Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Accelerated Diffusion Models via Speculative Sampling

Authors: Valentin De Bortoli, Alexandre Galashov, Arthur Gretton, Arnaud Doucet

ICML 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The proposed method achieves significant speed-ups for image generation on CIFAR10, and LSUN using pixel space diffusion models, without any loss of quality (Section 7). Furthermore, we show similar speed-ups in robotics for policy generation. In all of our experiments, we track two different types of metrics. First, we assess the quality of the output distribution obtained with the speculative sampling strategy (Wasserstein-2 in the low dimensional case, FID (Heusel et al., 2017) and IS (Salimans et al., 2016) in the image experiments and reward (Chi et al., 2023) in the robotics setting).
Researcher Affiliation Industry 1Google Deep Mind. Correspondence to: Valentin De Bortoli <EMAIL>.
Pseudocode Yes Algorithm 1 Speculative Sampling for LLM... Algorithm 2 REJECTION (p, q, X)... Algorithm 3 Speculative Sampling for DDM... Algorithm 4 REJECTION (p, q, Y )... Algorithm 5 INCORRECT REJECTION (p, q, X)... Algorithm 6 Speculative Sampling for DDM... Algorithm 7 (Temperature) REJECTION (p, q, Y )... Algorithm 8 REJECTION (p A, q A, YA)... Algorithm 9 REJECTION (p, q, Y )... Algorithm 10 REJECTION (p, q, Y )... Algorithm 11 Speculative Sampling for Unadjusted Langevin Diffusion
Open Source Code No The paper does not contain any explicit statement about providing open-source code for the described methodology, nor does it include a link to a code repository.
Open Datasets Yes The proposed method achieves significant speed-ups for image generation on CIFAR10, and LSUN using pixel space diffusion models, without any loss of quality (Section 7). Furthermore, we show similar speed-ups in robotics for policy generation. In our setting, we focus on the Push T dataset.
Dataset Splits No The paper mentions evaluating FID scores on "50k training samples" for CIFAR10 and latent CIFAR-10, and running "1000 episodes" for the Push T dataset. However, it does not specify the actual training/validation/test splits used for these datasets or for the low-dimensional GMM experiments (e.g., percentages, counts for each split).
Hardware Specification No The paper does not specify the hardware (e.g., GPU models, CPU types, or cloud instances) used for running its experiments.
Software Dependencies No The paper mentions using Python for implementation or specific libraries without providing specific version numbers (e.g., Python 3.8, PyTorch 1.9, CUDA 11.1). For example, it states: "The velocity of the diffusion model is parameterized with a sequence of MLPs. For all MLP we use the Ge LU activation function."
Experiment Setup Yes We train the model for 1M steps with the Adam optimizer and a learning rate of 10-4 and EMA decay of 0.9999. In our low-dimensional setting... The velocity of the diffusion model is parameterized with a sequence of MLPs. For all MLP we use the Ge LU activation function. ...The time embedding and the label embedding are then concatenated into a conditioning embedding. The conditioning embedding and the input xt of the velocity network are then processed independently with 3 MLP layers with output dimension (64, 64, 128). ... The batch size is set to 128. ... The channel multipliers are given by (1, 2, 2, 2). The channel size is 256. We consider a dropout rate of 0.2. The normalization layers are RMS normalization layers. For the attention layers we consider 8 heads. The number of residual blocks is 2.