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..

NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering

Authors: Zhihao Huang, Xi Qiu, Yukuo Ma, Yifu Zhou, Junjie Chen, Hongyuan Zhang, Chi Zhang, Xuelong Li

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on the Image Net-256 benchmark validate NFIG s effectiveness, demonstrating superior performance (FID: 2.81) and a notable 1.25 speedup compared to the strong baseline VAR-d20.
Researcher Affiliation Collaboration Zhihao Huang1,2 Xi Qiu2 Yukuo Ma2,4 Yifu Zhou1,2 Junjie Chen2 Hongyuan Zhang2,3, Chi Zhang2, Xuelong Li2, 1 Northwest Polytechnical University 2 Tele AI, China Telecom 3 University of Hong Kong 4 Beihang University
Pseudocode No The paper describes its methodology through detailed textual explanations and mathematical formulations, supported by architectural diagrams in Figure 2. It does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes The source code is available at https://github.com/Pride-Huang/NFIG.
Open Datasets Yes For the purpose of our experiments, we use the ILSVRC 2012 subset of Image Net [33]
Dataset Splits Yes This subset focuses on 1k object categories, with each category having approximately 1.2k training images, 50 validation images, and 100 test images.
Hardware Specification Yes Our model was implemented using the Py Torch framework [44] and trained on NVIDIA H100 graphics cards.
Software Dependencies No Our model was implemented using the Py Torch framework [44] and trained on NVIDIA H100 graphics cards. The paper mentions PyTorch but does not specify a version number or any other software dependencies with their respective versions.
Experiment Setup Yes Optimization was performed using the Adam optimizer, setting the learning rate to 8 10 5 and the batch size to 768. Training of the model ran for 350 epochs on the Image Net dataset. For inference, we configured CFG to 4.5 and top_k to 990.