Research map / Probabilistic and scientific modeling / Variational and generative sampling
Generative Modeling via Drifting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Drifting Models, which evolve the pushforward distribution during training via an anti-symmetric drifting field so that a single-pass generator reaches FID 1.54 (latent) and 1.61 (pixel) on ImageNet 256x256.
Problem
Generative models usually realize the pushforward map iteratively at inference time (diffusion/flows), making generation multi-step; the paper seeks a native one-step generator trained from scratch with high quality.
Approach
Instead of iterating at inference, the generator's training iterates the pushforward: a kernel-based anti-symmetric drifting field V_{p,q} = V+ - V- (attraction to data, repulsion from generated samples) defines a frozen target x+V, and the network regresses toward it with a stop-gradient MSE loss; feature-space drifting with a self-supervised latent-MAE encoder and CFG-as-negatives support traini
Key results
- One-step latent-space generator (DiT-L/2 with SD-VAE) reaches FID 1.54 on ImageNet 256x256, outperforming prior 1-NFE methods such as iMeanFlow (1.72).
- Pixel-space one-step generator (L/16) reaches FID 1.61, competitive with multi-step pixel methods such as PixelDiT/16 (1.61) and better than StyleGAN-XL (2.30).
- On robotic control, the 1-NFE Drifting Policy matches or exceeds 100-NFE Diffusion Policy on most tasks (e.g., BlockPush Phase 1 0.56 vs 0.36).
Surprising findings
- Breaking anti-symmetry of the drifting field (e.g., 1.5x attraction) causes FID to degrade catastrophically (8.46 to 41.05), showing the balance is essential.
- Attraction-only drifting yields FID 177.14, and the method failed to work on ImageNet without a feature encoder.
Limitations
- Converse implication (V=0 implies q=p) not guaranteed in theory; identifiability only heuristic.
- Requires a pretrained or customized feature encoder; fails on ImageNet without one.
- Design choices (kernel, encoder, architecture) acknowledged as likely suboptimal.
Experimental setup
Datasets: ImageNet ImageNet (robotics: Diffusion Policy benchmarks)
Benchmarks: ImageNet 256x256 Robomimic/Diffusion Policy tasks (Lift, Can, ToolHang, PushT, BlockPush, Kitchen)
Metrics: FID IS success rate
Baselines: DiT-XL/2 SiT-XL/2 SiT-XL/2+REPA LightningDiT-XL/2 RAE+DiT^DH-XL/2 iCT-XL/2 Shortcut-XL/2 MeanFlow-XL/2 AdvFlow-XL/2 iMeanFlow-XL/2 ADM-G SiD
Builds on: mean-shift InfoNCE classifier-free guidance MoCo DiT
Closest papers by meaning
- A solvable model of learning generative diffusion: theory and insights (NeurIPS 2025 · similarity 0.88)
- Drift Flow Matching (NeurIPS 2026 · similarity 0.87)
- Mean Flows for One-step Generative Modeling (NeurIPS 2025 · similarity 0.87)
- Neural Diffusion Models (ICML 2024 · similarity 0.87)
- Towards Theoretical Understandings of Self-Consuming Generative Models (ICML 2024 · similarity 0.87)
- Kernel-Gradient Drifting Models (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.