Awesome GitHub RepositoriesDiffusion Acceleration

Enables real-time video generation using pyramid attention broadcast.

Reduces token count for faster stable diffusion inference.

Applies structural pruning to reduce model complexity.

Uses token-wise feature caching to speed up transformers.

Provides a universal caching technique for transformer models.

Rethinks encoder roles to improve inference efficiency.

Uses bounded difference approximation for adaptive inference speed.

Integrates softmax and linear attention for efficient processing.

Compresses attention layers specifically for transformer-based models.

Optimizes attention mechanisms in transformers for faster generation.

Decomposes temporal attention to reduce inference overhead.

Implements fast-forward caching for transformer architectures.

Provides memory-efficient caching for video generation tasks.

Caches feature maps to accelerate model inference.

Parallelizes denoising processes to reduce total inference time.

Merges tokens based on structure to accelerate transformers.

Automates the optimization of time steps and architectures.