ICLR 2023（投稿） | 扩散模型相关论文分类整理

引言：ICLR是人工智能领域顶级会议之一，会议主题包括深度学习、统计和数据科学，以及一些重要的应用，例如：计算机视觉、计算生物学、语音识别、文本理解、游戏和机器人等。ICLR 2023将于2023年5月1日至5月5日在卢旺达基加利举行。官方的论文接受列表尚未公开，从投稿论文来看，扩散模型依然热度不减，是出现频率较高，且平均评分也较高的热点之一。

本文选取了与扩散模型相关的100多篇论文，按照不同的研究主题进行了分类整理，以供参考。ICLR 2023投稿论文openreview链接如下：

ICLR 2023 Conference | OpenReview

https://openreview.net/group?id=ICLR.cc/2023/Conference

1. 高效采样

• • Dynamic Scheduled Sampling with Imitation Loss for Neural Text Generation

• • Truncated Diffusion Probabilistic Models and Diffusion-based Adversarial Auto-Encoders

• • Denoising Diffusion Samplers

• • Denoising MCMC for Accelerating Diffusion-Based Generative Models

• • DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models

• • Quasi-Taylor Samplers for Diffusion Generative Models based on Ideal Derivatives

• • Fast Sampling of Diffusion Models with Exponential Integrator

• • Accelerating Guided Diffusion Sampling with Splitting Numerical Methods

• • Boomerang: Local sampling on image manifolds using diffusion models

• • Markup-to-Image Diffusion Models with Scheduled Sampling

2. 和其它生成模型结合

• • Diffusion-GAN: Training GANs with Diffusion

• • in Conversation based on offline reinforcement learning

• • FastDiff 2: Dually Incorporating GANs into Diffusion Models for High-Quality Speech Synthesis

• • Reduce, Reuse, Recycle: Compositional Generation with Energy-Based Diffusion Models and MCMC

• • Geometric Networks Induced by Energy Constrained Diffusion

• • Progressive Image Synthesis from Semantics to Details with Denoising Diffusion GAN

• • Flow Matching for Generative Modeling

• • SPI-GAN: Denoising Diffusion GANs with Straight-Path Interpolations

• • Building Normalizing Flows with Stochastic Interpolants

• • Guiding Energy-based Models via Contrastive Latent Variables

• • Your Denoising Implicit Model is a Sub-optimal Ensemble of Denoising Predictions

• • Thinking fourth dimensionally: Treating Time as a Random Variable in EBMs

3. 在CV、NLP领域的应用

• • Novel View Synthesis with Diffusion Models

• • Pyramidal Denoising Diffusion Probabilistic Models

• • Compositional Image Generation and Manipulation with Latent Diffusion Models

• • Towards the Detection of Diffusion Model Deepfakes

• • DifFace: Blind Face Restoration with Diffused Error Contraction

• • Restoration based Generative Models

• • Generative Modelling with Inverse Heat Dissipation

• • Deep Watermarks for Attributing Generative Models

• • Learning multi-scale local conditional probability models of images

• • Images as Weight Matrices: Sequential Image Generation Through Synaptic Learning Rules

• • Self-conditioned Embedding Diffusion for Text Generation

• • Sequence to sequence text generation with diffusion models

• • DiffusER: Diffusion via Edit-based Reconstruction

• • SDMuse: Stochastic Differential Music Editing and Generation via Hybrid Representation

• • Universal Speech Enhancement with Score-based Diffusion

• • Score-based Generative 3D Mesh Modeling

• • CAN: A simple, efficient and scalable contrastive masked autoencoder framework for learning visual representations

• • Neural Volumetric Mesh Generator

• • SketchKnitter: Vectorized Sketch Generation with Diffusion Models

• • $DDM^2$: Self-Supervised Diffusion MRI Denoising with Generative Diffusion Models

• • Neural Image Compression with a Diffusion-based Decoder

• • Lossy Compression with Gaussian Diffusion

• • Distilling Model Failures as Directions in Latent Space

• • Lossy Image Compression with Conditional Diffusion Models

• • Quantized Compressed Sensing with Score-Based Generative Models

• • Out-of-distribution Detection with Diffusion-based Neighborhood

4. 在多模态领域的应用

• • DreamFusion: Text-to-3D using 2D Diffusion

• • Diffusion-based Image Translation using disentangled style and content representation

• • CUSTOMIZING PRE-TRAINED DIFFUSION MODELS FOR YOUR OWN DATA

• • Human Motion Diffusion Model

• • Prosody-TTS: Self-Supervised Prosody Pretraining with Latent Diffusion For Text-to-Speech

• • Text-Guided Diffusion Image Style Transfer with Contrastive Loss Fine-tuning

• • Meta-Learning via Classifier(-free) Guidance

• • KNN-Diffusion: Image Generation via Large-Scale Retrieval

• • DiffEdit: Diffusion-based semantic image editing with mask guidance

• • Training-Free Structured Diffusion Guidance for Compositional Text-to-Image Synthesis

• • Discrete Contrastive Diffusion for Cross-Modal Music and Image Generation

• • Unified Discrete Diffusion for Simultaneous Vision-Language Generation

• • ResGrad: Residual Denoising Diffusion Probabilistic Models for Text to Speech

• • Re-Imagen: Retrieval-Augmented Text-to-Image Generator

• • Prompt-to-Prompt Image Editing with Cross-Attention Control

5. 与强化学习结合

• • Diffusion Policies as an Expressive Policy Class for Offline Reinforcement Learning

• • Offline Reinforcement Learning via High-Fidelity Generative Behavior Modeling

• • Provably Efficient Reinforcement Learning for Online Adaptive Influence Maximization

• • Variational Reparametrized Policy Learning with Differentiable Physics

• • Is Conditional Generative Modeling all you need for Decision Making?

6. 分子图建模

• • Diffusion Probabilistic Modeling of Protein Backbones in 3D for the motif-scaffolding problem

• • Protein structure generation via folding diffusion

• • Pre-training Protein Structure Encoder via Siamese Diffusion Trajectory Prediction

• • DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking

• • Pocket-specific 3D Molecule Generation by Fragment-based Autoregressive Diffusion Models

• • Equivariant 3D-Conditional Diffusion Models for Molecular Linker Design

• • Structure-based Drug Design with Equivariant Diffusion Models

• • Equivariant Energy-Guided SDE for Inverse Molecular Design

• • 3D Equivariant Diffusion for Target-Aware Molecule Generation and Affinity Prediction

• • Exploring Chemical Space with Score-based Out-of-distribution Generation

• • Protein Sequence and Structure Co-Design with Equivariant Translation

7. 扩散模型理论与理解

• • Information-Theoretic Diffusion

• • Analyzing diffusion as serial reproduction

• • Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions

• • Diffusion Models Already Have A Semantic Latent Space

• • Unifying Diffusion Models' Latent Space, with Applications to CycleDiffusion and Guidance

• • Understanding DDPM Latent Codes Through Optimal Transport

• • Interpreting Neural Networks Through the Lens of Heat Flow

• • gDDIM: Generalized denoising diffusion implicit models

8.扩散模型泛化与拓展

• • Soft Diffusion: Score Matching For General Corruptions

• • Where to Diffuse, How to Diffuse and How to get back: Learning in Multivariate Diffusions

• • Blurring Diffusion Models

• • Diffusion Probabilistic Fields

• • Neural Diffusion Processes

• • Pseudoinverse-Guided Diffusion Models for Inverse Problems

• • Removing Structured Noise with Diffusion Models

• • f-DM: A Multi-stage Diffusion Model via Progressive Signal Transformation

• • Iterative α-(de)Blending: Learning a Deterministic Mapping Between Arbitrary Densities

• • Score-Based Graph Generative Modeling with Self-Guided Latent Diffusion

• • Self-Guided Diffusion Models

• • From Points to Functions: Infinite-dimensional Representations in Diffusion Models

• • Score Matching via Differentiable Physics

• • Approximated Anomalous Diffusion: Gaussian Mixture Score-based Generative Models

• • Action Matching: A Variational Method for Learning Stochastic Dynamics from Samples

• • Autoregressive Generative Modeling with Noise Conditional Maximum Likelihood Estimation

• • DIFFUSION GENERATIVE MODELS ON SO(3)

• • Diffusion Posterior Sampling for General Noisy Inverse Problems

9.扩散模型迁移

• • Transferring Pretrained Diffusion Probabilistic Models

• • Cold Diffusion: Inverting Arbitrary Image Transforms Without Noise

• • Dual-Domain Diffusion Based Progressive Style Rendering towards Semantic Structure Preservation

• • Dual Diffusion Implicit Bridges for Image-to-Image Translation

• • Learning to Learn with Generative Models of Neural Network Checkpoints

• • Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model

10.特殊结构数据的建模

• • Autoregressive Diffusion Model for Graph Generation

• • Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioning

• • TabDDPM: Modelling Tabular Data with Diffusion Models

• • ChiroDiff: Modelling chirographic data with Diffusion Models

• • Modeling Temporal Data as Continuous Functions with Process Diffusion

• • Domain Specific Denoising Diffusion Probabilistic Models for Brain Dynamics

• • Discrete Predictor-Corrector Diffusion Models for Image Synthesis

• • Diffusion-based point cloud generation with smoothness constraints

• • Computational Doob h-transforms for Online Filtering of Discretely Observed Diffusions

• • Imitating Human Behaviour with Diffusion Models

• • Learning Diffusion Bridges on Constrained Domains

• • DiGress: Discrete Denoising diffusion for graph generation

• • Score-based Continuous-time Discrete Diffusion Models

• • Brain Signal Generation and Data Augmentation with a Single-Step Diffusion Probabilistic Model

11. 鲁棒性与稳定性

• • DensePure: Understanding Diffusion Models towards Adversarial Robustness

• • Defending against Adversarial Audio via Diffusion Model

• • PointDP: Diffusion-driven Purification against 3D Adversarial Point Clouds

• • Diffusion Adversarial Representation Learning for Self-supervised Vessel Segmentation

• • Improving Adversarial Robustness by Contrastive Guided Diffusion Process

• • Robustness for Free: Adversarially Robust Anomaly Detection Through Diffusion Model

• • (Certified!!) Adversarial Robustness for Free!

• • The Biased Artist: Exploiting Cultural Biases via Homoglyphs in Text-Guided Image Generation Models

• • Expected Perturbation Scores for Adversarial Detection

• • Input Perturbation Reduces Exposure Bias in Diffusion Models

• • Stable Target Field for Reduced Variance Score Estimation

12.扩散模型的隐私保护

• • Membership Inference Attacks Against Text-to-image Generation Models

• • Differentially Private Diffusion Models

13. 其它方向

• • OCD: Learning to Overfit with Conditional Diffusion Models

• • Denoising Diffusion Error Correction Codes

• • Neural Lagrangian Schrodinger Bridge: Diffusion Modeling for Population Dynamics

• • Diffusion Models for Causal Discovery via Topological Ordering

• • Transport with Support: Data-Conditional Diffusion Bridges

• • A Score-Based Model for Learning Neural Wavefunctions