aiqbench analysis
This arXiv preprint proposes a unified control framework for diffusion models that links inference-time guidance with reward-based fine-tuning. The reported gains are on ImageNet with DiT-XL/2 (FID drops from 7.58 to 4.44 at 25 steps) and on Z-Image-Turbo, but they come from the authors' own experiments and have not been independently verified. Readers working on diffusion sampling or reward alignment should check whether the method's advantage holds across other architectures and evaluators, and whether its cost compared with standard guidance is acceptable.
Variational Optimal Control for Diffusion Guidance
arXiv:2610.37146v1 Announce Type: new Abstract: We present a variational control framework that connects sampling-time guidance and reward-based model adaptation through a shared Jacobian pullback. Starting from terminal reward maximization with a path-space KL penalty, the Gibbs variational principle and Doob transform yield a correction to the generative dynamics. A local denoising approximation exposes the pullback that transports output preferences to the controlled variable. This construction gives Variational Optimal Guidance (VOG) in state space and extends to Distilled Variational Optimal Control (DVOC) by amortizing locally improved endpoints into model parameters. The resulting controls share a local variational rule: VOG acts through the denoising Jacobian, while DVOC realizes the parameter pullback by endpoint distillation. Under a linear MMSE model, VOG's exact correction lies in the data subspace. On ImageNet, VOG reduces DiT-XL/2 FID from 2.27 to 2.04 at 250 sampling steps and from 7.58 to 4.44 at 25 steps. After 20 updates on Z-Image-Turbo, DVOC improves HPSv3++ and four held-out evaluator scores over the untrained model, with the highest corresponding means among the trained methods compared. These results connect inference-time control and reward adaptation through a common geometric principle.
Abstract:We present a variational control framework that connects sampling-time guidance and reward-based model adaptation through a shared Jacobian pullback. Starting from terminal reward maximization with a path-space KL penalty, the Gibbs variational principle and Doob transform yield a correction to the generative dynamics. A local denoising approximation exposes the pullback that transports output preferences to the controlled variable. This construction gives Variational Optimal Guidance (VOG) in state space and extends to Distilled Variational Optimal Control (DVOC) by amortizing locally improved endpoints into model parameters. The resulting controls share a local variational rule: VOG acts through the denoising Jacobian, while DVOC realizes the parameter pullback by endpoint distillation. Under a linear MMSE model, VOG's exact correction lies in the data subspace. On ImageNet, VOG reduces DiT-XL/2 FID from 2.27 to 2.04 at 250 sampling steps and from 7.58 to 4.44 at 25 steps. After 20 updates on Z-Image-Turbo, DVOC improves HPSv3++ and four held-out evaluator scores over the untrained model, with the highest corresponding means among the trained methods compared. These results connect inference-time control and reward adaptation through a common geometric principle.
Machine Learning (cs.LG)
arXiv:2610.37146 [cs.LG]
(or arXiv:2610.37146v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.37146 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 276 words
Related coverage
Physics-Guided Generative AI for Property-Targeted 3D Porous Media Design
arXiv · Jul 28, 2026
The Principles of Diffusion Models
arXiv · Aug 28, 2026
Diffusion Model in Latent Space for Medical Image Segmentation Task
arXiv · Sep 9, 2026
Learning to Re-Draft: A Variational Stackelberg Game for Discrete Diffusion
arXiv · Sep 30, 2026