SEEDS: Exponential SDE Solvers for Fast High-Quality Sampling from Diffusion Models
Authors: Martin Gonzalez, Nelson Fernandez Pinto, Thuy Tran, elies Gherbi, Hatem Hajri, Nader Masmoudi
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our approach on several image generation benchmarks, showing that SEEDS outperform or are competitive with previous SDE solvers. Contrary to the latter, SEEDS are derivative and training free, and we fully prove strong convergence guarantees for them. |
| Researcher Affiliation | Collaboration | Martin Gonzalez IRT System X Nelson Fernandez Air Liquide Thuy Tran IRT System X Elies Gherbi IRT System X Hatem Hajri IRT System X & Safran Nader Masmoudi New York University |
| Pseudocode | Yes | Algorithm 1 Iterative procedure Algorithm 2 SEEDS-1(Fθ, xs, s, t) Algorithm 3 SEEDS-2(Fθ, xs, s, t) Algorithm 4 SEEDS-3(Fθ, xs, s, t) with 0 < r1 < r2 < 1 |
| Open Source Code | Yes | Our code is publicly available in this link. |
| Open Datasets | Yes | We compare SEEDS with several previous methods on discretely and continuously pre-trained DPMs. We report results of many available sources, such as DDPM [12], Analytic DDPM [2], PNDM [20], GGF [15], DDIM [33], g DDIM [41], DEIS [40] and DPM-Solver [22]. For CIFAR, Celeb A and FFHQ, we use baseline pre-trained models [34, 16]. For Image Net, we use the optimized pre-trained model from [16]. |
| Dataset Splits | No | The paper mentions using pre-trained models and evaluating on benchmarks. It describes training configurations for their experiments but does not explicitly provide details about specific training/validation/test dataset splits used for reproduction, beyond relying on standard benchmark practices. |
| Hardware Specification | Yes | During the experiments, we used three Linux-based servers with 60GB memory each and 4 GPUs NVIDIA V100 32GB, 4 GPUs NVIDIA V100 16GB, and 2 GPUs NVIDIA V100 32GB, respectively. Table 12 shows the detail of the configuration utilized for each experiment. |
| Software Dependencies | No | The paper mentions using common machine learning frameworks implicitly (e.g., PyTorch through context of DPMs) and references other solvers like DPM-Solver and EDM, but does not specify exact version numbers for any software dependencies. |
| Experiment Setup | Yes | For continuously trained models, SEEDS use the EDM discretization [16, Eq. 5] with default parameters and does not use the last-step iteration trick, meaning that the last iteration of SEEDS is trivial. For discretely trained models, SEEDS use the linear step schedule in the interval [λt0, λt N ] interval following [22, Sec. 3.3, 3.4]. All the reported SEEDS results were obtained using the noise prediction mode. In all experiments we fix the parameter c2 = 0.5 for SEEDS-2 and c2 = 1/3 for SEEDS-3. |