ComfyUI-CacheDiT provides a one-click solution to accelerate Diffusion Transformer (DiT) models within the ComfyUI environment. It targets users of ComfyUI who leverage DiT architectures for image and video generation, offering significant speedups (1.4-2.0x) with minimal configuration and no perceivable quality degradation. The primary benefit is reducing inference times for computationally intensive DiT models.

How It Works

The node implements an intelligent caching strategy inspired by llm-scaler. After an initial warmup phase, it selectively reuses previously computed intermediate results based on a skip_interval and the current inference step. If the conditions are met, cached data is utilized; otherwise, new computations are performed, and the result is cached for future steps. This approach minimizes redundant computations, leading to substantial performance gains.

Quick Start & Requirements

• Installation: Clone the repository into ComfyUI/custom_nodes/ and install dependencies via pip install -r requirements.txt.
• Prerequisites: ComfyUI, Python environment. Specific hardware requirements depend on the DiT models being accelerated.
• Links:
  • ComfyUI: https://github.com/comfyanonymous/ComfyUI
  • Cache-DiT (inspiration): https://github.com/vipshop/cache-dit

Highlighted Details

• Achieves 1.4-2.0x speedup across various DiT models including Z-Image, Z-Image-Turbo, Qwen-Image-2512, Flux.2 Klein, LTX-2, and WAN2.2 14B.
• Features "one-click" acceleration with automatic parameter tuning, requiring zero manual configuration for most models.
• Includes dedicated nodes (LTX2 Cache Optimizer, Wan Cache Optimizer) for specialized architectures like LTX-2 and WAN2.2 14B to ensure optimal performance and temporal consistency.
• Reports minimal impact on image quality when acceleration is properly configured using default settings.

Maintenance & Community

No specific details regarding maintainers, sponsorships, partnerships, or community channels (like Discord/Slack) are provided in the README.

Licensing & Compatibility

• License: Apache 2.0.
• Compatibility: Designed for ComfyUI. The Apache 2.0 license is permissive, generally allowing for commercial use and integration into closed-source projects.

Limitations & Caveats

The speedup benefit is significantly reduced for inference tasks with very low step counts (less than 6 steps) due to warmup overhead. Model auto-detection may occasionally fail, requiring manual selection of the model_type preset. A 0% cache hit rate can occur if the model is not detected, inference steps are too short, or specific log messages are absent. Support for distilled low-step models beyond Z-Image-Turbo requires further validation.