Generative Modeling via Drifting

2026/02/04 by Mingyang Deng, He Li, Tianhong Li +2 · 8 voices · 1 citation
#cs.LG #cs.CV

paper · pdf

Abstract

Generative modeling can be formulated as learning a mapping f such that its pushforward distribution matches the data distribution. The pushforward behavior can be carried out iteratively at inference time, for example in diffusion and flow-based models. In this paper, we propose a new paradigm called Drifting Models, which evolve the pushforward distribution during training and naturally admit one-step inference. We introduce a drifting field that governs the sample movement and achieves equilibrium when the distributions match. This leads to a training objective that allows the neural network optimizer to evolve the distribution. In experiments, our one-step generator achieves state-of-the-art results on ImageNet at 256 x 256 resolution, with an FID of 1.54 in latent space and 1.61 in pixel space. We hope that our work opens up new opportunities for high-quality one-step generation.

Citations

• Improved Mean Flows: On the Challenges of Fastforward Generative Models
• Adversarial Flow Models
• PixelDiT: Pixel Diffusion Transformers for Image Generation
• There is No VAE: End-to-End Pixel-Space Generative Modeling via Self-Supervised Pre-training
• Contrastive Flow Matching
• Mean Flows for One-step Generative Modeling
• Inductive Moment Matching
• Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models
• Normalizing Flows are Capable Generative Models
• Simpler Diffusion (SiD2): 1.5 FID on ImageNet512 with pixel-space diffusion
• One Step Diffusion via Shortcut Models
• Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think
• Flow map matching with stochastic interpolants: A mathematical framework for consistency models
• Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation
• One-step Diffusion with Distribution Matching Distillation
• Improved Techniques for Training Consistency Models
• Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models
• Stochastic Interpolants: A Unifying Framework for Flows and Diffusions
• Scaling up GANs for Text-to-Image Synthesis
• Diffusion Policy: Visuomotor Policy Learning via Action Diffusion
• Diffusion policy: Visuomotor policy learning via action diffusion
• Understanding Diffusion Objectives as the ELBO with Simple Data Augmentation
• Simple diffusion: End-to-end diffusion for high resolution images
• ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders
• Scalable Diffusion Models with Transformers
• Flow Matching for Generative Modeling
• Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow
• Classifier-Free Diffusion Guidance
• StyleGAN-XL: Scaling StyleGAN to Large Diverse Datasets
• High-Resolution Image Synthesis with Latent Diffusion Models
• High-Resolution Image Synthesis with Latent Diffusion Models
• Masked Autoencoders Are Scalable Vision Learners
• Diffusion Models Beat GANs on Image Synthesis
• RoFormer: Enhanced Transformer with Rotary Position Embedding
• Learning Transferable Visual Models From Natural Language Supervision
• Taming Transformers for High-Resolution Image Synthesis
• Score-Based Generative Modeling through Stochastic Differential Equations
• Exploring Simple Siamese Representation Learning
• Exploring Simple Siamese Representation Learning
• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• Query-Key Normalization for Transformers
• Denoising Diffusion Probabilistic Models
• Improved Baselines with Momentum Contrastive Learning
• A Simple Framework for Contrastive Learning of Visual Representations
• GLU Variants Improve Transformer
• Root Mean Square Layer Normalization
• Large Scale GAN Training for High Fidelity Natural Image Synthesis
• Representation Learning with Contrastive Predictive Coding
• Group Normalization
• The Unreasonable Effectiveness of Deep Features as a Perceptual Metric
• Coulomb GANs: Provably Optimal Nash Equilibria via Potential Fields
• Density estimation using Real NVP
• Deep Residual Learning for Image Recognition
• Variational Inference with Normalizing Flows
• Training generative neural networks via Maximum Mean Discrepancy optimization
• Deep Unsupervised Learning using Nonequilibrium Thermodynamics
• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• Generative Moment Matching Networks

Cited by

• Amortized Moment Matching for Visual Generation

Discussions

• New He -- potentially a BFD arxiv.org/abs/2602.047... [bsky, 5 points, 0 comments]
• imo the original paper (arxiv.org/abs/2602.04770) is well written, but there are already some implementations/blog posts about it (github.com/Algomancer/M...) [bsky, 3 points, 1 comments]
• Generative Modeling via Drifting [hn, 2 points, 1 comments]
• Sergio Charles @eigentopology 1/10 "Thread: Drifting Models - The Math Behind Training-Time Evolution 🧵 New paradigm for generative modeling from Kaiming He's team. Let's break down the mathematics. [bsky, 1 points, 1 comments]
• It is a comparison between - MMD: a kernel (2017) arxiv.org/abs/1801.01401 - Drifting: a stochastic kernel (2026) arxiv.org/abs/2602.04770 [bsky, 1 points, 2 comments]
• I might have missed it, but Drifting Model seems to be getting a lot of attention. Some say that this is the end of the diffusion model, as it can generate high-quality images in one step. It may also [bsky, 0 points, 0 comments]
• new SotA diffusion model usurper just dropped arxiv.org/abs/2602.047... [bsky, 0 points, 1 comments]
• Neat optimization arxiv.org/abs/2602.047... [bsky, 0 points, 0 comments]

Related

• DriftXpress: Faster Drifting Models via Projected RKHS Fields
• Towards Practical Plug-and-Play Diffusion Models
• Conditional Image Generation with Pretrained Generative Model
• Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation
• Neural Diffusion Models