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..
Foresight: Adaptive Layer Reuse for Accelerated and High-Quality Text-to-Video Generation
Authors: Muhammad Adnan, Nithesh Kurella, Akhil Arunkumar, Prashant Nair
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our comprehensive experiments on NVIDIA A100 GPUs for text-to-video generation benchmarks demonstrate the effectiveness of Foresight compared to static reuse baselines on Open Sora, Latte, and Cog Video X. |
| Researcher Affiliation | Collaboration | Muhammad Adnan The University of British Columbia EMAIL Nithesh Kurella d-Matrix EMAIL Akhil Arunkumar d-Matrix EMAIL Prashant J. Nair The University of British Columbia EMAIL |
| Pseudocode | Yes | Algorithm 1 Foresight |
| Open Source Code | Yes | The source code of Foresight is available at https://github.com/STAR-Laboratory/foresight. |
| Open Datasets | Yes | For comprehensive evaluation, we assess Foresight on the VBench benchmark suite [Huang et al., 2024]... We report results on the UCF-101 [Soomro et al., 2012] and Eval Crafter [Liu et al., 2024d] prompt sets |
| Dataset Splits | Yes | For evaluation, we select the first 50 prompts from each category, totaling 550 prompts. |
| Hardware Specification | Yes | We evaluated Foresight on three popular text-to-video models... using NVIDIA A100 80GB GPUs with Flash Attention [Dao et al., 2022] enabled. |
| Software Dependencies | No | using NVIDIA A100 80GB GPUs with Flash Attention [Dao et al., 2022] enabled. |
| Experiment Setup | Yes | Open-Sora used rflow [Liu et al., 2022] sampling with 30 denoising steps and a CFG scale of 7.5. Latte and Cog Video X used DDIM [Song et al., 2020] with 50 steps, and CFG scales of 7.5 and 6.0, respectively. ... We report two Foresight configurations, varying reuse steps from N = 1 to N = 2 and computation intervals from R = 2 to R = 3, with a fixed scaling factor (γ = 0.5) balancing speed and quality. |