ComfyUI-Lightning

The codes are modified from SageAttention, ComfyUI-TeaCache, comfyui-flux-accelerator and Comfy-WaveSpeed.

I use follow tips to accelerate FLUX inference speed:

1. Add TeaCache or FBCache or MBCache;
2. Skip some unnessasery blocks;
3. Compile and quantize model;
4. Use fast CuDNN attention kernels;
5. Use SageAttention;
6. Fix AttributeError: 'SymInt' object has no attribute 'size' to speed up recompilation after a resolution change.

MBCache extends FBCache and is applied to cache multiple blocks. More details see above given repositories.

Usage