Project status: the caching and compilation nodes below are stable and widely used, but active development on this repository is intermittent — please treat the unchecked item above as an idea rather than a roadmap. Issues and pull requests are still welcome, and the Discord server is the fastest way to get a question answered.

Discord Server

Comfy Registry

Installation

cd custom_nodes
git clone https://github.com/chengzeyi/Comfy-WaveSpeed.git

Usage

Demo Workflows

You can find demo workflows in the workflows folder.

NOTE: The Compile Model+ node requires your computation to meet some software and hardware requirements, please refer to the Enhanced torch.compile section for more information. If you have problems with the compilation node, you can remove it from the workflow and only use the Apply First Block Cache node. The Apply First Block Cache node can still bring you a significant speedup.

Dynamic Caching (First Block Cache)

Inspired by TeaCache and other denoising caching algorithms, we introduce First Block Cache (FBCache) to use the residual output of the first transformer block as the cache indicator. If the difference between the current and the previous residual output of the first transformer block is small enough, we can reuse the previous final residual output and skip the computation of all the following transformer blocks. This can significantly reduce the computation cost of the model, achieving a speedup of up to 2x while maintaining high accuracy.

To use first block cache, simply add the wavespeed->Apply First Block Cache node to your workflow after your Load Diffusion Model node and adjust the residual_diff_threashold value to a suitable value for your model, for example: 0.12 for flux-dev.safetensors with fp8_e4m3fn_fast and 28 steps. It is expected to see a speedup of 1.5x to 3.0x with acceptable accuracy loss.

It supports many models like FLUX, LTXV (native and non-native), HunyuanVideo (native), SD3.5 and SDXL, feel free to try it out and let us know if you have any issues!

Some configurations for different models that you can try:

| Model | Steps | residual_diff_threashold | | - | - | - | | flux-dev.safetensors with fp8_e4m3fn_fast | 28 | 0.12 | | ltx-video-2b-v0.9.1.safetensors | 30 | 0.1 | | hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors | 20 | 0.1 | | sd3.5_large_fp8_scaled.safetensors | 30 | 0.12 | | sd_xl_base_1.0.safetensors | 25 | 0.2 |

NOTE: SDXL First Block Cache is incompatible with the FreeU Advanced node pack and will not function properly if it is installed and enabled.

See Apply First Block Cache on FLUX.1-dev for more information and detailed comparison on quality and speed.

Usage of First Block Cache

Enhanced `torch.compile`

To use the Enhanced torch.compile, simply add the wavespeed->Compile Model+ node to your workflow after your Load Diffusion Model node or Apply First Block Cache node. The compilation process happens the first time you run the workflow, and it takes quite a long time, but it will be cached for future runs. You can pass different mode values to make it runs faster, for example max-autotune or max-autotune-no-cudagraphs. One of the advantages of this node over the original TorchCompileModel node is that it works with LoRA.

It is suggested to pass --gpu-only when launching your ComfyUI if you are using this node, for example, if you are using comfy-cli:

comfy launch -- --gpu-only

If you encounter any problems with frequent compilation after changing the resolution or text prompt, you could try enabling the dynamic option of the Compile Model+ node. Or you could launch your ComfyUI with environment variable TORCH_LOGS=recompiles_verbose to debug the reason for recompilation.

NOTE: torch.compile might not be able to work with model offloading well, you could try passing --gpu-only when launching your ComfyUI to disable model offloading.

NOTE: torch.compile does not work on Windows offcially, you should not use this node if you are facing this problem, or search on Google to find out how to make it work.

NOTE: Compiling a model with FP8 quantization does not work on pre-Ada GPUs like RTX 3090, you should try using FP16/BF16 models or removing the compilation node.

Usage of Enhanced torch.compile

Running these models without a local GPU

This plugin makes local ComfyUI inference faster. If you don't have a GPU to run these models on in the first place, the same model families are available on WaveSpeed AI, an AI image & video generation platform.

Browser playgrounds — no install, no API key:

Image generator — Nano Banana 2, Seedream, GPT Image, Z-Image and more, switchable from one page
Video generator — Seedance, Wan, Kling, Veo and more
Full model catalog

Popular endpoints:

From ComfyUI, the wavespeed-comfyui plugin exposes the whole hosted catalog through a single node whose widgets are generated from each model's schema. It is a separate project from this one — this repository optimizes local inference, that one calls the hosted API.

From the terminal or your own code, see the CLI and the Python / JavaScript SDKs.

Others

Use with `diffusers`

Please refer to ParaAttention for more information.