Ok, but to give some context please also put the link to the source thread, since they are testing wf and very likely to be modified soon :
https://github.com/kijai/ComfyUI-WanVideoWrapper/issues/1519#issuecomment-3447933556

get the GGUF creator from that City96 guy and convert them to GGUF yourself but I'm about to look at this as this was my exact idea and already have it built I just need Co-pilot to create a node that automates it for me as these nodes rarely have what I am trying to accomplish so one must get savvy with python if you ever want to break from the weirdness.

https://github.com/city96/ComfyUI-GGUF/tree/main/tools

I have 10 SVI LoRAs. Two of them are 2.1, the rest are 2.2. Search for SVI PRO.

https://github.com/vita-epfl/Stable-Video-Infinity/tree/svi_wan22?tab=readme-ov-file#-model-preparation

This sounds like it would work. Do you have a workflow you can share?

@evert_guy I have a FLF workflow that uses SVI posted on my models page, specifically for capping off long generations using this method. Go to my models page and download that and take a look at how I set up the first input to grab end frames and send them to an encoder. You easily adapt this to encode, say, 81 frames from the end and send it to SVI node as your previous latent. Don't do what I said above with the anchor. Instead, send the LAST frame of the video to the encoder that sets your I2V input image- that's your new anchor (NOT the previous generation's starting image latent- that will make the video start back at the beginning) The SVI encoder combines the anchor (incoming image) and prev (incoming motion) into one embed, because you can only have one embed, of course. You can also pull both inputs from the same encoder - depends on how your setup is configured already, whatever is easier. This stuff gave me fits at first, but I've got it down I think. One thing you should definitely do: put a save latent node onto all of your workflows and make a folder for them. That way anything you want to extend in the future, you don't have to worry about re-encoding. I'll be posting my own 4-stage SVI workflow that does this, I'm still finishing the automated prompt/LoRA switching, which is a monster, stupidly way too big. But I will post it. I think instead of incorporating the above into it, I'll make a little WF specifically for making input latents to use in the main one.

Hope this made sense. If it didn't, I promise you it does work, and you can make it work despite my babbling. I believe in you soooooo much. Like, a lot.

@Ponder_Stibbons Thanks so much for the reply, I'll look through the workflow and see if I can make it work.

Sounds like what happens when you just use the last frame image to generate the next video. Thought SVI was suppose to avoid that.

@Plasmadose That also happens, at least for me, when I use this workflow. I will test other workflows with SVI models in the future. But thanks for your answer.