Jasonzzt/ComfyUI-CacheDiT

Cache-DiT Node for Comfyui

What it solves

It reduces the time required to generate images and videos by accelerating Diffusion Transformer (DiT) models. It addresses the high computational cost of these models, providing speedups typically ranging from 1.4x to 2.0x without requiring complex configuration.

How it works

The project implements intelligent caching of transformer results. After an initial "warmup" phase where the first few steps are fully computed, the system reuses cached results at specific intervals (defined by a skip_interval) instead of recomputing them for every single step of the diffusion process.

For specific models like MiniMax H3, it uses a more advanced adaptive approach based on residual differences to decide when to reuse the cache, ensuring better quality preservation for joint audio-video generation.

Who it’s for

Users of ComfyUI who work with DiT-based image and video generation models (such as Flux.2, WAN2.2, LTX-2, and MiniMax H3) and want to decrease generation times.

Highlights

  • Significant Speedups: Achieves 1.4x to 2.0x acceleration across various models.
  • Zero Configuration: Includes an "Auto" preset that detects the model and applies the best settings.
  • Broad Model Support: Verified for a wide range of models including Z-Image, Qwen-Image, Flux.2 Klein, MiniMax H3, LTX-2, and WAN2.2 14B.
  • Dynamic Control: Offers advanced nodes for fine-tuning cache behavior (e.g., residual thresholds and warmup steps) for specific workflows.

Related

  • Project
    chengzeyi/Comfy-WaveSpeed

    An inference optimization suite for ComfyUI that uses dynamic caching and enhanced torch.compile to speed up image and video generation.

  • Project
    vipshop/cache-dit

    A PyTorch-native inference engine that accelerates Diffusion Transformers (DiTs) using hybrid caching, parallelism, and quantization to significantly reduce inference latency.

  • Project
    wildminder/ComfyUI-DyPE

    A ComfyUI custom node pack that enables ultra-high-resolution (4K+) image generation for Diffusion Transformers like FLUX and Qwen without requiring retraining.

  • Project
    xdit-project/xDiT

    xDiT is a scalable inference engine for Diffusion Transformers (DiTs) that uses hybrid parallelization and cache acceleration to enable real-time high-quality image and video generation across multiple GPUs.

  • Project
    ssitu/ComfyUI_UltimateSDUpscale

    A set of ComfyUI nodes that perform image-to-image diffusion on large images using a tiled approach to improve detail and reduce hardware requirements.