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..
BADiff: Bandwidth Adaptive Diffusion Model
Authors: Xi Zhang, Hanwei Zhu, Yan Zhong, Jiamang Wang, Weisi Lin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our approach significantly improves the visual fidelity of bandwidth-adapted generations compared to naive early-stopping, offering a promising solution for efficient image delivery in bandwidth-constrained environments. ... We validate the effectiveness of BADiff through extensive experiments comparing our method to strong baselines, including standard diffusion models with post-generation compression and naive early-stopping approaches. Our results demonstrate that BADiff consistently achieves superior tradeoffs among perceptual quality, computational efficiency, and bandwidth efficiency... |
| Researcher Affiliation | Collaboration | 1Nanyang Technological University 2Alibaba Group EMAIL |
| Pseudocode | Yes | Algorithm 1 BADiff Training Algorithm 2 BADiff Sampling |
| Open Source Code | Yes | Code is available at: https://github.com/xzhang9308/BADiff. |
| Open Datasets | Yes | Datasets. We train and evaluate on three standard diffusion benchmarks CIFAR-10 [23], CELEBA-HQ [18], and LSUN-CHURCH/BEDROOM [51]. |
| Dataset Splits | Yes | All splits and preprocessing follow the original DDPM protocol [13, 42]. |
| Hardware Specification | Yes | We report end-to-end sampling latency (ms/image) on CIFAR-10 (32 32) using an NVIDIA RTX-4090 GPU. ... All experiments are run on NVIDIA RTX 4090 GPUs with 24GB VRAM each. |
| Software Dependencies | Yes | We use Py Torch 2.1.0 with torch.compile enabled for maximum inference speed, and CUDA version 11.8. |
| Experiment Setup | Yes | Each model is trained for 800,000 iterations using a batch size of 64 images per GPU. We use automatic mixed precision (AMP) to accelerate training and reduce memory consumption. ... We adopt the Adam optimizer [20] with default coefficients β1=0.9, β2=0.999. The learning rate is set to 1 10 4 and kept constant throughout training. No learning rate decay or warmup is applied. Gradient clipping is used with a max norm of 1.0. ... Table 7: Key hyperparameters. |