Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

WAN 2.1 I2V/T2V. I tested the rel L1 thresh values of 0.12, 0.25, and 0.3, but none of them could accelerate the process. However, setting FirstBlockCache to 0.1 can double the speed, but the quality drops significantly. #8

Open
zaitaxiang2012 opened this issue Mar 4, 2025 · 14 comments

Comments

@zaitaxiang2012
Copy link

WAN 2.1 I2V/T2V. I tested the rel L1 thresh values of 0.12, 0.25, and 0.3, but none of them could accelerate the process. However, setting FirstBlockCache to 0.1 can double the speed, but the quality drops significantly.

Image

@lldacing
Copy link
Owner

lldacing commented Mar 4, 2025

Image

#7

@zaitaxiang2012
Copy link
Author

Image

Image

Image

There are obvious quality issues with TeaCache, and FirstBlockCache didn't improve the speed.

@shiraishi-png
Copy link

Image

Image

Image

There are obvious quality issues with TeaCache, and FirstBlockCache didn't improve the speed.

Did you try using the "ApplyTeaCachePatchAdvanced" node?

Image

I don't know what value to set "start_at" but I assume the quality drop is caused when the value is set to 0 or a not high enough number, because kijai nodes come with start_at set to "6".

@lldacing
Copy link
Owner

lldacing commented Mar 4, 2025

try add steps to 30.

@lldacing
Copy link
Owner

lldacing commented Mar 4, 2025

start_at and end_at is range in 0-1, like control's start_percent and end_percent

@lldacing
Copy link
Owner

lldacing commented Mar 4, 2025

For i2v, it seems like the width and the height of the node WanImageToVideo will effect quality of the TeaCache in my test.

@Djanghost
Copy link

This node is still unusable with Wan, Kijai's version of teacache remains the most efficient for the time being.

@lldacing
Copy link
Owner

lldacing commented Mar 4, 2025

This node is still unusable with Wan, Kijai's version of teacache remains the most efficient for the time being.

It seems like kj used another way of TeaCache, I will try it tomorrow.

@kakaxixx
Copy link

kakaxixx commented Mar 5, 2025

https://github.com/ali-vilab/TeaCache/tree/main/TeaCache4Wan2.1 Can author take a look at its implementation method?"

@lldacing
Copy link
Owner

lldacing commented Mar 5, 2025

https://github.com/ali-vilab/TeaCache/tree/main/TeaCache4Wan2.1 Can author take a look at its implementation method?"

thks, I will check it later.

@lldacing
Copy link
Owner

lldacing commented Mar 6, 2025

Submitted, please update the latest version.

@zaitaxiang2012
Copy link
Author

Thank you. I will provide the latest test results.

@zaitaxiang2012
Copy link
Author

First of all, thank you for your efforts. However, after testing, the video quality still cannot be satisfactory.

@lldacing
Copy link
Owner

lldacing commented Mar 8, 2025

First of all, thank you for your efforts. However, after testing, the video quality still cannot be satisfactory.

Did you try change wan_coeffiecents?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

5 participants