Generative Modeling via Drifting

TL;DR

Drifting Models are a new one-step generative paradigm from Kaiming He’s group that evolves the pushforward distribution of a learned map during training, rather than learning to invert a fixed noising process as diffusion/flow models do. A “drifting field” governs how samples move toward the data distribution and reaches equilibrium when the two distributions match; the network is optimized to follow this field, which naturally yields one-step inference. On ImageNet 256×256 the one-step generator achieves FID 1.54 in latent space and 1.61 in pixel space — competitive with (and at one-NFE, better than) multi-step diffusion baselines.

Key claims

Generative modeling is reformulated as learning a single map f whose pushforward of a base distribution matches the data distribution — sidestepping the iterative pushforward of diffusion/flow models [Abstract].
A drifting field governs sample movement during training; equilibrium of the field corresponds to distribution match, giving a principled objective for training a one-step generator from scratch [Abstract].
The training objective lets the optimizer itself evolve the pushforward distribution, rather than fixing it via a pre-specified noising process [Abstract].
One-step (1-NFE) generation on ImageNet 256×256 reaches FID 1.54 in latent space and 1.61 in pixel space, at CFG scale 1.0 [Project page, Uncurated samples].
The pixel-space result demonstrates the recipe works without a VAE [Project page].

Method

A Drifting Model parameterizes a map f and trains it so that its pushforward of a simple base distribution evolves over training steps toward the data distribution. The mechanism is a drifting field — a vector field on sample space whose stationary distribution is the data distribution and that vanishes (equilibrium) when the pushforward matches the target. The neural network optimizer is treated as the agent that moves the pushforward along this field, making the training dynamics themselves the “sampler.” Three initialization regimes are studied (Middle / Far-Away / Collapsed), suggesting the choice of where the pushforward distribution starts in sample space is a design knob distinct from objective choice.

Inference is a single forward pass: draw a base sample and apply f. The paper releases a JAX codebase with pretrained weights and an interactive 2D notebook demonstrating the algorithm on toy distributions.

Results

ImageNet 256×256, class-conditional, 1 NFE, CFG 1.0:
- FID 1.54 in latent space.
- FID 1.61 in pixel space.
Pixel-space results are particularly notable given that latent-space one-step methods (SoFlow, MeanFlow, Consistency Models) typically rely on a VAE; Drifting matches them in pixel space without one.
Headline framing in the project page is “state-of-the-art” for one-step ImageNet 256² [Project page].

(No public arxiv body at filing time — the project page is the authoritative source for the headline numbers; method details below trace the project-page summary plus the bibtex entry pointing at arXiv:2602.04770.)

Why it’s interesting

This is a different reduction of the generative-modeling problem from diffusion or flow matching. Diffusion/flow fix a noising process and learn its inverse; consistency / MeanFlow / SoFlow (see SoFlow: Solution Flow Models for One-Step Generative Modeling) fix a flow ODE and learn its solution operator. Drifting fixes neither — it lets the pushforward distribution evolve under a drifting field whose equilibrium is the data, and treats SGD itself as the sampler. For Luma that’s interesting both as an architectural-and-loss research direction and as a near-term reference point for one-step image (and possibly video) generation: a Kaiming-lab result claiming FID 1.54 at 1 NFE on ImageNet-256 is a high-water mark the few-step literature will have to engage with. The pixel-space FID 1.61 result is also directly relevant to the Pixel-space diffusion thread — a one-step pixel generator competitive with multi-step latent diffusion would change the practical cost calculus for high-resolution generation.