Efficient Inference-time Scaling For Flow Models: Enhancing Sampling Diversity And Compute Allocation

Recent advancements successful AI scaling laws person shifted from simply expanding exemplary size and training information to optimizing inference-time computation. This approach, exemplified by models for illustration OpenAI o1 and DeepSeek R1, enhances exemplary capacity by leveraging further computational resources during inference. Test-time fund forcing has emerged arsenic an businesslike method successful LLMs, enabling improved capacity pinch minimal token sampling. Similarly, inference-time scaling has gained traction successful diffusion models, peculiarly successful reward-based sampling, wherever iterative refinement helps make outputs that amended align pinch personification preferences. This method is important for text-to-image generation, wherever naïve sampling often fails to afloat seizure intricate specifications, specified arsenic entity relationships and logical constraints.

Inference-time scaling methods for diffusion models tin beryllium broadly categorized into fine-tuning-based and particle-sampling approaches. Fine-tuning improves exemplary alignment pinch circumstantial tasks but requires retraining for each usage case, limiting scalability. In contrast, particle sampling—used successful techniques for illustration SVDD and CoDe—selects high-reward samples iteratively during denoising, importantly improving output quality. While these methods person been effective for diffusion models, their exertion to travel models has been constricted owed to nan deterministic quality of their procreation process. Recent work, including SoP, has introduced stochasticity to travel models, enabling particle sampling-based inference-time scaling. This study expands connected specified efforts by modifying nan reverse kernel, further enhancing sampling diverseness and effectiveness successful flow-based generative models.

Researchers from KAIST propose an inference-time scaling method for pretrained travel models, addressing their limitations successful particle sampling owed to a deterministic generative process. They present 3 cardinal innovations: (1) SDE-based procreation to alteration stochastic sampling, (2) VP interpolant conversion to heighten sample diversity, and (3) Rollover Budget Forcing (RBF) for adaptive computational assets allocation. Experimental results show that these techniques amended reward alignment successful tasks for illustration compositional text-to-image generation. Their attack outperforms anterior methods, demonstrating nan advantages of inference-time scaling successful travel models, peculiarly erstwhile mixed pinch gradient-based techniques for differentiable rewards for illustration artistic image generation.

Inference-time reward alignment intends to make high-reward samples from a pretrained travel exemplary without retraining. The nonsubjective is to maximize nan expected reward while minimizing deviation from nan original information distribution utilizing KL regularization. Since nonstop sampling is challenging, particle sampling techniques, commonly utilized successful diffusion models, are adapted. However, travel models trust connected deterministic sampling, limiting exploration. To reside this, inference-time stochastic sampling is introduced by converting deterministic processes into stochastic ones. Additionally, interpolant conversion enhances hunt abstraction by aligning travel exemplary sampling pinch diffusion models. A move compute allocation strategy further optimizes ratio during inference-time scaling.

The study presents experimental results connected particle sampling methods for inference-time reward alignment. The study focuses connected compositional text-to-image and quantity-aware image generation, utilizing FLUX arsenic nan pretrained travel model. Metrics specified arsenic VQAScore and RSS measure alignment and accuracy. Results bespeak that inference-time stochastic sampling improves efficiency, pinch interpolant conversion further enhancing performance. Flow-based particle sampling yields high-reward outputs compared to diffusion models without compromising image quality. The projected RBF method optimizes fund allocation, achieving nan champion reward alignment and accuracy results. Qualitative and quantitative findings corroborate its effectiveness successful generating precise, high-quality images.

In conclusion, nan study introduces an inference-time scaling method for travel models, incorporating 3 cardinal innovations: (1) ODE-to-SDE conversion for enabling particle sampling, (2) linear-to-VP interpolant conversion to heighten diverseness and hunt efficiency, and (3) RBF for adaptive compute allocation. While diffusion models use from stochastic sampling during denoising, travel models require tailored approaches owed to their deterministic nature. The projected VP-SDE-based procreation efficaciously integrates particle sampling, and RBF optimizes compute usage. Experimental results show that this method surpasses existing inference-time scaling techniques, improving capacity while maintaining high-quality outputs successful flow-based image and video procreation models.

Check out the Paper. All in installments for this investigation goes to nan researchers of this project. Also, feel free to travel america on Twitter and don’t hide to subordinate our 85k+ ML SubReddit.