Yo so I've been messing around with AnimateDiff in ComfyUI for vid2vid and here's what I've figured out so far.

AnimateDiff was open-sourced back in July 2023, and right now it’s one of the more solid options for AI video generation. I mainly go the ComfyUI route because you can hook up ControlNet for vid2vid, and also use Prompt Scheduling to swap prompts mid-clip—two things the A1111 plugin just can’t do.

Let’s get the hardware requirements out of the way first: vid2vid needs 10G VRAM, pure text-to-video can kinda scrape by with 8G. When setting it up, you’ll need to git clone a few custom nodes, then hop into Manager to install Auxiliary Preprocessors and FizzNodes. For the motion module, go with the v1.5 v2 one; any SD1.5 base model works, but for anime-style stuff I usually slap on DarkSushiMix.

A few parameters worth remembering: keep context_length at 16 for the most stability, context_overlap defaults to 4—so it groups the first 16 frames, then 12 to 28, overlapping like that to keep things smooth; context_stride in theory makes the image more consistent, but in practice it just makes the footage jittery, so just leave it at 1. Crank denoise to 1 for max effect, lower it if you want it closer to the original video. For output, go with h265-mp4 for the best quality. Oh, and the generated videos don’t carry ComfyUI’s workflow metadata that you can drag back in, so if you want to save the workflow, remember to enable save_image to capture one frame.

I actually figured out that 16-frame number through trial and error — anything higher or lower just doesn’t work as well.

10G VRAM for vid2vid, is that gonna be super slow?

Yeah it’s slow, but if you drop the resolution to 640x360 it helps a bit.

The best part of this setup is being able to swap prompts mid-generation with prompt scheduling.