This is a repost from my page at OpenArt.ai.
Workflow is available in ko-fi (click here)
๐ฅ Video demo link
๐ Tutorial (incl. resources): ย https://youtu.be/4826j---2LU
SAMPLES
(I cannot make them bigger because of the 5MB size limitations!)
What this workflow does
You can easily transform any boring video or clip and create a fantastic animation, completely change objects or change scenes. It is ideal to create reels in Instagram or Tik Tok videos.
It is a relatively simple workflow that uses the new RAVE method in combination with AnimateDiff.
Workflow development and tutorials not only take part of my time, but also consume resources. Please consider a donation or to use the services of one of my affiliate links:
๐จ Use Runpod and I will get credits! ๐ค
Run ComfyUI without installation with
๐ ThinkDiffusion
๐ RunDiffusion
How to use this workflow (download)
The tutorial contains more details in how to get the different resources
Download the workflow and the resources (videos) if you want to follow the tutorial
Install custom nodes via custom manager
Install required models (checkpoints, ControlNet, AnimateDiff)
Update ComfyUI and restart
Check that the models are properly loaded (check that the names are the same as your model names)
Do a test run with a low number of frames (16) to check everything works
Adjust your settings, change models, add CN's, add IP Adapters, etc., ย until you get satisfactory results
Run the complete animation: beware of VRAM use of RAVE KSampler.
Tips about this workflow
This workflow is setup to work with AnimateDiff version 3. For other versions, it is not necessary to use the Domain Adapter (Lora). You only need to deactivate or bypass the Lora Loader node.
Unsampling originally works with Depth, but with LooseDepth you can achieve bigger transformation of objects and scenes.
The workflow works good with other controlnets besides ControlGIF.
IP Adapter also provides nice consistent results when applied with the ย first image generated from the RAVE KSampler.
I did not manage yet to get it working nicely with SDXL, any suggestion/trick is appreciated.
The RAVE Ksampler also uses quite VRAM. I manage to process 96 frames with a 4090 24 GB with SD1.5 type of videos.
Instructions for Openart. AI runnable workflow
1- Load your video and do not use many frames. You may want to start rescale to 0.5 for HD videos. Do not process too many frames
2- Change/use the right models/checkpoints: Realistic vision (checkpoint), Depth or LooseControl and AnimateDiff_controlnet (instead of controlGif) for controlNet, v3sd1.5 adapter (Lora).
3- Change the prompt and play with it!
Additional information
RAVE method: ย https://rave-video.github.io/
Rave ComfyUI node: ย https://github.com/spacepxl/ComfyUI-RAVE?tab=readme-ov-file
AnimateDiff Evolved: ย https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved
AnimateDiff version 3: ย https://github.com/guoyww/AnimateDiff