one-step-generation

This paper proposes Seq-Drifting, a framework that learns a map from Gaussian noise to text in a single forward pass by applying drifting to sequences of continuous token representations, achieving a generative perplexity of 36.01 on LM1B with large speedups over few-step and multi-step diffusion language models.

本文证明了 Drifting Models 与 Distribution Matching Distillation (DMD/DMD2) 在优化目标上的等价性，并基于此视角提出了改进的 Multi-Bandwidth Distribution Matching Distillation (MBDMD) 方法，用于提升一步生成模型的性能。

The paper introduces Reward-Weighted Transport Distillation (RWTD), a post-training method for one-step generative models that needs only generated samples and scalar rewards, blending tilted current and reference distributions via feature-space optimal transport. It boosts the GenEval score of the SanaSprint 1.6B one-step backbone from 0.73 to 0.80 while preserving compositional capabilities.

The paper presents a method for one-step code generation by making language continuous, using diffusion models, and distilling the trajectory, with accompanying code release.

Introduces Three-Body Scattering Modeling (TBSM), a framework for one-step generative modeling that learns a transport field. Achieves competitive results on ImageNet-256 and scales to text-to-image models with up to 20B parameters.

Researchers propose Latent Thought Flows, compressing 256 text tokens into 8 continuous latents for one-step generation, achieving a better Pareto frontier than autoregressive baselines.

This paper introduces Representation Distribution Matching (RDM), a method for one-step image generation by matching feature distributions under pretrained encoders, achieving state-of-the-art results on ImageNet and enabling post-training of FLUX.2 into a one-step generator with improved performance.

Introduces Cross-Space Distillation, a method to transfer knowledge from modern high-capacity diffusion models to compact student models with different latent spaces using a lightweight latent interface called Bridge, enabling quality improvements without modifying the student backbone.

Stanford and ByteDance introduce W-Flow, a single-step generative model that uses Wasserstein gradient flows to achieve state-of-the-art one-step ImageNet 256x256 generation (1.29 FID) with 100x faster sampling than multi-step diffusion models.

One-Forcing improves one-step video generation by augmenting the DMD objective with an auxiliary GAN loss, achieving state-of-the-art performance with reduced training costs.

Introduces Discrete MeanFlow, a method for one-step generation in discrete state spaces by learning conditional transition kernels of continuous-time Markov chains, avoiding iterative denoising.

Researchers extend MeanFlow one-step image generation from class labels to flexible text inputs by integrating highly-discriminative LLM-based text encoders, enabling efficient text-conditioned synthesis with improved performance.