Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion
This addresses the problem of improving sequence generation for applications like video and planning, offering a novel hybrid approach that is incremental in integrating existing methods.
The paper tackles sequence generative modeling by introducing Diffusion Forcing, a training paradigm that combines next-token prediction with full-sequence diffusion, enabling variable-length generation and guided sampling. The result includes capabilities like rolling out sequences beyond training horizons and achieving performance gains in decision-making tasks, with a proof of optimizing a variational lower bound on subsequence likelihoods.
This paper presents Diffusion Forcing, a new training paradigm where a diffusion model is trained to denoise a set of tokens with independent per-token noise levels. We apply Diffusion Forcing to sequence generative modeling by training a causal next-token prediction model to generate one or several future tokens without fully diffusing past ones. Our approach is shown to combine the strengths of next-token prediction models, such as variable-length generation, with the strengths of full-sequence diffusion models, such as the ability to guide sampling to desirable trajectories. Our method offers a range of additional capabilities, such as (1) rolling-out sequences of continuous tokens, such as video, with lengths past the training horizon, where baselines diverge and (2) new sampling and guiding schemes that uniquely profit from Diffusion Forcing's variable-horizon and causal architecture, and which lead to marked performance gains in decision-making and planning tasks. In addition to its empirical success, our method is proven to optimize a variational lower bound on the likelihoods of all subsequences of tokens drawn from the true joint distribution. Project website: https://boyuan.space/diffusion-forcing