Problem
Research questions and friction points this paper is trying to address.
Classifier-Free Guidance
Latent Diffusion Models
Inference-time Enhancement
Compositional Alignment
Rectified-Flow Transformers
Revisiting Classifier-Free Guidance Methods in Latent Diffusion Models
🤖 AI Summary
This study addresses ongoing debates regarding the efficacy of training-free inference enhancement methods in modern Transformers by systematically evaluating eight Classifier-Free Guidance (CFG) derivatives on open-weight Rectified-Flow Transformers and compositional alignment benchmarks. The findings reveal that no evaluated method consistently outperforms standard CFG; improvements from Adaptive Projected Guidance largely fall within error margins, while attention perturbation techniques demonstrate unstable performance. By delineating the effectiveness boundaries of training-free guidance approaches, this work establishes that standard CFG remains the most robust and cost-effective baseline for current applications. These results provide critical empirical evidence to inform future research directions in inference-time scaling and model alignment, clarifying misconceptions about newer alternatives' superiority over established guidance mechanisms.
📝 Abstract
Inference-time quality-enhancement methods are an effective and widely adopted means of improving diffusion models without expensive retraining. We study a family of training-free techniques conceptually rooted in Classifier-Free Guidance (CFG), most of which were originally proposed on older U-Net diffusion models and validated using metrics that assess image quality in isolation, without accounting for compositional alignment or semantic correspondence between the generated image and its associated text prompt. We re-evaluate eight such methods on two open-weight rectified-flow transformers under a fixed per-model protocol and three compositional-alignment benchmarks. No method consistently improves on CFG across the measured criteria. APG obtains several nominal best scores, but the corresponding gains often remain within the estimated evaluation uncertainty. Attention-perturbation methods provide isolated gains on SD3.5 Medium and more frequent degradations on FLUX.2 [klein] 4B Base, while CFG remains a competitive lower-cost baseline.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Classifier-Free Guidance
Training-free Inference
Rectified-Flow Transformers
Compositional Alignment
Benchmark Re-evaluation
🔎 Similar Papers
No similar papers found.
