Warped Diffusion: Solving Video Inverse Problems with Image Diffusion Models

Authors: Giannis Daras, Weili Nie, Karsten Kreis, Alex Dimakis, Morteza Mardani, Nikola Kovachki, Arash Vahdat

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of our method for video inpainting and 8 video super-resolution, outperforming existing techniques based on noise transformations. We provide generated video results in the following URL: https://giannisdaras.github.io/warped_diffusion.github.io/. We extensively validate our method on video inpainting and super-resolution.
Researcher Affiliation Collaboration Giannis Daras UT Austin Weili Nie NVIDIA Karsten Kreis NVIDIA Alexandros G. Dimakis UT Austin Morteza Mardani NVIDIA Nikola B. Kovachki NVIDIA Arash Vahdat NVIDIA
Pseudocode Yes Algorithm 1 Warped Diffusion Temporal Consistency with Equivariance Self Guidance
Open Source Code No We do not yet release our code for noise warping and sampling guidance, but we are working on open-sourcing it.
Open Datasets Yes We train models with and without correlated noise on the COYO dataset [12] for 100k steps.
Dataset Splits No The paper mentions a 'test split of the COYO dataset' but does not specify a distinct validation set split (e.g., percentages or counts for training, validation, and test splits).
Hardware Specification Yes We train all our models on 16 A100 GPUs on a SLURM-based cluster. We perform all our sampling experiments on a single A-100 GPU.
Software Dependencies No The paper mentions software like Stable Diffusion XL, RAFT model, and PyTorch, but does not provide specific version numbers for these software components or programming languages (e.g., 'PyTorch 1.9', 'Python 3.8').
Experiment Setup Yes We use the following training hyperparameters: Training resolution: 1024 1024. Batch size: 64. Latent resolution: 128. Optimizer Adam with Weight Decay. Optimizer parameters: Learning rate: 5e 6 β1 = 0.9 β2 = 0.999 Weight Decay: 1e 2 ϵ = 1e 08 Max Gradient Norm (Gradient clipping): 1.0