Thanks for making this, camera control in general is super useful

How many steps did you train? Did you use diffusion-pipe or musubi-tuner?

God bless you

I notice same video has 5 seconds another 3 seconds Why?

You should now make a slime effect where something turns into slime and gets pulled apart.

Would it be possible to make a variation of this where the background also rotates?

can you make a lora, the subject is moving or walking and the camera rotates 360 around it

It might be useful for image-to-3d. Just need to use photogrammetry with it.

I noticed a lot of the time I don't get the full 360 degree rotation (usually its about 270 degrees). Does anyone have any tips on how to fix this?

it performs best when you actually describe what is happening in the image. otherwise the subject spins and twists but not like a "360 microwave" :D

WanVideoSampler

'NoneType' object is not callable

how can I fix this? already 2 days and I can't find a fix!

Doesn't work with native ComfyUI workflow.

Did videos used for the training had the same frame count ? Because as it is a true 360 only happens by pure luck.

Any training code we can look at? I'd love to fine tune this thing further

So that did work for me, but the character in the image is moving, making some movements, i couldnt make a static 360 of the character, any hints ?

Works decently. I'd like to see different versions, like one for rotating vertically, and one for a full 360 camera movement where the subject is stationary. If you look up the term "stereograph" or "wigglegram", that would be a decent style too (basically it's just 2 still images used to quickly tilt the view to give a perception of depth). That would pretty much fill in the general possibilities. I think some training could be added that also considers out-of-frame elements, because currently it doesn't really know what to do with subjects that are partially clipped from the image.

I need help with this... I get an error on "Load WanVideo T5 TextEncoder" Node, my settings are:
model_name : umt5-xxl-enc-bf16.safetensors
precision : fp16
load_device : main_device
quantization : disabled

My best wan lora so far.

720p i2v variant would be nice. (Upscaled 480p Looks not so realistic)

Hey everyone 👋

It does a really cool job generating 360° rotation videos from a single image — but I’ve hit a wall and hoping someone here might have some insight.

The Issue:

Since it only uses one input image (usually the front view), the AI doesn’t actually know what the back of the statue looks like. So, it just “hallucinates” the backside, often in a way that looks totally different from the actual statue. In my case, the result is quite far off from the real design.

What I’ve Tried:

• I provided a clean front image of the statue.

• The generated rotation video looks great from the front and sides, but once it gets to the back, it’s a different design entirely.

What I’m Hoping to Achieve:

1. Is there a way to provide more than one image (like front + back) so the model understands what the full 3D object looks like?

2. Would training a custom LoRA of this statue help? For example, feeding it multiple angles (front, side, back) so that it learns the actual geometry and doesn’t make stuff up?

has anyone thought of using the frames of these outputs to train loras actually seems incredibly useful

Could you please provide this example prompt for testing?

Amazing LoRA dude, and you did this on 30 seconds of videos? You mind sharing some more info? Any images, # of repeats, resolution, # of frames to process (chunk 17/33/49, etc.). Would love to get some training advice since there is so little out there for WAN.

Edit: before downvoting this, read my most recent comments near the bottom. PS: stop downvoting people just because you're too numb in the head to take the time to see where they're coming from.

For as basic as video loras are at this point, how is this one of the top downloaded? No offence intended, but concepts don't really get much more basic than turning on the spot.. It's not a hit on the model, or the uploader, just curious how there's DBZ action loras, and this is the most downloaded, are people(in general) just really boring, or what?

Edit: Ok, I just got my first 3D printer, and now my eyes are fully open. This probably is the most practical one on here. The key is that it's not necessarily for making amazing videos, but for it's practical use in contrast to the others. Since most consumer hardware can't even make videos with any practical use as videos, this one does stand out for what it does.

I'll give it a thumbs up, and probably try it out myself sometime shortly.

anyone having vertigo after seeing the examples? my head was spinning

My model rotated, but it did not rotate 360 ​​degrees, but only rotated 120 degrees. I set it to 360 degrees. How can I solve this problem?

Rad. Can you point me in the direction of a lora training workflow using a video dataset?

Thank you for lora. How about license? Can it be used in commercial generations?

can right rotation?Almost output left rotation

1.3B t2v wan 2.1 vace?

can we have a wan 2.2 version?

Only ever get about 270 degrees with the settings you've recommended - never the full 360. Adding frames or rephrasing it as fx 720 degrees makes zero difference. Could you provide some sort of solution to this? I see a number of other people having the same problem.

can you do one for wan 2.2

Hello, I am interested in learning more about LoRa. Could you please share insights on how you trained the model, the amount of data used for training, and the duration of the training process? Thank you!