DPM-Solver-v3: Improved Diffusion ODE Solver with Empirical Model Statistics

Authors: Kaiwen Zheng, Cheng Lu, Jianfei Chen, Jun Zhu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that DPM-Solver-v3 achieves consistently better or comparable performance in both unconditional and conditional sampling with both pixelspace and latent-space DPMs, especially in 5 10 NFEs. We achieve FIDs of 12.21 (5 NFE), 2.51 (10 NFE) on unconditional CIFAR10, and MSE of 0.55 (5 NFE, 7.5 guidance scale) on Stable Diffusion, bringing a speed-up of 15% 30% compared to previous state-of-the-art training-free methods.
Researcher Affiliation Collaboration Kaiwen Zheng 1, Cheng Lu 1, Jianfei Chen1, Jun Zhu 123 1Dept. of Comp. Sci. & Tech., Institute for AI, BNRist Center, THBI Lab 1Tsinghua-Bosch Joint ML Center, Tsinghua University, Beijing, China 2Shengshu Technology, Beijing 3Pazhou Lab (Huangpu), Guangzhou, China
Pseudocode Yes We provide the pseudocode of the local approximation and global solver in Algorithm 1 and Algorithm 2, which concisely describes how we implement DPM-Solver-v3. [...] Algorithm 3 (n + 1)-th order singlestep solver
Open Source Code Yes Code is available at https://github.com/thu-ml/DPM-Solver-v3.
Open Datasets Yes We conduct extensive experiments on diverse image datasets... [24] (CIFAR10), [55] (LSUN-Bedroom), [9] (Image Net-256), [26] (MS-COCO2014) are cited as datasets used.
Dataset Splits No The paper mentions drawing K datapoints for EMS estimation and evaluating FID/MSE on 50k or 10k samples, but does not explicitly state the train/validation/test splits of the datasets used for model training or evaluation in terms of percentages or counts.
Hardware Specification Yes The total time for EMS computing is 7h on 8 GPU cards of NVIDIA A40. ... Table 4 shows the runtime of DPM-Solver-v3 and some other solvers on a single NVIDIA A40 under different settings.
Software Dependencies No We utilize the forward-mode automatic differentiation (torch.autograd.forward_ad) provided by Py Torch [39] to compute the Jacobian-vector-products (JVPs). No specific version numbers for PyTorch or other software dependencies are provided.
Experiment Setup Yes EMS computing We estimate the EMS at N = 1200 uniform timesteps λj0, λj1, . . . , λj N by drawing K = 4096 datapoints xλ0 q0... We use start time ϵ = 10 3 (NFE 10) and ϵ = 10 4 (NFE>10), end time T = 1 and adopt the uniform log SNR timestep schedule. For DPM-Solver-v3, we use the 3rd-order predictor with the 3rd-order corrector by default.