ComfyUI-CacheDiT Speed Boosts DiT models
ComfyUI-CacheDiT is a new custom node that accelerates Diffusion Transformer (DiT) models in ComfyUI. It delivers 1.4 up to 2 times faster generation speeds through intelligent caching, with no manual configuration needed.
Developer Jasonzzt created this tool to address speed bottlenecks common with large transformer-based models. The node works by storing and reusing intermediate computations, reducing the processing load during both image and video generation.
Speed ups and compatibility features of ComfyUI-CacheDiT
- Automatic acceleration from 1.4x to 2.0x depending on the model.
- Zero-configuration setup with auto-detection for most models.
- Support for popular models including Flux.2 Klein, LTX-2, and WAN2.2 14B.
- Built-in performance dashboard to monitor cache efficiency.
- Simple three-step workflow for image models.
Users working with video generation models or high-resolution image synthesis can save significant rendering time. The tool handles different model architectures automatically, making it practical for workflows that switch between multiple diffusion models.
Technical details and general limitations
The extension uses a 'warmup' phase during the first few inference steps before activating its caching mechanism. According to the developer, quality impact remains minimal when using default settings.
Some models require dedicated optimizer nodes rather than the standard accelerator. LTX-2 uses dual latent paths for audio and video, while WAN2.2 14B employs a Mixture of Experts architecture with separate high-noise and low-noise components. The developer notes that extremely low step counts under six steps may not benefit much from caching, as the warmup overhead reduces potential gains.
Get ComfyUI-CacheDiT on GitHub.
