FORA: Fast-Forward Caching in Diffusion Transformer Acceleration Diffusion transformers (DiT) have become the de facto choice for generating high-quality images and videos due to their scalability. However, their increased size leads to higher inference costs, making them less suitable for real-time applications. This paper presents FORA, a simple yet effective caching mechanism that accelerates DiT by exploiting the repetitive nature of the diffusion process. FORA stores and reuses intermediate outputs from the attention and MLP layers across denoising steps, reducing computational overhead without requiring model retraining. Experiments show that FORA significantly speeds up diffusion transformers while minimally affecting performance metrics such as IS Score and FID. FORA represents a significant advancement in deploying diffusion transformers for real-time applications. The paper introduces FORA, a caching strategy tailored for transformer-based diffusion models. This mechanism capitalizes on the repetitive aspects of the diffusion process by preserving and reusing intermediate outputs from attention and MLP layers during inference. FORA significantly cuts down computational overhead and integrates seamlessly with existing DiT models without necessitating retraining. As a result, it effectively reduces computational costs while maintaining output quality. The paper performs experiments to assess the performance of FORA, demonstrating notable improvements in inference speed and computational efficiency. Our findings underscore FORA’s potential to make high-performance generative models more suitable for real-time use. FORA implements a static caching mechanism, a straightforward yet powerful approach, to accelerate the sampling process in diffusion models. This method operates on a simple principle: recompute and cache features at regular intervals, and reuse these cached features for a predetermined number of subsequent time steps. At the core of this mechanism is a single hyperparameter N, which we call the cache interval. This interval determines how frequently the model recomputes and caches new features. Specifically, N is an integer that can range from 1 to T - 1, where T is the total number of sampling time steps in the diffusion process. The static caching process unfolds as follows: initialization, caching condition, recomputation and caching, feature reuse, and cycle repetition. The effectiveness of static caching hinges on the choice of the cache interval N. A smaller N leads to more frequent recomputations, potentially preserving more accuracy but offering less computational savings. Conversely, a larger N increases computational efficiency but may impact the quality of the generated outputs. In our experiments, we found that the optimal value of N depends on the specific requirements of the task and the desired trade-off between speed and quality. Through extensive testing, we determined that setting N max to 7 provides a good balance. Beyond this value, we observed a significant degradation in the Fréchet Inception Distance (FID) score, indicating a decline in the quality of generated images. The paper presents comprehensive experimental results, demonstrating that FORA significantly improves inference speed while maintaining image quality. FORA is shown to be effective for both class-conditional and text-conditional imageFORA: Fast-Forward Caching in Diffusion Transformer Acceleration Diffusion transformers (DiT) have become the de facto choice for generating high-quality images and videos due to their scalability. However, their increased size leads to higher inference costs, making them less suitable for real-time applications. This paper presents FORA, a simple yet effective caching mechanism that accelerates DiT by exploiting the repetitive nature of the diffusion process. FORA stores and reuses intermediate outputs from the attention and MLP layers across denoising steps, reducing computational overhead without requiring model retraining. Experiments show that FORA significantly speeds up diffusion transformers while minimally affecting performance metrics such as IS Score and FID. FORA represents a significant advancement in deploying diffusion transformers for real-time applications. The paper introduces FORA, a caching strategy tailored for transformer-based diffusion models. This mechanism capitalizes on the repetitive aspects of the diffusion process by preserving and reusing intermediate outputs from attention and MLP layers during inference. FORA significantly cuts down computational overhead and integrates seamlessly with existing DiT models without necessitating retraining. As a result, it effectively reduces computational costs while maintaining output quality. The paper performs experiments to assess the performance of FORA, demonstrating notable improvements in inference speed and computational efficiency. Our findings underscore FORA’s potential to make high-performance generative models more suitable for real-time use. FORA implements a static caching mechanism, a straightforward yet powerful approach, to accelerate the sampling process in diffusion models. This method operates on a simple principle: recompute and cache features at regular intervals, and reuse these cached features for a predetermined number of subsequent time steps. At the core of this mechanism is a single hyperparameter N, which we call the cache interval. This interval determines how frequently the model recomputes and caches new features. Specifically, N is an integer that can range from 1 to T - 1, where T is the total number of sampling time steps in the diffusion process. The static caching process unfolds as follows: initialization, caching condition, recomputation and caching, feature reuse, and cycle repetition. The effectiveness of static caching hinges on the choice of the cache interval N. A smaller N leads to more frequent recomputations, potentially preserving more accuracy but offering less computational savings. Conversely, a larger N increases computational efficiency but may impact the quality of the generated outputs. In our experiments, we found that the optimal value of N depends on the specific requirements of the task and the desired trade-off between speed and quality. Through extensive testing, we determined that setting N max to 7 provides a good balance. Beyond this value, we observed a significant degradation in the Fréchet Inception Distance (FID) score, indicating a decline in the quality of generated images. The paper presents comprehensive experimental results, demonstrating that FORA significantly improves inference speed while maintaining image quality. FORA is shown to be effective for both class-conditional and text-conditional image