Data Warmup: Complexity-Aware Curricula for Efficient Diffusion Training

One-step diffusion distillation achieves strong performance on small datasets but remains insufficient for high-resolution or complex text-to-image settings. Extending to multistep distillation often requires training data, suffers from early-step degradation, inconsistent outputs across sampling budgets, and incurs high computational cost. We identify that these limitations stem from simultaneous training, where a single adapted network approximates multiple score functions of different student distributions at once, causing biased distribution matching and unstable optimization. Therefore, we propose Progressive Multistep Data-free Distillation (PMDD), which assigns a dedicated adapted network per student distribution while avoiding the memory overhead of maintaining multiple adapted networks. Combined with a stage-wise unforget loss to preserve earlier-step generations, PMDD reduces optimization bias in the student, provides faster convergence, and a single unified model that supports multiple inference budgets. Extensive experiments show that PMDD matches or outperforms data-free few-step baselines at every multistep budget, reaches FID 1.99 on ImageNet 64x64 in two steps, below its 512-NFE teacher (2.32), and achieves FID 7.76 on zero-shot COCO 512x512, while requiring 4-5x less training compute than data-free DMDv2 to support three inference budgets.

Diffusion models achieve impressive image-generation quality but remain expensive at inference time. Diffusion distillation reduces sampling steps, yet many distilled models, including SDXL-Lightning and distribution matching distillation methods, suffer from degraded Fréchet Inception Distance (FID). We analyze this phenomenon through a PAC-style generalization bound. Our analysis suggests that aggressive early-step redirection of the velocity field makes the distillation target harder to learn, enlarging the train-test gap. As a result, early-step output distributions differ between training and inference, causing distribution mismatch in the intermediate noisy latent used as next-step inputs. We empirically validate this mechanism by showing reduced diversity in both intermediate features and final outputs. To address this issue, we propose EMPURPLE, a simple training-free method that recycles intermediate latents sampled from the original model. EMPURPLE is model-agnostic and improves FID by 7% to 20% across DMD2, Hyper-SD, FlashSD, and SDXL-Lightning. The repo is: https://github.com/TheLovesOfLadyPurple/Empurple-Training-Free-Algorithm-To-enhance-Diversity-of-The-Diffusion-Distillation-Model

Distributional Diffusion Models (DDMs) replace the standard mean-prediction denoiser with a distributional denoiser trained via a scoring rule objective, learning a stochastic approximation to $p(x_1 \mid x_t)$ rather than its conditional mean. However, scaling DDMs to modern image-generation settings faces two obstacles: (i) multi-particle training incurs overhead that scales with the number of particles, (ii) DDMs use globally fixed scoring rule hyperparameters, forcing a single trade-off across sampling budgets. We mitigate these limitations by deferring particle expansion to late transformer layers, and the hyperparameter trade-off by introducing time-dependent scoring rule schedules informed by the dynamical regimes of~\citet{Biroli2024}. Combined with a DiT-based latent setup, these changes make DDM training practical on class-conditional ImageNet-$256^2$, achieving 4.48 FID at 4 steps and 2.38 at 50 steps with DiT-XL/2, from a single model trained from scratch in one stage, without a teacher, self-distillation or JVPs. The result is a stochastic few-step generator whose FID does not degrade as the sampling budget grows from 4 to 50 NFE, and the same recipe transfers to text-to-image generation. Code and pre-trained models available at https://github.com/CompVis/iDDM.