Seems to download Controlnet processors on each run? about cog-comfyui

[ParthGandhi]

I'm also sporadically seeing model errors:

Executing node 1234, title: IPAdapter Unified Loader FaceID, class type: IPAdapterUnifiedLoaderFaceID
!!! Exception during processing!!! Error while deserializing header: MetadataIncompleteBuffer

[fofr]

Do you have a prediction link for the error?

[ParthGandhi]

Its a private fork but heres the entire stacktrace:

REDACTED

[ParthGandhi]

Ok i've run into another issue. Just summarizing my questions:

1. Is there a way to avoid redownloading the controlnet files each time?
2. Why do I get sporadic errors?
3. I'm finding the Face Analysis nodes nodes to be very slow at 30-40sec/image. Any idea why?
4. This cog (both yours and my private fork) appears much much slower than it should be for anything more than a basic workflow. I can't seem to figure out why.

For (4), the example workflow using dreamshaper-lightning at 8 steps and 2cfg takes 7sec vs 5sec on my A10G (24GB VRAM). I'm only including prediction time.

Thank you for the help!

[fofr]

My guess with the face analysis nodes is that you are running them on CPU rather than GPU.