ComfyUI_Patches_ll please support Wan2.1 video model

please support Wan2.1 video model..thanks

Mar 01 '25 05:03 rakib91221

Work in progressing. I need make about 100 wan videos to get coefficients of the TeaCache. I am looking for a suitable cloud service.

Mar 03 '25 02:03 lldacing

Please update latest version, I have commit coefficients based on 36 wan videos.

Mar 03 '25 12:03 lldacing

Thank you @lldacing What are the recommended settings and how it should be implemented with the native ComfyUI nodes?

Mar 03 '25 16:03 Hearmeman24

Please update latest version, I have commit coefficients based on 36 wan videos.

So for it to work correctly you need to generate more videos? Should this work for any of the 4 models afterwards and even if the are ggufs? Because the firstblock does something for me with i2v 420p Q4_0 but at residual = 0.15 its extremely blurry but faster and teacache doesnt seem to really work for me for some reason no matter what residual?

Mar 03 '25 16:03 wsbagnsv1

Okay seems like the issue with Teacache was some bug with another node, but its gone now so yeah

Mar 03 '25 18:03 wsbagnsv1

Thank you @lldacing What are the recommended settings and how it should be implemented with the native ComfyUI nodes?

Only native ComfyUI nodes are supported. You just need add VideoForwardOverrider and ApplyTeaCachePatch node after UnetLoader node, and can see my inadequate test result on ApplyTeaCachePatch's tooltip.

Mar 04 '25 00:03 lldacing

Please update latest version, I have commit coefficients based on 36 wan videos.

So for it to work correctly you need to generate more videos? Should this work for any of the 4 models afterwards and even if the are ggufs? Because the firstblock does something for me with i2v 420p Q4_0 but at residual = 0.15 its extremely blurry but faster and teacache doesnt seem to really work for me for some reason no matter what residual?

Generating more videos is to get more stable coefficients(The first block doesn't need coefficients.). I generated the coefficient based on the t2v1.3B model. Other models have not been tested. GGUF can also work, just adjust the residual appropriately.

Mar 04 '25 00:03 lldacing

my test result

Mar 04 '25 02:03 lldacing

my test result

1.3B model not working

Mar 05 '25 06:03 rakib91221

my test result

1.3B model not working

Need to optimize the implementation of TeaCache, It's unstable. Doesn't FirstBlock work either？

Mar 05 '25 15:03 lldacing

Please update the latest version. Have optimized the implementation of the TeaCache for wanvideo

Mar 06 '25 10:03 lldacing

Please update the latest version. Have optimized the implementation of the TeaCache for wanvideo

can you tell me please what wrong this node

Mar 07 '25 04:03 rakib91221

Please update the latest version. Have optimized the implementation of the TeaCache for wanvideo

can you tell me please what wrong this node

When wan_coefficients is set to disabled, the rel_l1_thresh value is set to about 0.025
When wan_coefficients is set to other values, put the mouse on rel_l1_thresh and you will find some recommended values. @rakib91221

Mar 07 '25 08:03 lldacing