Welcome to my presentation archive! This repository contains slides and notes from various presentations I have delivered on topics such as Diffusion Models and NeRf. Each presentation is accompanied by a brief summary of its content and objectives. Feel free to explore, and I hope these resources provide insights and value!
Each presentation can be accessed via the links provided below. You can download the files to view the slides and review the content.
📽 Note: To play embedded videos in the PDF slides, please use Adobe Acrobat Reader or another compatible viewer that supports multimedia content.
If you're interested in specific topics or want to dive deeper into any particular presentation, feel free to reach out or check the Resources section.
- Fast Sampling of Diffusion Models via Operator Learning
- Alleviating Exposure Bias in Diffusion Models through Time-Shift Sampling
- Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference
- MVDiffusion++: A Dense High-resolution Multi-view Diffusion Model for Single or Sparse-view 3D Object Reconstruction
- Flow Matching for Generative Modeling
- Instant Neural Graphics Primitives with a Multiresolution Hash Encoding
- Neuralangelo: High-Fidelity Neural Surface Reconstruction
- RadSplat: Radiance Field-Informed Gaussian Splatting for Robust Real-Time Rendering with 900+ FPS
- FRESCO: Spatial-Temporal Correspondence for Zero-Shot Video Translation
- FlowMo: Variance-Based Flow Guidance for Coherent Motion in Video Generation
- Date: October 16, 2023
- File: Fast Sampling of Diffusion Models via Operator Learning (PDF)
- Venue: ICML 2023
- Description: Introduces a method for accelerating diffusion model sampling using neural operators and temporal convolutions. This approach enables efficient learning of diffusion trajectories while minimizing computational complexity.
- Key Topics: Diffusion Models
- Date: October 30, 2023
- File: Alleviating Exposure Bias in Diffusion Models through Time-Shift Sampling (PDF)
- Venue: ICLR 2024
- Description: Proposes a method to reduce exposure bias in diffusion probabilistic models by adjusting time steps during sampling, improving alignment between training and inference distributions.
- Key Topics: Diffusion Models
- Date: November 16, 2023
- File: LCM (PDF)
- Venue: arXiv Preprint
- Description: Introduces LCM for high-resolution image synthesis with few-step inference, leveraging consistency distillation and skipping-step techniques to achieve faster and more efficient generative modeling.
- Key Topics: Diffusion Models
- Date: November 30, 2023
- File: Retentive Network (PDF)
- Venue: arXiv Preprint
- Description: Proposes the Retentive Network (RetNet), a foundational architecture for large language models designed to improve training parallelism, reduce inference costs, and maintain strong performance while addressing limitations of Transformers.
- Key Topics: LLMs
- Date: May 17, 2024
- File: Instant-NGP (PDF)
- Venue: SIGGRAPH 2022
- Description: Presents a method for representing neural graphics primitives using multiresolution hash encoding, enabling efficient, high-quality approximations with compact neural networks.
- Key Topics: NeRF and Splatting
- Date: May 29, 2024
- File: Sora (PDF)
- Venue: arXiv Preprint
- Description: Explores Sora's development as a text-to-video generative AI model, its underlying technologies, applications in various industries, challenges in safe and unbiased generation, and potential future advancements in text-to-video AI.
- Key Topics: Diffusion Models
- Date: June 7, 2024
- File: Neuralangelo (PDF)
- Venue: CVPR 2023
- Description: Proposes a framework for reconstructing high-fidelity 3D surfaces from RGB images using neural volume rendering, leveraging numerical gradients and coarse-to-fine optimization for detailed and continuous surfaces.
- Key Topics: NeRF and Splatting
- Date: June 28, 2024
- File: FRESCO (PDF)
- Venue: CVPR 2024
- Description: Proposes a framework for high-quality, temporally coherent video translation using spatial-temporal correspondences and an inversion-free approach to adapt pre-trained image diffusion models for video editing.
- Key Topics: Diffusion Models
- Date: July 26, 2024
- File: RadSplat (PDF)
- Venue: arXiv Preprint
- Description: Introduces a real-time rendering framework combining radiance field priors with Gaussian splatting to achieve high-quality visualization of large-scale 3D scenes at exceptional speed.
- Key Topics: NeRF and Splatting
MVDiffusion++: A Dense High-resolution Multi-view Diffusion Model for Single or Sparse-view 3D Object Reconstruction
- Date: August 23, 2024
- File: MVDiffusion++ (PDF)
- Venue: ECCV 2024
- Description: Proposes a high-resolution multi-view diffusion framework designed to reconstruct 3D objects from single or sparse-view inputs, leveraging correspondence-aware attention and efficient self-attention mechanisms.
- Key Topics: Diffusion Models
- Date: November 6, 2024
- File: Flow Matching (PDF)
- Venue: ICLR 2023
- Description: Proposes Flow Matching as an efficient training framework for continuous normalizing flows, leveraging vector fields of fixed probability paths to simplify training and sampling while achieving state-of-the-art generative modeling.
- Key Topics: Diffusion Models
- Date: June 11, 2025
- File: Force Prompting (PDF)
- Venue: arXiv Preprint 2025
- Description: : Proposes Force Prompting, a method for controlling video diffusion models using localized and global force inputs. Trained on limited synthetic data, the model generalizes to diverse objects and scenes without 3D assets, enabling intuitive, physics-aware video generation with emergent mass understanding.
- Key Topics: Video Generation, Diffusion Models, Physics-based Control
- Date: June 20, 2025
- File: FlowMo (PDF)
- Venue: arXiv Preprint
- Description: Presents FlowMo, a training-free method that improves motion coherence in text-to-video generation by leveraging variance-based flow guidance. The approach extracts temporal representations directly from pre-trained flow matching models without requiring additional training or external inputs.
- Key Topics: Video Generation, Flow Matching, Temporal Coherence
- Date: July 18, 2025
- File: 4Real-Video-V2 (PDF)
- Venue: arXiv Preprint
- Description: Proposes 4Real-Video-V2, a parameter-efficient 4D scene generation method that uses fused view-time attention and a feedforward reconstructor for text-to-4D video generation. Unlike prior methods, it requires no extra weights beyond a pre-trained video model and replaces explicit optimization with fast feedforward 3D Gaussian splatting.
- Key Topics: 4D Scene Generation, Text-to-Video, Feedforward Reconstruction, Dynamic Gaussians
Here are some additional resources related to the topics covered in these presentations:
- Recommended Reading: TODO.
I welcome feedback and suggestions! If you have any questions or would like to discuss any of the topics covered, please feel free to reach out or open an issue. If you find these presentations helpful, consider giving this repo a ⭐️ star!
Thank you for visiting my presentation archive repository. Enjoy exploring the content!