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].

Learning few-step posterior samplers by unfolding and distillation of diffusion models

Authors: Charlesquin Kemajou Mbakam, Marcelo Pereyra, Jonathan Spence

TMLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate our proposed unfolded and distilled samplers through extensive experiments and comparisons with the state of the art, where they achieve excellent accuracy and computational efficiency... In Section 4, we report a series numerical experiments and comparisons with competing approaches from the state of the art. ... Table 1 and Table 2 below summarize the performance metrics for our proposed method... 4.3 Ablation Studies
Researcher Affiliation Academia Charlesquin Kemajou Mbakam EMAIL School of Mathematical and Computer Sciences & Maxwell Institute for Mathematical Sciences Heriot-Watt University Jonathan Spence EMAIL School of Mathematics & Maxwell Institute for Mathematical Sciences University of Edinburgh Marcelo Pereyra EMAIL School of Mathematical and Computer Sciences & Maxwell Institute for Mathematical Sciences Heriot-Watt University
Pseudocode Yes Algorithm 1 Conditional Diffusion Sampling Require: Observation y, Time-grid 0 = t0 < t1 < < t N = T, 1: Sample xt N N(0, I) ▷ Initialize reversed diffusion 2: for n = N, . . . , 1 do 3: Set ˆx0 x θ tn,K(xtn, y) using Lϑ (see Figure 2) ▷ Unfolded sample targeting p0(x0 | xtn, y) 4: Sample xtn 1 ptn 1(xtn 1 | ˆx0, xtn) ▷ Reverse DDIM step 5: end for 6: return xt0
Open Source Code Yes The code is available at https://github.com/charles-kmc/UD2M.
Open Datasets Yes Datasets and experimental conditions. We use the following two public datasets in our experiments Image Net (Russakovsky et al., 2015) and LSUN Bedroom (Yu et al., 2015) which have been used extensively in prior work related to image restoration with DMs and CMs in particular.
Dataset Splits Yes For our experiments with the Image Net dataset, we used one million images from the training set for model training, whereas for the LSUN Bedroom dataset, we used 1.2 million images for training. For computing performance metrics, we use a test set of 1500 images from Image Net and a test set of 300 images from LSUN bedroom.
Hardware Specification No We acknowledge the use of the HWU high-performance computing facility (DMOG) and associated support services in the completion of this work. (No specific hardware model numbers provided). And in Table 10, it states "GPU (GB)" but doesn't mention specific models like NVIDIA A100.
Software Dependencies No The paper discusses the implementation of models and optimizers like U-NET, Adam, and Lo RA, and refers to pre-trained models from ADM and DDPM, but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes With regards to training, we use the training objective detailed in Section 3.2 and Lo RA adaptation with rank 5, while keeping original weights θ frozen. ... To facilitate reproducibility, comprehensive details regarding batch size (bs), learning rate (lr), optimizer and specific weights for loss term are summarized in Table 9. Table 9: Model 1 refers to the ADM pre-trained on the Image Net dataset, while Model 2 denotes the DDPM pre-trained on the LSUN dataset. lr bs optimizer weight decay ωℓ2 ωGP ωGS Model 1 1e 4 2 Adam W 0.01 1 0.1 0.01 Model 2 1e 4 4 Adam W 0.01 1 0.1 0.01 RAMψ 1e 5 Adam W 0.01 -