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..
ViewPoint: Panoramic Video Generation with Pretrained Diffusion Models
Authors: Zixun Fang, Kai Zhu, Zhiheng Liu, Yu Liu, Wei Zhai, Yang Cao, Zheng-Jun Zha
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our method can synthesize highly dynamic and spatially consistent panoramic videos, achieving state-of-the-art performance and surpassing previous methods. 4 Experiments |
| Researcher Affiliation | Collaboration | 1 Mo E Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China 2 Tong Yi Lab 3 HKU |
| Pseudocode | No | The paper describes the method using mathematical formulas and descriptive text in sections like "3.2 View Point Map", "3.3 Pano-Perspective Attention", and "3.4 Overlapping Fusion", but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | Justification: We will release the codes and weights, but not now as we are still in the process of organizing them. |
| Open Datasets | Yes | Our model is trained on 4 panorama datasets, including one image dataset, Flickr360 [3], and three video datasets, WEB360 [39], ODV360 [3], and 360+x [4]. |
| Dataset Splits | No | All datasets are resized to a resolution of 512 1024, and during training, video data is divided into clips of 49 frames each. We evaluate our approach and previous methods on the ODV360 [3] dataset across five dimensions: "subject consistency", "imaging quality", "motion smoothness" and "dynamic degree". The paper does not explicitly specify the training/validation/test splits used for the datasets. |
| Hardware Specification | Yes | The training process is executed on 8 NVIDIA A100 GPUs, using a batch size of 1 and a learning rate of 1e 4. inference-time latency on a single NVIDIA A800(80GB). |
| Software Dependencies | No | We first inflate the patch embedding layer of the powerful video generation model, Wan2.1 [35], from 16 to 33 channels to accommodate the input data, and then fine-tune the entire model. ... therefore, we use Qwen-VL [2] to annotate the remaining three datasets... The paper mentions specific models used (Wan2.1, Qwen-VL) but does not provide specific version numbers for software libraries or frameworks like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | The training process is executed on 8 NVIDIA A100 GPUs, using a batch size of 1 and a learning rate of 1e 4. We employ a joint image-video training strategy, treating images as videos with only one frame. |