{z}),\mathbf{y})$ and outputting $f_\theta(\mathbf{z})$ yields exactly the desired data-space posterior, provided the normalization exists. Because this pullback posterior is often smoother and lower-dimensional than the data-space posterior, the paper trains a diffusion model directly in $\mathbf{z}$-space with the trajectory balance objective, treating $r(f_\theta(\mathbf{z}),\mathbf{y})$ as a black-box reward. The result is an amortized, gradient-free sampler that applies to GANs, hierarchical VAEs, continuous normalizing flows, and diffusion-style priors, demonstrated on conditional image generation, text-to-image reward fine-tuning, and protein structure generation.

What carries the argument

The load-bearing objects are Proposition 3.1's pullback identity and the trajectory balance objective. The identity says the data-space posterior is the pushforward of the noise-space posterior, so a correct noise-space sampler automatically yields correct data-space samples through $f$. The trajectory balance objective, $\mathcal{L}_{\mathrm{TB}}(\tau;y,\phi)=\left(\log \frac{Z_\phi(y)p_\phi^F(\tau|y)}{R(z_1|y)p_B(\tau|z_1)}\right)^2$, matches the distribution over denoising trajectories defined by the trained sampler against the fixed noising kernel weighted by the target density $R(z_1|y)=p(z_1)r(f(z_1),y)$; when it is zero on every trajectory in the continuous-time limit, the sampler's terminal marginal is the target. The objective is off-policy, so the reward function only needs to be queried, not differentiated, and exploration can be biased toward high-reward regions via replay buffers.

What would settle it

Take a low-dimensional prior with an exactly computable posterior, train the outsourced diffusion sampler for an increasing number of steps, and compare its samples against the true pullback posterior with a two-sample test or estimated KL divergence; if the divergence does not decrease toward zero, or if a test reward concentrates on latent directions the sampler is architecturally prevented from representing, then the method is sampling an approximation rather than the claimed posterior.

Watch

Extended reading notes

Core claim

The central claim is a pullback identity (Proposition 3.1): if $z'$ is sampled with density proportional to $p(z)r(f(z),y)$, then $x'=f(z')$ is distributed with density proportional to $p(x)r(x,y)$. The paper argues this turns an intractable data-space posterior into a tractable noise-space sampling target, without ever computing the pushforward density and without requiring $f$ to be injective or smooth. The proposed outsourced diffusion sampler trains a diffusion model on that noise-space target using the trajectory balance objective (Eq. 4), with the prior density $p(z)$ and the constraint evaluated through $f$ as the only required queries. In experiments with unconditional CIFAR-10 generators, the NVAE and StyleGAN3 face generators, Stable Diffusion 3, and a protein backbone flow model, the method produces conditional samples with higher reward and competitive diversity relative to latent HMC and, where applicable, adjoint matching.

Load-bearing premise

The practical advantage rests on the empirical claim that the noise-space posterior is smoother and easier to sample than the data-space posterior, together with the assumption that off-policy trajectory balance training converges to the target within a practical compute budget; in the NVAE experiment there is also the approximation that the first four latent levels capture all reward-relevant variation.

Editorial extensions

If this is right

  • Any prior expressible as a deterministic transformation of Gaussian noise — VAE, hierarchical VAE, GAN, normalizing flow, continuous normalizing flow, or diffusion/latent diffusion — can be conditioned on a constraint by training one noise-space sampler, with no change to the algorithm.
  • Because only queries of $r(f(z),y)$ are needed, the method applies to constraints whose gradients are unavailable or undefined, such as the discrete secondary-structure diversity reward in the protein experiment.
  • Conditioning the diffusion sampler on $y$ amortizes the cost over the constraint variable: after training, posterior samples for a new $y$ are generated by a fast neural SDE rollout rather than by MCMC.
  • A distilled one-step version of the trained sampler matches or improves on the multi-step sampler's FID in the CIFAR-10 experiment, so the posterior inference cost can be reduced to a single network evaluation.
  • The method avoids the bias that the paper reports for naive adjoint matching on flow models trained with minibatch optimal transport couplings or non-Gaussian sources.

Reading between the lines

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

  • If the smoothness observation generalizes, the same pullback recipe could turn any high-dimensional inverse problem with an expressive pretrained generator into a lower-dimensional variational problem, potentially reducing the training compute needed for reward fine-tuning; this is an extension the paper motivates but does not test.
  • Because the target $R(z|y)$ factorizes as prior times reward through $f$, one trained noise-space sampler could be reused across many reward functions for the same prior by multiplying rewards during training, a compositional property the paper leaves implicit.
  • For non-injective generators, the pullback posterior places mass on every noise preimage of a data point; the paper's mode-mixing results suggest this latent-space spread may act as a useful regularizer, but a direct comparison of output diversity against an exact data-space posterior would be needed to confirm it.
  • A natural stress test is to apply the method to a latent diffusion prior where the omitted high-level latents carry reward-relevant information; the NVAE experiment restricts to the first four latent levels, so a mismatch would show up as reward that the sampler cannot reach.
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

4 major / 4 minor

Summary. The paper proposes "outsourced diffusion sampling" for posterior inference under generative models that are deterministic transformations x=f_theta(z) of a Gaussian noise variable. For a constraint r(x,y), the posterior p(x|y) proportional to p(x)r(x,y) is pulled back to the noise space: Proposition 3.1 states that if z' is drawn from a density proportional to p(z)r(f(z),y), then f(z') is distributed as the desired posterior. The paper trains a diffusion sampler in z-space using the trajectory balance objective (Eq. 4), treating R(z|y)=p(z)r(f(z),y) as an unnormalized target, and transforms samples by f. It demonstrates the approach on CIFAR-10 class-conditional generation with SN-GAN and flow-matching priors, FFHQ text-conditioned generation with StyleGAN3 and NVAE priors, reward-aligned text-to-image generation with Stable Diffusion 3, and protein secondary-structure diversification with FoldFlow 2, plus a one-step distillation variant. The exposition includes the noise-outsourcing lemma (Prop. 2.1), a proof of Prop. 3.1, and an appendix showing that relative trajectory balance is recovered as a special case.

Significance. Proposition 3.1 is proved in the appendix and gives a clean, correct reduction: posterior sampling in data space is exactly equivalent to sampling the pullback target R(z|y) when the noise-space sampler is exact. The framework is genuinely model-agnostic, works with black-box non-differentiable constraints, and is demonstrated across four application domains with large pretrained priors; the code is provided. The Appendix A.1 connection between the proposed objective and RTB is a useful unifying observation. If the exactness caveats raised below are addressed, this would be a broadly applicable amortized inference recipe. At present, however, the central claim of sampling fidelity is weakened by an unquantified bias in the trajectory-balance training objective and by experiment-specific approximations (latent truncation, best-checkpoint selection) that are not flagged in the main text.

major comments (4)
  1. [§4.2, Eq. (4)] The trajectory-balance objective in Eq. (4) has an irreducible minimum for the finite-time VP-SDE backward policy used in all experiments (Appendix B.1), because the forward path measure starts from p(z0)=N(0,I), while the target path measure R(z1|y)p_B(tau|z1) has initial marginal q(z0)=Z^{-1}∫R(z1|y)p_B(z0|z1) dz1, which is generally not N(0,I) for a nontrivial reward r. Consequently no policy can drive L_TB to zero on all trajectories, and the global optimum of Eq. (4) is a biased approximation of R(z|y), not an exact sampler. The paper reports no quantitative comparison of the learned z-marginal against R(z|y) (Tables 3-6 only report reward, FID, and diversity), so the magnitude of this bias is unknown. Please add such a diagnostic, e.g., effective sample size or sliced-Wasserstein distance between samples and R/Z on held-out z, or modify the objective so that the initial marginal of the target path measure matches the sampler's initial distribution.
  2. [§5.2 and Appendix B.3] The NVAE experiment samples noise only for the first 4 of 36 latent groups (z in R^{4x20x8x8}), setting or ignoring the remaining latents. This makes the target a conditional distribution given the omitted latent levels, not the full posterior p(x|y); the main text does not flag this as an approximation and labels the outputs 'posterior samples.' The HMC baseline targets the same truncated distribution, so the comparison is internally fair, but the claim of posterior inference under an HVAE prior is not established. Please quantify the effect of the omitted latents (e.g., by measuring reward sensitivity to those levels or by marginalizing them in a tractable toy setting) or explicitly reframe the experiment as conditional sampling from a truncated latent posterior.
  3. [§5.4 and Appendix B.5] The protein experiment selects the model checkpoint with the highest diversity score during training ('we evaluate the model achieving the highest diversity score during training'; Appendix B.5 saves models every 100 iterations and selects the best by reward). This is a form of selection on the evaluation metric: the reported outcome is the best of many checkpoints, not the performance of a fixed training procedure, and the random-walk MCMC baseline is not given the same selection. With only 3 seeds and large standard deviations (reward 0.422±0.225, pairwise TM-score 0.4407±0.0706), the claim that outsourced diffusion 'samples diverse protein structures ... more frequently than the baselines' is not supported as stated. Report performance at a pre-specified checkpoint or budget with a validation protocol, and report all training curves.
  4. [Tables 3-5] The main comparison tables report point estimates without error bars or numbers of seeds. In Table 3, for example, the I-CFM reward difference between latent HMC (-2.80) and outsourced diffusion (-3.35) and the SN-GAN FID differences could easily be within run-to-run noise. Since the paper's headline claims are comparative ('comparing favorably with other inference methods'), single-run differences of this size are not informative. Add repeated-seed statistics (mean±std over at least 3 seeds) for Tables 3-5, or state clearly when a result is a single run.
minor comments (4)
  1. [Abstract and §5.1] The abstract and §1 claim the method 'compares favorably with other inference methods'; in Table 3, adjoint matching achieves better FID (19.45 vs 34.28) and reward (-3.09 vs -3.35) on the same I-CFM prior. Suggest qualifying the claim, e.g., 'competitive across settings, with advantages where gradients are unavailable.'
  2. [§2, Prop. 2.1] Proposition 2.1 states the Gaussian noise-outsourcing lemma, but no proof or precise citation is given for the Gaussian-specific extension; since this is foundational for the paper's framing, include a proof sketch or a specific reference.
  3. [Abstract and Figs. 1-2] The claim that noise-space posteriors are 'smoother' than data-space posteriors is supported only by illustrative marginal plots; consider adding a quantitative measure (e.g., effective dimension or mode count) or stating it explicitly as a motivating heuristic.
  4. [Typos and copyediting] Typos and copyediting: 'classifer-free guidance' in the Figure 29/32/35/38 captions, 'Furthemore' in Appendix C.1, 'CMF posterior' should be 'CFM posterior' in Appendix C.1, 'University of Tront' in the Krizhevsky reference, and 'Maching Learning' in the Song et al. (2023) reference.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation is present: Prop. 3.1 is proved from measure theory, the TB objective is prior work, and the only self-citation is a non-load-bearing appendix connection to RTB.

full rationale

The central derivation is self-contained and not circular. Proposition 3.1 is a standard measure-theoretic pushforward identity proved directly in Appendix A.2; the pullback posterior R(z|y)=p(z)r(f(z),y) is defined exactly as the density whose pushforward under f is proportional to p(x)r(x,y), so no fitted quantity is relabelled as a prediction. The training objective (Eq. 4) is the trajectory-balance loss from prior work (Malkin et al. 2022; Lahlou et al. 2023; Sendera et al. 2024; Berner et al. 2025), and the paper's exactness claim is explicitly asymptotic ('in the continuous-time limit'), not derived from the present experiments. The only overlapping-author citation that could raise a self-citation flag is Appendix A.1, which shows that autoregressive outsourced sampling under a diffusion prior recovers the relative trajectory balance objective of Venkatraman et al. (2024); this is presented as a connection and is not load-bearing for Prop. 3.1 or for the main experiments, which compare against external baselines (HMC, adjoint matching, CFG, DPS, RTB). No fitted constant is renamed as a prediction and no uniqueness theorem is imported to forbid alternatives. Two non-circular limitations should be noted. First, with the fixed 25-step VP-SDE backward policy (Appendix B.1), the target path measure's initial marginal is generally not the fixed N(0,I) of Eq. (3), so the TB loss in Eq. (4) may have an irreducible minimum and the learned sampler can be biased; this is an approximation and correctness concern, not a circularity. Second, the NVAE experiment samples only the first 4 latent levels (Appendix B.3), so it targets an approximate posterior; again an approximation, not a circular step. The score of 2 reflects only the minor, non-load-bearing self-citation in Appendix A.1.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities. The learned partition function Z_phi(y) and diffusion sampler parameters are variational parameters, not free parameters in the derivation. The ledger instead lists the reward temperature and latent-space truncation, both of which alter the target distribution and are chosen by hand, plus the background results the central claim rests on.

free parameters (2)
  • Reward inverse temperature beta = 4 (CIFAR), 100 (FFHQ), 30 (SD3), 400 (protein)
    The target density R(z|y) = p(z) r(f(z), y) is replaced by p(z) r(f(z), y)^beta; beta is chosen by hand per experiment and changes the posterior being sampled from the stated untempered posterior.
  • NVAE latent truncation level = 4 latent groups out of 36
    The FFHQ/NVAE experiment samples only the first 4 latent groups, so the sampler targets an approximate posterior that ignores potentially reward-relevant higher-level latents; this is a model choice rather than a derived necessity.
assumptions (4)
  • standard math Noise outsourcing lemma for Gaussians (Prop 2.1): any generative model with marginally Gaussian latents can be written as a deterministic function of a higher-dimensional Gaussian noise variable.
    Invoked in Section 2 to justify the representation x = f(z); cited to Austin (2015) and stated without proof in the paper.
  • standard math Measure-theoretic pullback of posteriors (Prop 3.1): sampling z from the pulled-back density and pushing forward through f yields the exact data-space posterior.
    This is the core exactness result; it is proved in Appendix A.2 using standard pushforward measure arguments.
  • domain assumption Trajectory balance objective is asymptotically consistent for diffusion samplers in the continuous-time limit.
    The paper relies on consistency of the TB objective and off-policy training, citing Berner et al. (2025) and Lahlou et al. (2023); this convergence is not proved in the present paper.
  • domain assumption The constraint function r(x, y) is positive, finite, and queryable as a black box, so log r is well-defined and finite for the z-space target.
    Assumed throughout Section 3 and used in the reward definition and Algorithm 1; not stated as an explicit limitation, though negative or zero rewards would break the formulation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Outsourced diffusion sampling: Efficient posterior inference in latent spaces of generative models." pith.science (2026). https://pith.science/paper/TJTCOVKD

@misc{pith2026250206999,
  author       = {Pith},
  title        = {Pith review of: Outsourced diffusion sampling: Efficient posterior inference in latent spaces of generative models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TJTCOVKD}},
  note         = {Machine review of arXiv:2502.06999}
}
abstract

Any well-behaved generative model over a variable $\mathbf{x}$ can be expressed as a deterministic transformation of an exogenous ('outsourced') Gaussian noise variable $\mathbf{z}$: $\mathbf{x}=f_\theta(\mathbf{z})$. In such a model (\eg, a VAE, GAN, or continuous-time flow-based model), sampling of the target variable $\mathbf{x} \sim p_\theta(\mathbf{x})$ is straightforward, but sampling from a posterior distribution of the form $p(\mathbf{x}\mid\mathbf{y}) \propto p_\theta(\mathbf{x})r(\mathbf{x},\mathbf{y})$, where $r$ is a constraint function depending on an auxiliary variable $\mathbf{y}$, is generally intractable. We propose to amortize the cost of sampling from such posterior distributions with diffusion models that sample a distribution in the noise space ($\mathbf{z}$). These diffusion samplers are trained by reinforcement learning algorithms to enforce that the transformed samples $f_\theta(\mathbf{z})$ are distributed according to the posterior in the data space ($\mathbf{x}$). For many models and constraints, the posterior in noise space is smoother than in data space, making it more suitable for amortized inference. Our method enables conditional sampling under unconditional GAN, (H)VAE, and flow-based priors, comparing favorably with other inference methods. We demonstrate the proposed outsourced diffusion sampling in several experiments with large pretrained prior models: conditional image generation, reinforcement learning with human feedback, and protein structure generation.

Figures

Figures reproduced from arXiv: 2502.06999 by the authors.

Figure 2
Figure 2. Marginal densities of a diffusion sampler of the posteriors from the CNF example in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. CIFAR-10 samples generated using SN-GAN (top row) and CFM (bottom row) priors and posterior samples from trained outsourced diffusion models for the ‘Dog’ and ‘Car’ classes. the outsourced diffusion model offers significant advantages in training efficiency, requiring approximately 5 hours on an A100 GPU compared to 12 hours for adjoint matching (see Appendix C.2). Moreover, we show how we can distill one-step outso… view at source ↗
Figure 5
Figure 5. Sampled images from the FFHQ priors and outsourced diffusion posteriors for different prompts. More examples in Appendix D [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figures from the paper (18 more)
Figure 6
Figure 6. Figure 6: Sampled images from the SD3 prior and outsourced dif￾fusion posterior for different prompts. More examples in Ap￾pendix E. 5.4. Protein Secondary Structure Diversity Setup. Many protein generative models have been proposed to tackle the problem of designing novel yet r…
Figure 7
Figure 7. Figure 7: Protein samples with pertinent secondary structures high￾lighted: 𝛼-helices (blue), 𝛽-sheets (red), and coils (green). strated the effectiveness of this method in a variety of set￾tings, there are many questions for future work. One natural direction is to extend the m…
Figure 8
Figure 8. Figure 8: CIFAR-10 samples generated Adjoint Matching (Adj.) for the Dog and Car classes. need to train a model, outsourced diffusion is actually more memory-efficient than even a single HMC chain, as it eliminates the need to compute gradients through the ODE integrator. We vis…
Figure 9
Figure 9. Figure 9: Posterior samples via the original outsourced diffusion sampler and the distilled one-step sampler in three experiments. We use the prompt ”A cat and a dog.” for the stable diffusion experiments, and ”An old man” for the SN-GAN experiments. C. Additional Experiments C.…
Figure 10
Figure 10. Figure 10: NVAE prior [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]
Figure 12
Figure 12. Figure 12: NVAE HMC [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 16
Figure 16. Figure 16: NVAE HMC [PITH_FULL_IMAGE:figures/full_fig_p023_16.png]
Figure 20
Figure 20. Figure 20: NVAE HMC [PITH_FULL_IMAGE:figures/full_fig_p024_20.png]
Figure 24
Figure 24. Figure 24: NVAE HMC [PITH_FULL_IMAGE:figures/full_fig_p025_24.png]
Figure 28
Figure 28. Figure 28: Prior [PITH_FULL_IMAGE:figures/full_fig_p026_28.png]
Figure 31
Figure 31. Figure 31: Prior [PITH_FULL_IMAGE:figures/full_fig_p026_31.png]
Figure 34
Figure 34. Figure 34: Prior [PITH_FULL_IMAGE:figures/full_fig_p027_34.png]
Figure 35
Figure 35. Figure 35: Classifer-Free Guidance [PITH_FULL_IMAGE:figures/full_fig_p027_35.png]
Figure 36
Figure 36. Figure 36: Outsourced Diffusion E.4. A human with a horse face and a human with a wolf face [PITH_FULL_IMAGE:figures/full_fig_p027_36.png]
Figure 37
Figure 37. Figure 37: Prior [PITH_FULL_IMAGE:figures/full_fig_p027_37.png]
Figure 38
Figure 38. Figure 38: Classifer-Free Guidance [PITH_FULL_IMAGE:figures/full_fig_p027_38.png]
Figure 39
Figure 39. Figure 39: Outsourced Diffusion 27 [PITH_FULL_IMAGE:figures/full_fig_p027_39.png]
Figure 40
Figure 40. Figure 40: Samples from HMC chain of length 1000 after 100 steps of burn-in, for CIFAR class ’Airplane’ with CNF prior. We find that in the latent space, MCMC smoothly traverses through different modes. The samples (d,o,q,r,s) are distinctly identifiable as airplanes and (c) is …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Na\"ive PAINE: Lightweight Text-to-Image Generation Improvement with Prompt Evaluation

    cs.CV 2026-03 conditional novelty 6.0 of 10

    A lightweight predictor ranks initial noises by expected human-preference score for a prompt, selecting the best few for diffusion generation and reporting prompt difficulty.

Reference graph

Works this paper leans on

18 extracted references · 10 canonical work pages · cited by 1 Pith paper

  1. [5]

    Alias-free generative adver- sarial networks

    Karras, T., Aittala, M., Laine, S., H ¨ark¨onen, E., Hellsten, J., Lehtinen, J., and Aila, T. Alias-free generative adver- sarial networks. Neural Information Processing Systems (NeurIPS), 2021a. Karras, T., Laine, S., and Aila, T. A Style-Based Generator Architecture for Generative Adversarial Networks . IEEE Transactions on Pattern Analysis & Machine In...

    arXiv 1939

  2. [6]

    and Cipcigan, F

    Phillips, D. and Cipcigan, F. MetaGFN: Exploring dis- tant modes with adapted metadynamics for continuous GFlowNets. arXiv preprint arXiv:2408.15905,

    arXiv

  3. [7]

    and Raginsky, M

    Tzen, B. and Raginsky, M. Neural stochastic differential equations: Deep latent Gaussian models in the diffusion limit. arXiv preprint arXiv:1905.09883,

    arXiv 1905

  4. [8]

    L., Tseng, A

    Uehara, M., Zhao, Y ., Black, K., Hajiramezanali, E., Scalia, G., Diamant, N. L., Tseng, A. M., Biancalani, T., and Levine, S. Fine-tuning of continuous-time diffusion models as entropy-regularized control. arXiv preprint arXiv:2402.15194,

    arXiv

  5. [10]

    Posterior estimation can also be achieved through stochastic optimization (Graikos et al., 2022; Mardani et al.,

    or using approximations specialized for constraints arising from linear inverse problems (Kawar et al., 2021; Kadkhodaie & Simoncelli, 2021; Song et al., 2022; Chung et al., 2023). Posterior estimation can also be achieved through stochastic optimization (Graikos et al., 2022; Mardani et al.,

    work page 2021

  6. [11]

    Amortized methods

    Methods related to sequential Monte Carlo, which treat a modification to the denoising transition kernel as a proposal, have also been proposed (Doucet et al., 2022; Dou & Song, 2024; Chen et al., 2025). Amortized methods. Because generation proceeds in a long sequence of sampling steps, and the modes of the posterior are not known a priori, these methods...

    work page 2022

  7. [12]

    Outsourcing noise in diffusion models

    and relative trajectory balance (Venkatraman et al., 2024). Outsourcing noise in diffusion models. Generation of data x= x0 is modeled as a Markov process x𝑇→···→ x1→ x0, where x𝑇∼N( 0,𝐼𝑑data) and the transition from x𝑡 to x𝑡−1 is conditionally spherical-Gaussian. Via the reparametrization trick, the trajectory of latent variables can be expressed as a fu...

    work page 2024

  8. [13]

    1.26 0.14 B.5

    SD 1.5 RLHF results, averaged across prompts Sampler E[log𝑟(x, y)](↑) CLIP diversity(↑) Prior -0.17 0.18 DDPO 1.37 0.09 DPOK 1.23 0.13 RTB 1.4 0.11 Outsourced Diff. 1.26 0.14 B.5. Protein Structure Prediction The specific form of the constraint function𝑟div(x) (which can also be thought of as a reward) is adapted from (Huguet et al., 2024). It is a monoto...

    work page 2024

Show all 18 references
  1. [18]

    Outsourced Diffusion

    We found this to improve numerical stability in the diffusion training. Outsourced Diffusion. In the outsourced diffusion experiments, an inverse temperature of𝛽= 400.0 was used to allow for improvement in this sparse reward setting. The diffusion sampler used 20 sampling step...

    2020

  2. [32]

    To fine-tune flow matching models with a classifier, we use adjoint matching following the approach described in Domingo-Enrich et al

    Adjoint Matching. To fine-tune flow matching models with a classifier, we use adjoint matching following the approach described in Domingo-Enrich et al. (2024). Due to the unavailability of the open source code at this time, we made our own implementation which we found works ...

    2024

  3. [64]

    Classifier-Free Guidance. Since SD3 is trained as both an unconditional and conditional model, we can use CFG to approximately sample from lowered temperature conditional distribution: ˆv𝜃(x𝑡, y)=(1+𝑤)v𝜃(x𝑡, y)− 𝑤v𝜃(x𝑡) (8) Increasing the guidance scale 𝑤 generally guides the ...

    2024

  4. [2004]

    Theory and Method Details A.1

    13 Outsourced Diffusion Sampling: Efficient Posterior Inference in Latent Spaces of Generative Models A. Theory and Method Details A.1. Outsourced Sampling for Diffusion Priors We describe posterior sampling, outsourced posterior inference and diffusion sampling when the prior...

    2023

  5. [2011]

    doi: 10.1201/b10905

    ISBN 9780429138508. doi: 10.1201/b10905. URL http://dx.doi.org/10. 1201/b10905. Bugallo, M. F., Elvira, V ., Martino, L., Luengo, D., Miguez, J., and Djuric, P. M. Adaptive importance sampling: The past, the present, and the future. IEEE Signal Processing Magazine, 34(4):60–79,

    doi

  6. [2015]

    This architecture is naturally well-suited for latent spaces structured as image feature maps

    for the diffusion model. This architecture is naturally well-suited for latent spaces structured as image feature maps. However, for other types of latent representations, we found that simply reshaping them into𝐻×𝑊×𝐶 feature maps and treating them as images yielded surprising...

    2024

  7. [2019]

    and Salimans, T

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

    arXiv

  8. [2021]

    Domingo-Enrich, C., Drozdzal, M., Karrer, B., and Chen, R. T. Q. Adjoint matching: Fine-tuning flow and diffusion generative models with memoryless stochastic optimal control. arXiv preprint arXiv:2409.08861,

    arXiv

  9. [2023]

    For methods where it is applicable, we also report the estimated ELBO

    and diffusion fine-tuning with RTB (Venkatraman et al., 2024). For methods where it is applicable, we also report the estimated ELBO. Model Sampler E[log𝑝(y| x)](↑) FID(↓) ELBO(↑) I-CFM Prior −5.88 84 .79 −24.04 DPS −2.22 84 .96 - RTB −4.20 90 .77 −147.69 Latent HMC −2.80 46 ....

    2024

  10. [2024]

    URL https://openreview

    ISSN 2835-8856. URL https://openreview. net/forum?id=oYIjw37pTP. Berner, J., Richter, L., Sendera, M., Rector-Brooks, J., and Malkin, N. From discrete-time policies to continuous- time diffusion samplers: Asymptotic equivalences and faster training. arXiv preprint arXiv:2501.06148,

Pith tools

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