Introduction

I tried to implement the first block cache optimization for the Wan2.1 models. The optimization is based ParaAttention repository.

Installation

Follow installation instructions from DiffSynth-Studio.

Usage

https://github.com/user-attachments/assets/2c2e88de-a8c1-48bd-a85c-80f8ccf858b8


from fbcache import fbcache_forward
# ...
# load models and initialize the pipeline
# ...

fb_cache_thr = 0.06
#Add FB cache to the pipeline
enable_fbcache = True
pipe.dit.__class__.enable_fbcache = enable_fbcache
pipe.dit.__class__.fb_cache_thr = fb_cache_thr
pipe.dit.__class__.first_block_cache = None
pipe.dit.__class__.last_block_cache = None
pipe.dit.__class__.forward = fbcache_forward

Thoughts

Threshold is different for each model. I tried to find a good threshold for each model by trial and error.

For Wan2.1 T2V 14B, I found that the threshold is 0.06 or 0.07 works well.

For Wan2.1 I2V 1.3B, I found that the threshold is 0.03 or 0.04 works well.

Results

Wan 2.1 T2V 14B

No FBCache	0.04 FBCache	0.07 FBCache	0.09 FBCache
1216.17 sec	1090.77 sec (1.11X)	657.64 sec (1.85X)	359.47 sec (Not usable)
Videos are available at assets.

No FBCache result

Wan2.1-T2V-14B__fbcache_0.00__1216.17.mp4

0.07 FBCache result

Wan2.1-T2V-14B__fbcache_0.07__657.64.mp4

Wan 2.1 I2V 1.3B

No FBCache	0.03 FBCache	0.04 FBCache	0.06 FBCache
227.14 sec	200.35 sec (1.13X)	167.06 sec (1.36X)	80.41 sec (Not usable)
Videos are available at assets.

No FBCache result

Wan2.1-T2V-1.3B__fbcache_0.00__227.14.mp4

0.04 FBCache result

Wan2.1-T2V-1.3B__fbcache_0.04__167.06.mp4

Wan 2.1 I2V 14B (480P)

I was not able to run it on my setup. Need testing.