OpenNovelty.org

FACM: Flow-Anchored Consistency Models

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
Image GenerationConsistency Model
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: Training the network exclusively on a shortcut objective leads to the catastrophic forgetting of the instantaneous velocity field that defines the flow. Our solution is to explicitly anchor the model in the underlying flow, ensuring high trajectory fidelity during training. We introduce the Flow-Anchored Consistency Model (FACM), where a Flow Matching (FM) task serves as a dynamic anchor for the primary CM shortcut objective. Key to this Flow-Anchoring approach is a novel expanded time interval strategy that unifies optimization for a single model while decoupling the two tasks to ensure stable, architecturally-agnostic training. 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 1.70 with just one step (NFE=1) on ImageNet 256x256. To address the challenge of scalability, we develop a memory-efficient Chain-JVP that resolves key incompatibilities with FSDP. This method allows us to scale FACM training on a 14B parameter model (Wan 2.2), accelerating its Text-to-Image inference from 2x40 to 2-8 steps. Our code and pretrained models will be available to the public.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces Flow-Anchored Consistency Models (FACM), which anchor consistency training in flow matching to address training instability. It resides in the 'Flow-Anchored and Flow-Integrated Consistency Models' leaf, which contains only three papers total, indicating a relatively sparse research direction within the broader taxonomy of fifty papers. This leaf sits under 'Flow-Based and Hybrid Consistency Frameworks', a branch exploring integration of flow matching with consistency objectives. The small number of sibling papers suggests this specific approach of explicitly anchoring consistency models to flow dynamics represents an emerging rather than crowded research area.

The taxonomy reveals that FACM's leaf is part of a larger branch containing five other leaves addressing flow-based methods, unified frameworks, and trajectory distribution approaches. Neighboring categories include 'Multistep and Trajectory-Based Consistency Approaches' (four papers) and 'Training Efficiency and Stability Improvements' (three papers), both under the core theory branch. The scope note for FACM's leaf explicitly includes 'approaches anchoring consistency training in flow matching or combining flow and consistency objectives', distinguishing it from pure consistency models (which belong in core theory) and pure diffusion distillation (which belongs in distillation categories). This positioning suggests FACM bridges flow matching and consistency modeling in a way that differs from purely theoretical analysis or distillation-only methods.

Among seventeen candidates examined across three contributions, the analysis found two refutable pairs. The core FACM training strategy examined ten candidates with one appearing to provide overlapping prior work, while nine were non-refutable or unclear. The expanded time interval strategy examined three candidates with none refutable, suggesting stronger novelty for this component. The memory-efficient Chain-JVP examined four candidates with one refutable. These statistics indicate that within the limited search scope, most contributions show substantial differentiation from examined prior work, though the core training strategy has at least one candidate with apparent overlap. The small candidate pool (seventeen total) means these findings reflect top semantic matches rather than exhaustive coverage.

Based on the limited search of seventeen candidates, the work appears to occupy a relatively novel position within flow-anchored consistency modeling, particularly for the time interval decoupling strategy. The taxonomy structure confirms this is an emerging rather than saturated research direction. However, the analysis acknowledges its scope limitations: findings derive from top-K semantic search plus citation expansion, not comprehensive field coverage. The presence of one or two refutable candidates per contribution suggests some conceptual overlap exists within the examined literature, though the majority of candidates did not clearly refute the contributions.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
17
Contribution Candidate Papers Compared
2
Refutable Paper

Research Landscape Overview

Core task: Few-step generative modeling with consistency models. The field has evolved around the central challenge of generating high-quality samples in very few inference steps, departing from the many-step requirements of traditional diffusion models. The taxonomy reveals several major branches: foundational consistency model theory and training methods establish the core mathematical framework (Consistency Models[20], Improved Training Techniques[16]); flow-based and hybrid consistency frameworks explore connections to continuous normalizing flows and rectified flows (Simplifying Continuous-Time[3], Flow Map Matching[39]); distillation and acceleration techniques focus on compressing pretrained diffusion models into faster samplers (Multistep Consistency[1], Latent Consistency[5]); latent space consistency models adapt these ideas to compressed representations for efficiency (Improved Latent Training[14], Hyper-sd[15]); and domain-specific applications demonstrate practical deployment across audio, video, motion, and other modalities (AudioLCM[13], Swiftvideo[22], MotionPCM[10]). Additional branches address inverse problems, geometric constraints, reinforcement learning integration, adversarial training, few-shot adaptation, and sampling optimization, reflecting the breadth of research directions. Within the flow-based and hybrid consistency frameworks, a particularly active line of work investigates how to anchor consistency models to flow trajectories rather than diffusion processes, aiming to leverage the straighter paths and simpler dynamics of rectified or optimal transport flows. Flow-Anchored Consistency[0] sits squarely in this cluster, proposing methods that integrate flow matching with consistency training to achieve robust few-step generation. Nearby works such as Flow-Anchored[9] and Flow Map Matching[39] explore similar themes of coupling flow-based ODE solvers with consistency objectives, while Simplifying Continuous-Time[3] examines theoretical simplifications in continuous-time formulations. The main trade-offs in this area revolve around balancing the computational cost of flow simulation during training against the quality and diversity of few-step samples at inference, with Flow-Anchored Consistency[0] emphasizing tighter integration between flow dynamics and consistency constraints compared to more distillation-focused approaches like Latent Consistency[5] or adversarial hybrids such as Consistency-GAN[8].

Claimed Contributions

Flow-Anchored Consistency Model (FACM) training strategy

The authors propose a training strategy that combines two objectives: a Flow Matching loss that anchors the model to the instantaneous velocity field for stability, and a Consistency Model loss that learns efficient one-step generation. This approach addresses training instability by preventing catastrophic forgetting of the underlying flow.

10 retrieved papers
Can Refute
Expanded time interval strategy for task decoupling

The authors introduce a method that maps the time domain to an expanded interval [0,2], using t for the consistency task and 2-t for the flow matching task. This strategy decouples the two learning objectives into distinct domains while maintaining a continuous target, enabling stable training without architectural modifications.

3 retrieved papers
Memory-efficient Chain-JVP for scalable training

The authors develop a Chain-JVP implementation that computes Jacobian-vector products sequentially by module rather than materializing all parameters at once. This resolves memory bottlenecks with Fully Sharded Data Parallel training, enabling scalability to models with over 10 billion parameters.

4 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Flow-Anchored Consistency Model (FACM) training strategy

The authors propose a training strategy that combines two objectives: a Flow Matching loss that anchors the model to the instantaneous velocity field for stability, and a Consistency Model loss that learns efficient one-step generation. This approach addresses training instability by preventing catastrophic forgetting of the underlying flow.

Contribution

Expanded time interval strategy for task decoupling

The authors introduce a method that maps the time domain to an expanded interval [0,2], using t for the consistency task and 2-t for the flow matching task. This strategy decouples the two learning objectives into distinct domains while maintaining a continuous target, enabling stable training without architectural modifications.

Contribution

Memory-efficient Chain-JVP for scalable training

The authors develop a Chain-JVP implementation that computes Jacobian-vector products sequentially by module rather than materializing all parameters at once. This resolves memory bottlenecks with Fully Sharded Data Parallel training, enabling scalability to models with over 10 billion parameters.

Table of Contents