SyncTweedies: A General Generative Framework Based on Synchronized Diffusions

Authors: Jaihoon Kim, Juil Koo, Kyeongmin Yeo, Minhyuk Sung

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We verify that Sync Tweedies offers the broadest applicability to diverse applications and superior performance compared to the previous state-of-the-art for each application. Our project page is at https://synctweedies.github.io. ... We present quantitative and qualitative results in Table 3 and Figure 4, respectively.
Researcher Affiliation Academia Jaihoon Kim Juil Koo Kyeongmin Yeo Minhyuk Sung KAIST {jh27kim,63days,aaaaa,mhsung}@kaist.ac.kr
Pseudocode No No explicit pseudocode or algorithm blocks labeled 'Algorithm' or 'Pseudocode' were found in the paper.
Open Source Code Yes Our project page is at https://synctweedies.github.io.
Open Datasets Yes We generate 360 panorama images from 360 depth maps obtained from the 360Mono Depth [43] dataset. ... We use 429 pairs of meshes and prompts used in TEXTure [44] and Text2Tex [10]. ... We use 3D Gaussian splats trained with multi-view images from the Synthetic Ne RF dataset [39]... MVDiffusion [56], which is finetuned using 3D scenes in the Scan Net [13] dataset.
Dataset Splits No The paper describes splits for generation/evaluation (e.g., '50 views for texture generation and evaluate the results from 150 unseen views') but does not specify explicit training, validation, and test dataset percentages or counts in a general sense for their method's training.
Hardware Specification Yes We use the NVIDIA RTX A6000 for the runtime comparisons. ... Additionally, we present qualitative results of arbitrary-sized image generation using Intel Gaudi-v2 in Figure 11, along with a comparison of computation times between Intel Gaudi-v2 and NVIDIA A6000 in Figure 12.
Software Dependencies No The paper mentions specific models like 'Depth-conditioned Control Net [64]', 'Stable Diffusion [47]', and 'Deep Floyd [14]', but does not list specific software dependencies with version numbers (e.g., PyTorch 1.9, CUDA 11.1).
Experiment Setup Yes For all diffusion synchronization processes, we use a fully deterministic DDIM [53] sampling with 30 steps, unless specified otherwise. ... We set the CFG weight [22] to 30 and t to 0.8. For other settings, we follow the 3D mesh texture generation experiment presented in Section 5.1.