Inference-Time Alignment of Diffusion Models with Direct Noise Optimization
read the original abstract
In this work, we focus on the alignment problem of diffusion models with a continuous reward function, which represents specific objectives for downstream tasks, such as increasing darkness or improving the aesthetics of images. The central goal of the alignment problem is to adjust the distribution learned by diffusion models such that the generated samples maximize the target reward function. We propose a novel alignment approach, named Direct Noise Optimization (DNO), that optimizes the injected noise during the sampling process of diffusion models. By design, DNO operates at inference-time, and thus is tuning-free and prompt-agnostic, with the alignment occurring in an online fashion during generation. We rigorously study the theoretical properties of DNO and also propose variants to deal with non-differentiable reward functions. Furthermore, we identify that naive implementation of DNO occasionally suffers from the out-of-distribution reward hacking problem, where optimized samples have high rewards but are no longer in the support of the pretrained distribution. To remedy this issue, we leverage classical high-dimensional statistics theory to an effective probability regularization technique. We conduct extensive experiments on several important reward functions and demonstrate that the proposed DNO approach can achieve state-of-the-art reward scores within a reasonable time budget for generation.
This paper has not been read by Pith yet.
Forward citations
Cited by 9 Pith papers
-
Flow-GRPO: Training Flow Matching Models via Online RL
Flow-GRPO is the first online RL method for flow matching models, raising GenEval accuracy from 63% to 95% and text-rendering accuracy from 59% to 92% with little reward hacking.
-
You've Got a Golden Ticket: Improving Generative Robot Policies With A Single Noise Vector
Optimizing a single constant initial noise vector for frozen generative robot policies improves success rates on 38 of 43 tasks by up to 58% relative improvement.
-
It's Never Too Late: Noise Optimization for Collapse Recovery in Trained Diffusion Models
Noise optimization during sampling recovers diversity in mode-collapsed diffusion models while preserving output fidelity.
-
Generative Refinement for Low-Budget Black-Box Optimization
SPARROW is a black-box optimization method that treats a fixed generative sampler as a structured proposal operator and applies rank-based selection over evaluated candidates to achieve low-budget optimization with as...
-
Don't Settle at the Mode! Mitigating Diversity Collapse in Pretrained Flow Models via Feature Self-Guidance
Feature self-guidance disperses internal features of flow models during batch generation and applies manifold regularization to increase output diversity while preserving condition alignment.
-
PG-MAP: Joint MAP Optimization for Inference-Time Alignment of Diffusion and Flow-Matching Models
PG-MAP formulates inference-time alignment as joint MAP optimization over conditioning c and latent z_t via forward-consistency coupling for diffusion and flow-matching models, reporting gains in PickScore and HPS metrics.
-
Do Less, Achieve More: Do We Need Every-Step Optimization for RL Fine-tuning of Diffusion Models?
AdaScope adaptively selects optimal RL intervention points during diffusion denoising by monitoring structural and semantic changes, delivering 66% higher performance at 59% lower cost than full-trajectory RL baselines.
-
Bias at the End of the Score
Reward models used as quality scorers in text-to-image generation encode demographic biases that cause reward-guided training to sexualize female subjects, reinforce stereotypes, and reduce diversity.
-
Improving Video Generation with Human Feedback
A human preference dataset and VideoReward model enable Flow-DPO and Flow-NRG to produce smoother, better-aligned videos from text prompts in flow-based generators.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.