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..

LeMiCa: Lexicographic Minimax Path Caching for Efficient Diffusion-Based Video Generation

Authors: Huanlin Gao, Ping Chen, Fuyuan Shi, Chao Tan, Zhaoxiang Liu, Fang Zhao, Kai Wang, Shiguo Lian

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on multiple text-to-video benchmarks demonstrate that Le Mi Ca delivers dual improvements in both inference speed and generation quality. Notably, our method achieves a 2.9 speedup on the Latte model and reaches an LPIPS score of 0.05 on Open-Sora, outperforming prior caching techniques.
Researcher Affiliation Industry Data Science & Artificial Intelligence Research Institute, China Unicom1 Unicom Data Intelligence, China Unicom2 EMAIL
Pseudocode Yes A Pseudocode: Lexicographic Minimax Path Selection Below we present the full pseudocode implementation of the lexicographic minimax path selection algorithm, as introduced in Section 3.3. The algorithm leverages dynamic programming to efficiently compute an optimal caching path from source s to target t under a step budget constraint B. Algorithm 1 Lexicographic Minimax Path Selection
Open Source Code Yes Our code is available at https://github.com/Unicom AI/Le Mi Ca
Open Datasets Yes We evaluate our method on representative diffusion-based video models: Open-Sora [53], Latte [24], and Cog Video X [45]. Baselines include -Di T [2], T-GATE [49], PAB [51], and Tea Cache [20]. To construct the DAG for Global Outcome-Aware error modeling, we sample 70 prompts (10 per attribute) from T2V-Comp Bench [41], following standard practice [41, 20].
Dataset Splits Yes The DAG construction and forward inference use distinct datasets to ensure fair and robust evaluation. Sampling is repeated 10 times with different seeds, and results are averaged to reduce bias. To construct the DAG for Global Outcome-Aware error modeling, we sample 70 prompts (10 per attribute) from T2V-Comp Bench [41]... Video quality is then evaluated on 50 selected VBench prompts, with average results reported.
Hardware Specification Yes Experiments are conducted on NVIDIA H100 GPUs using Py Torch.
Software Dependencies No The paper mentions 'Py Torch' as a software dependency but does not specify a version number for it or any other key software components.
Experiment Setup Yes Table 4: Model forward steps B under different configurations. Model Forward Steps. In this work, we control the acceleration efficiency of Le Mi Ca via the Model Forward Steps B. Smaller values of B reduce the denoising time, leading to higher speed-up ratios. We consider two variants: Le Mi Ca-slow, which emphasizes visual fidelity, and Le Mi Ca-fast, which prioritizes inference efficiency. The corresponding B values for each variant across different models are listed in Table 4.