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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
TADA: Improved Diffusion Sampling with Training-free Augmented DynAmics
Authors: Tianrong Chen, Huangjie Zheng, David Berthelot, Jiatao Gu, Joshua Susskind, Shuangfei Zhai
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We begin by benchmarking our method against Uni PC [34] and DPM-Solver++ [23] on both Image Net-64 and Image Net-512, using the EDM [17] and EDM2 [18] frameworks. To ensure fair and competitive evaluation, we conduct comprehensive ablations over all combinations of discretization schemes, solver variants, and solver orders available in the respective codebases, selecting the best-performing configurations for comparison. Full ablation results for the baselines are reported in the Supplementary Material. |
| Researcher Affiliation | Industry | Tianrong Chen, Huangjie Zheng, David Berthelot, Jiatao Gu, Josh Susskind, Shuangfei Zhai Apple EMAIL |
| Pseudocode | Yes | Algorithm 1 TADA sampling |
| Open Source Code | Yes | The code is available at https://github.com/apple/ml-tada. |
| Open Datasets | Yes | Specifically, we assess diffusion models such as EDM [17], EDM2 [18], in both pixel and latent spaces on the Image Net-64 and Image Net-512 datasets [8]. Additionally, we evaluate the flow matching model Stable Diffusion 3 [10] in the latent space. |
| Dataset Splits | No | The paper mentions Image Net-64 and Image Net-512 datasets, and Stable Diffusion 3, but does not explicitly provide specific training/test/validation splits (e.g., percentages, sample counts, or clear citations for specific splits). |
| Hardware Specification | No | The paper discusses computational costs but does not provide specific hardware details such as GPU/CPU models, processor types, or memory specifications used for the experiments. |
| Software Dependencies | No | The paper mentions using 'DPM-Solverv3 [36]' for baselines and refers to 'the official Diffusers implementation', but it does not specify version numbers for general software dependencies like Python, PyTorch, or other libraries. |
| Experiment Setup | Yes | For the fair comparision, for all baselines, we controlled ฯmin = 0.002 and ฯmax = 80 as suggested in the original EDM and EDM2 paper. For our method, TADA, we consistently employ the simplest multistep exponential integrator with third-order solvers, using the same polynomial discretization across all experiments. To ensure consistency, our method uses the same SNR-based discretization scheme as SD3, which is also the default across all baseline implementations. For the sake of consistency, we employ a second-order multi-step solver, regardless of CFG scale. |