Flow-Anchored Consistency Models

2025-07-04

AI Summary

Continuous-time consistency models (CMs) suffer from training instability in few-step generation, primarily due to overreliance on shortcut paths at the expense of physically grounded instantaneous velocity fields. To address this, we propose Flow-Anchoring—a training strategy that explicitly incorporates a Flow Matching objective as an anchor, enforcing faithful modeling of the underlying probability flow velocity field without architectural modifications and maintaining compatibility with mainstream architectures. Our method jointly optimizes the CM’s shortcut-path objective and the Flow Matching loss, augmented by knowledge distillation from a pretrained LightningDiT model. On ImageNet 256×256, our approach achieves an FID of 1.32 in two steps and 1.76 in a single step—substantially outperforming existing few-step CMs. To our knowledge, this is the first work to achieve both stability and high-fidelity image synthesis in one- and two-step generation.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Continuous-time Consistency Models (CMs) promise efficient few-step generation but face significant challenges with training instability. We argue this instability stems from a fundamental conflict: by training a network to learn only a shortcut across a probability flow, the model loses its grasp on the instantaneous velocity field that defines the flow. Our solution is to explicitly anchor the model in the underlying flow during training. We introduce the Flow-Anchored Consistency Model (FACM), a simple but effective training strategy that uses a Flow Matching (FM) task as an anchor for the primary CM shortcut objective. This Flow-Anchoring approach requires no architectural modifications and is broadly compatible with standard model architectures. By distilling a pre-trained LightningDiT model, our method achieves a state-of-the-art FID of 1.32 with two steps (NFE=2) and 1.76 with just one step (NFE=1) on ImageNet 256x256, significantly outperforming previous methods. This provides a general and effective recipe for building high-performance, few-step generative models. Our code and pretrained models: https://github.com/ali-vilab/FACM.

Problem

• Training instability in Continuous-time Consistency Models
• Conflict between shortcut learning and velocity field grasp
• Need for high-performance few-step generative models

Innovation

• Flow-Anchored Consistency Model for stability
• Flow Matching task anchors CM shortcut
• No architectural changes, compatible with standards