🚀 Awesome Acceleration GenAI

A curated list of awesome papers on acceleration techniques for Generative AI.

---

📑 Contents

• Cache Methods
• Token Merging & Sparse
• Optimal Timestep
• Distillation
• Image Generation
• Video Generation
• Others

---

⚡ Cache Methods

• Timestep Embedding Tells: It’s Time to Cache for Video Diffusion Model, CVPR 2025 [Paper] [Code]
• Adaptive Caching for Faster Video Generation with Diffusion Transformers, arXiv 2024 [Paper] [Code]
• PAB: Real-time video generation with pyramid attention broadcast, ICLR 2025 [Paper] [Code]
• SmoothCache: A Universal Inference Acceleration Technique for Diffusion Transformers, arXiv 2024 [Paper] [Code]
• Δ-DiT: Training-Free Acceleration for Diffusion Transformers, arXiv 2024 [Paper]
• T-Gate: Faster Diffusion via Temporal Attention Decomposition, TMLR 2025 [Paper] [Code]
• Unveiling Redundancy in Diffusion Transformers (DiTs): A Systematic Study, arXiv 2024 [Paper] [Code]
• Accelerating Diffusion Transformer via Error-Optimized Cache, arXiv 2025 [Paper]
• ToCa: Token-wise Feature Caching for Diffusion Transformers, ICLR 2025 [Paper] [Code]
• DuCa: Dual Feature Caching for Diffusion Transformers, arXiv 2024 [Paper] [Code]
• UniCP: Unified Caching and Pruning Framework for Efficient Video Generation, arXiv 2025 [Paper]
• FasterCache: Training-Free Video Diffusion Acceleration with High Quality, arXiv 2024 [Paper] [Code]
• Region-Adaptive Sampling for Diffusion Transformers, arXiv 2025 [Paper] [Code]
• CacheQuant: Comprehensively Accelerated Diffusion Models, CVPR 2025 [Paper] [Code]
• Cache Me if You Can: Accelerating Diffusion Models through Block Caching, CVPR 2024 [Paper] [Code]
• BlockDance: Reuse Structurally Similar Spatio-Temporal Features, arXiv 2025 [Paper]
• Learning-to-Cache: Layer Caching for Diffusion Transformers, NeurIPS 2024 [Paper] [Code]
• TaylorSeers: Forecasting for Accelerating Diffusion Models, arXiv 2025 [Paper] [Code]
• Profiling-Based Feature Reuse for Video Diffusion Models, arXiv 2025 [Paper] [Code]
• SRDiffusion: Sketching-Rendering Cooperation, arXiv 2025 [Paper]
• AB-Cache: Adams-Bashforth Cached Feature Reuse, arXiv 2025 [Paper]
• EEdit: Rethinking Spatial and Temporal Redundancy for Efficient Image Editing, ICCV 2025 [Paper] [Code]
• CacheDiT: Training-free Cache Acceleration Toolbox for Diffusion Transformers, arXiv 2025 [Code]
• Confidence-Gated Taylor Forecasting for Acceleration, arXiv 2025 [Paper]
• HiCache: Hermite Polynomial-based Feature Caching, arXiv 2025 [Paper]
• ERTACache: Error Rectification and Timesteps Adjustment, arXiv 2025 [Paper]
• DiCache: Let Diffusion Model Determine Its Own Cache, arXiv 2025 [Paper]
• OmniCache: Global Trajectory-Oriented Cache Reuse, arXiv 2025 [Paper]
• EasyCache: Runtime-Adaptive Caching for Video Diffusion, arXiv 2025 [Paper] [Code]
• Forecast then Calibrate: Caching as ODE for Diffusion Transformers, arXiv 2025 [Paper]
• SpecDiff: Accelerating Diffusion Model Inference with Self-Speculation, arXiv 2025 [Paper]
• Evolutionary Caching to Accelerate Your Off-the-Shelf Diffusion Model, arXiv 2025 [Paper] [Code]
• SemCache:Adaptive Semantic-Aware Caching for Efficient Video Diffusion arXiv 2025 [Paper] [Code]

---

🔀 Token Merging & Sparse

• AsymRnR: Video Diffusion Transformers Acceleration with Asymmetric Reduction and Restoration, arXiv 2024 [Paper]
• Dynamic Token Carving (Jenga): Training-Free Efficient Video Generation, arXiv 2025 [Paper] [Code]
• VORTA: Efficient Video Diffusion via Routing Sparse Attention, arXiv 2025 [Paper]
• Sparse VideoGen: Accelerating with Spatial-Temporal Sparsity, ICML 2025 [Paper] [Code]
• Sparse VideoGen2: Semantic-Aware Permutation, arXiv 2025 [Paper] [Code]

---

⏱ Optimal Timestep

• Shortest Path in Denoising Diffusion, CVPR 2025 [Paper] [Code]
• BOSS: Bellman Optimal StepSize, ICLR 2024 [Paper] [Code]
• Optimal Stepsize for Diffusion Sampling, arXiv 2025 [Paper] [Code]

---

🎯 Distillation

• DKDM: Data-Free Knowledge Distillation for Diffusion Models, CVPR 2025 [Paper] [Code]
• Distilling Diversity and Control in Diffusion Models, arXiv 2025 [Paper] [Code]
• Glance: Accelerating Diffusion Models with 1 Sample, arXiv 2025 [Paper] [Code]

---

🖼 Image Generation

• EEdit: Efficient Image Editing by Redundancy Reduction, ICCV 2025 [Paper] [Code]

---

🎬 Video Generation

• Impossible Videos, arXiv 2025 [Paper] [Code]
• SRDiffusion: Sketching-Rendering Cooperation, arXiv 2025 [Paper]

---

📌 Others

• Inference-Time Scaling for Diffusion Models beyond Scaling Denoising Steps, CVPR 2025 [Paper]
• Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding, arXiv 2025 [Paper] [Code]

---

📑 Resources

https://github.com/VILA-Lab/Awesome-DLMs

---

✨ This list is continuously updated with the latest works.