Distribution Matching Distillation (DMD) provides an effective distribution-level correction for few-step generation, while relying on an auxiliary fake-score network to track the evolving generative distribution. Recent work combines DMD with a flow-map objective to exploit the benefits of both methods: mode-seeking and mode-covering behaviors. Since this fake-score network increases memory and update overhead, we study whether this explicit tracker can be avoided when the generator has a flow-map structure. In this work, we propose Fake-Score-network-Free DMD (FSF-DMD), a DMD-style correction for flow-map generators, replacing the auxiliary fake-score estimator with a generator-induced surrogate. We observe that this surrogate can be derived from the definition of the fake velocity under a certain perturbation assumption. Building on this insight, we derive a DMD-style objective without a fake-side network, extend it to backward simulation for flow maps, and to a self-teacher variant for training from scratch. In our ImageNet-1K $256\times 256$ experiments, FSF-DMD is competitive with the studied DMD baselines when initialized from pretrained networks and when trained from scratch.
Youngjoong Kim, Duhoe Kim, Woosung Kim, Jaesik Park
原文摘要
Consistency models have been proposed for fast generative modeling, achieving results competitive with diffusion and flow models. However, these methods exhibit instability and limited reproducibility when training from scratch, motivating subsequent work to explain and stabilize these issues. While these efforts have provided valuable insights, the explanations remain fragmented, and the theoretical relationships remain unclear. In this work, we provide a theoretical examination of consistency models by analyzing them from a flow map-based perspective. This joint analysis clarifies how training and convergence behavior can give rise to degenerate solutions. Building on these insights, we revisit self-distillation as a practical remedy for certain forms of suboptimal convergence. To avoid excessive gradient norms during its optimization, we reformulate its objective empirically, yielding improved Self-Distillation (iSD). We further extend iSD with training-time classifier-free guidance, yielding iSD-T, which improves performance and reproducibility in the studied settings. We demonstrate that iSD extends beyond image generation to diffusion-based policy learning, without reliance on pretrained diffusion models for initialization.
Compositional diffusion planners have demonstrated remarkable success on long-horizon diffusion planning by sampling short trajectories, namely chunks, and composing them into a long trajectory. However, averaging overlapping chunks can induce the mode-averaging problem, resulting in discontinuous trajectories. Several works aim to address this problem, whereas a non-negligible portion of plans remains infeasible. In this work, we propose Selective Compositional Diffuser (SCD), handling this problem by measuring reachability. To this end, we present a self-supervised learning method for the reachability classifier. This is trained on offline data and generated chunks jointly to address potential distributional mismatch, while reducing the bias inherited from the generated data. Based on this, we construct a graph whose edge costs represent whether consecutive chunks are reachable, and traverse the graph to find effective yet efficient routes. Lastly, we generate bridging trajectories and refine the composed trajectory for smooth transitions. We evaluate our method on OGBench stitching tasks. It practically reduces incompatible mode compositions and the mode-averaging problem compared to the studied baselines. Across tasks, SCD improves success rates, reaching state-of-the-art results on HumanoidMaze-Large and Giant tasks.