Stabilizing, Scaling & Enhancing MeanFlow for Large-scale Diffusion Distillation

• URL: http://arxiv.org/abs/2605.17834v1
• Date: Mon, 18 May 2026 04:16:15 GMT
• Title: Stabilizing, Scaling & Enhancing MeanFlow for Large-scale Diffusion Distillation
• Abstract summary: Diffusion models exhibit remarkable generative capability, but their high latency limits practical deployment.<n>MeanFlow has attracted considerable attention due to its concise formulation and remarkable performance.<n>We introduce a warm-up technique, in which the original differential solution of MeanFlow is replaced by a discrete solution.
• Score: 37.28053263183048
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Diffusion models exhibit remarkable generative capability, but their high latency limits practical deployment. Many studies have attempted to reduce sampling steps to accelerate inference. Among them, MeanFlow has attracted considerable attention due to its concise formulation and remarkable performance. Nevertheless, the instability of its optimization objective and the ''mean-seeking bias'' have limited its applicability to distill large-scale industrial models. To stabilize MeanFlow for distilling large-scale models, we first introduce a warm-up technique, in which the original differential solution of MeanFlow is replaced by a discrete solution. This design avoids training collapse caused by the MeanFlow target containing a stop-gradient term from an undertrained model. Once the model acquires a preliminary ability to fit the average velocity field, we switch the optimization objective back to the differential solution, enabling further refinement. Meanwhile, to alleviate the ''mean-seeking bias'' of MeanFlow under extremely few-step inference with complex target distributions, we incorporate trajectory distribution alignment as an auxiliary objective, encouraging the student model's trajectory distribution to align more closely with that of the teacher model. Our proposed distillation framework achieves superior performance compared to existing distillation approaches when applied to the text-to-image (T2I) model FLUX.1-dev (up to 12B parameters). Furthermore, when extended to the 80B-parameter state-of-the-art (SOTA) T2I model HunyuanImage 3.0, our method continues to demonstrate robust generalization and strong performance.

Related papers

• Curvature-Adaptive Consistency Flow Matching: Autonomous Trajectory Optimization via Reinforcement Learning [8.422673935994451]
  consistency distillation has significantly accelerated the inference of diffusion models.<n>We propose Curvature-Adaptive Consistency Flow Matching to address the limitations of static sampling.<n>Our approach achieves new state-of-the-art results on large-scale models such as FLUX and SDXL.
  arXiv  Detail & Related papers  (2026-06-21T08:49:50Z)
• One-Step Generative Modeling via Wasserstein Gradient Flows [56.266175641614986]
  W-Flow is a framework for training a generator that transforms samples from a simple reference distribution into samples from a target data distribution in a single step.<n>We set a new state of the art for one-step ImageNet 256$times$256 generation, achieving 1.29 FID, with improved mode coverage and domain transfer.
  arXiv  Detail & Related papers  (2026-05-12T08:29:44Z)
• P-Flow: Proxy-gradient Flows for Linear Inverse Problems [28.438072266272954]
  P-Flow is a framework that stabilizes the reconstruction process by leveraging a proxy gradient to update the source point.<n>Experiments across diverse restoration tasks demonstrate that P-Flow delivers competitive performance.
  arXiv  Detail & Related papers  (2026-05-08T17:07:36Z)
• InjectFlow: Weak Guides Strong via Orthogonal Injection for Flow Matching [1.974921946982281]
  Flow Matching (FM) has emerged as a leading approach for high-fidelity visual generation.<n>FM models are sensitive to dataset biases, which cause severe semantic degradation when generating out-of-distribution or minority-class samples.<n>We introduce InjectFlow, a novel, training-free method by injecting semantics during the initial velocity field.
  arXiv  Detail & Related papers  (2026-03-19T11:07:14Z)
• Joint Distillation for Fast Likelihood Evaluation and Sampling in Flow-based Models [100.28111930893188]
  Some of today's best generative models still require hundreds to thousands of neural function evaluations to compute a single likelihood.<n>We present fast flow joint distillation (F2D2), a framework that simultaneously reduces the number of NFEs required for both sampling and likelihood evaluation by two orders of magnitude.<n>F2D2 is modular, compatible with existing flow-based few-step sampling models, and requires only an additional divergence prediction head.
  arXiv  Detail & Related papers  (2025-12-02T10:48:20Z)
• Which Layer Causes Distribution Deviation? Entropy-Guided Adaptive Pruning for Diffusion and Flow Models [77.55829017952728]
  EntPruner is an entropy-guided automatic progressive pruning framework for diffusion and flow models.<n>Experiments on DiT and SiT models demonstrate the effectiveness of EntPruner, achieving up to 2.22$times$ inference speedup.
  arXiv  Detail & Related papers  (2025-11-26T07:20:48Z)
• FlowSteer: Guiding Few-Step Image Synthesis with Authentic Trajectories [82.90132015584359]
  ReFlow has theoretical consistency with flow matching but suboptimal performance in practical scenarios.<n>We propose FlowSteer, a method unlocks the potential of ReFlow-based distillation by guiding the student along teacher's authentic generation trajectories.
  arXiv  Detail & Related papers  (2025-11-24T07:13:23Z)
• Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation [2.3359837623080613]
  We show that diffusion training may be viewed as a form of generative pre-training.<n>We create a one-step generation model by fine-tuning a pre-trained model with 85% of parameters frozen.
  arXiv  Detail & Related papers  (2025-06-11T03:55:26Z)
• Mean Flows for One-step Generative Modeling [64.4997821467102]
  We propose a principled and effective framework for one-step generative modeling.<n>A well-defined identity between average and instantaneous velocities is derived and used to guide neural network training.<n>Our method, termed the MeanFlow model, is self-contained and requires no pre-training, distillation, or curriculum learning.
  arXiv  Detail & Related papers  (2025-05-19T17:59:42Z)
• Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow [65.51671121528858]
  Diffusion models have greatly improved visual generation but are hindered by slow generation speed due to the computationally intensive nature of solving generative ODEs. Rectified flow, a widely recognized solution, improves generation speed by straightening the ODE path. We propose Rectified Diffusion, which generalizes the design space and application scope of rectification to encompass the broader category of diffusion models.
  arXiv  Detail & Related papers  (2024-10-09T17:43:38Z)
• SlimFlow: Training Smaller One-Step Diffusion Models with Rectified Flow [24.213303324584906]
  We develop small, efficient one-step diffusion models based on the powerful rectified flow framework. We train a one-step diffusion model with an FID of 5.02 and 15.7M parameters, outperforming the previous state-of-the-art one-step diffusion model.
  arXiv  Detail & Related papers  (2024-07-17T16:38:45Z)
• DF2: Distribution-Free Decision-Focused Learning [30.288876294435294]
  Decision-focused learning (DFL) has emerged as a powerful approach for predict-then-optimize problems.<n>DFL faces three bottlenecks: model error, sample average approximation error, and approximation error.<n>We present DF2, the first decision-free learning method designed to mitigate these three bottlenecks.
  arXiv  Detail & Related papers  (2023-08-11T00:44:46Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.