Creates a 360 camera orbit given an input image. Trained with WAN2.1 I2V 14B 480p.
I trained with 256x256 res video because I can't afford the better hardware. The input videos are about ~65ish frames so that's how much you'll need to get a full orbit. It seems to work okay scaling up to larger sizes. It doesn't work for super complicated images but for basic images with a single subject, you can get a good output after a few tries. I marked it as version 0.1 because I plan to improve it more.
I use anywhere between 0.9 to 1.0 strength on the LoRA, and anywhere from 13 to 40+ steps for KSampler. Anything less and it starts not doing the full circle or moving around in weird ways.
I posted one image as a .webp that has the ComfyUI workflow that I used attached as metadata. (the skeleton bunny post here: https://civarchive.com/images/64762348)
This is my first LoRA so it might not be the best, let me know if you have any suggestions.
Example prompt:
r0t4tION orb1t A cute pink owl sleeping in a blue teaapot, with beautiful flowers in the background
Description
FAQ
Comments (6)
"I posted one image as a .webp that has the ComfyUI workflow"
You posted it where ? Only mp4 to download and the LoRA itself here. Am I missing something ?
Its attached to this skeleton bunny post https://civitai.com/posts/14433297 (see the "Other Metadata" section of that image, the Comfy nodes should be there). Let me know if it doesn't work, I can send you it directly
@samusynth All right. Totally missed this Other metadata > Nodes in the animation info. Thanks ;)
this is quiet nice lora!
however, this sometime makes the background blurry.
so, I prefer to use this as a helper for other similar loras.
for ex, you can use this 0.1~0.25 strength, so wan would understand and do the rotation even if the character is in the small room.
UPDATE: This work great thanks, it's now my favourite LoRA on CivitAi! Tested using image to video in Wan at strengths from 0.5 to 1.0 and 20 steps. Wan 2.1 needs a detailed description of the surroundings as it doesn't do a great job of recreating the background just from what was in the original image if you do a full 360 degree turn (actual rotation is around 380 degrees in total so the video can be edited at the 360 degree point and looped and would look smoother with some interpolation to double the frame rate of course). However, if you rotate around a person 360 degrees Wan will almost certainly create an extra arm as they turn so you end up with 3 arms! and other distorted limbs. The way around this is to either prompt for the person to be looking at the camera in which case they will actually keep turning towards the camera at a certain point which is pretty cool. OR, set the strenght to 0.5 which will only do an orbit of around 100 degrees creating a nice visual effect and not breaking the model by turning it too much! Thanks for making this.
Any chance of wan2.2 versions?
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.