TeaCache is a training-free caching method designed to accelerate inference for diffusion models, particularly video diffusion models, by leveraging timestep embedding differences. It targets researchers and developers working with generative AI who need to optimize inference speed without retraining.

How It Works

TeaCache estimates and caches intermediate outputs based on the dynamic changes in timestep embeddings during the diffusion process. This approach avoids redundant computations by intelligently reusing previously computed states, leading to significant speedups. Its advantage lies in its training-free nature and broad applicability across various diffusion model architectures.

Quick Start & Requirements

• Installation: Typically integrated into existing diffusion model frameworks (e.g., Diffusers, ComfyUI). Specific integration instructions vary by model and framework.
• Prerequisites: Python, PyTorch, and dependencies of the target diffusion model. GPU acceleration is highly recommended for practical inference speeds.
• Resources: Requires memory for caching, proportional to the model size and batch size.
• Links:
  • Project Page: https://liewfeng.github.io/TeaCache/
  • Paper: https://arxiv.org/abs/2411.19108

Highlighted Details

• CVPR 2025 Highlight (top 3.7% of papers).
• Supports a wide range of models including Wan2.1, Cosmos, CogVideoX1.5, FLUX, TangoFlux, and more for Text-to-Video, Image-to-Video, Text-to-Image, and Text-to-Audio tasks.
• Numerous community integrations for popular frameworks like ComfyUI and Diffusers.
• Achieves significant inference speedups across various supported models.

Maintenance & Community

The project is actively maintained with recent updates and contributions from institutions like Alibaba Group and universities. It has a growing community with many integrations and support for new models.

Licensing & Compatibility

The majority of the project is released under the Apache 2.0 license. Users must also adhere to the licenses of the underlying diffusion models it integrates with. Apache 2.0 is generally permissive for commercial use.

Limitations & Caveats

The effectiveness and specific implementation details of TeaCache may vary depending on the target diffusion model architecture. While training-free, it requires careful integration into existing inference pipelines.