This workflow allows you to generate videos from image or just prompt using a video reference.
What to know :
- I2V and T2V
- FP8 Models
- Lora loader
- 4 steps with Light2X Loras
- 32fps interpolation
- Use either Depth Anything, LineArt, Canny, or DWPose to use a video for reference
Summary :
By using a reference video you can now generate a video with Wan using either the Depth, LineArt, Canny edge or DWPose of the reference video. In addition, by using Light2X Loras it allows you to generate your videos in only 4 total steps, you can also add certain Loras to your generation using the loaders, reducing greatly the time to generate a wan 2.2 video. You can either use an image to use as a visual reference or use only a prompt, by deactivating the image loader nodes (More information in the guide). The final video is output at 32FPS. The models used are two FP8 wan models that are each about 15gb, so the Ultra-pro with 48gb of vram and 64gb of ram option is very much advised so you don't lose time having to load the models and load the sampling.
Guide :
1) Models : 
Model 1 = High Noise : First model to be loaded, it's better for the lora of this loader to be twice the weight of the 2nd Lora loader (so 2 for 1, 1 for 0.50 etc)
Model 2 = Low Noise : Always second model to be loaded.
Shift is good at 8, you can try different value.
VAE : Always Wan 2.1 vae with 14B
Clip : FP8 is good, FP16 would probably tank your generation time, in addition FP8 quality is very good by itself.
2) Image and Video Loaders :
Load your image if you want to use I2V. You can choose your dimensions with Image resize, it is set to give the width and height to the videos preprocessing and the video generation. You will have the preview of the image along.
Load your video into Load Video, keep the forced rate of 16 and whether you want to make a video longer than 5 seconds you can change the frame_load_cap to the number of total frames of your choosing.
You can choose which Preprocessing Control you prefer : Depth, LineArt, Canny or DWPose, you can see what they do above in the given images and videos, the image resize will resize the videos to the same dimensions of your given image and you can check the output preprocessed control video with Video Combine.
3) Prompting and video Length :
Write your prompt in the CLIP Text Encode green, then choose your video length, Wan is originally output at 16fps, so 5 seconds = 81 fps, 10 seconds = 161fps.
4) Sampling :
The sampling is done in two samplers, one for high noise model, and the second for low noise model, 4 steps are enough (BE SURE TO INDICATE THE SAME NUMBER IN BOTH SAMPLER'S "STEPS" ABOVE CFG), CFG of 1, euler/simple (although you can try different sampler/scheduler), first sampler start at 0 and end at 2, second sampler start at 2 and end at 4 or 10000 as long as it's at least 4. If you want to increase the steps, divide it in two, for example 8 steps would be 0 to 4, and 4 to 8.
Output :
The VAE decode your generation which is then combined in the first node at 16fps and then send to the second node to be interpolated at 32fps.
How to use Text To Video :
For T2V you only need a few tweaks :
First deactivate by selecting and pressing ctrl+b the three nodes of the Load image group.
Disconnect the link of the width and height input of the Image Resize node in Load video group, choose your dimensions like in the image, and create output links for width and height from the node to Wan22FunControlToVideo :
Wan22FunControlToVideo will create a latent with the chosen dimensions and the samplers will generate the videos from the prompt, example of a quick t2v using the video of the cat from the cover :
(Hair are kind of a weakness of wan so a cat is not the best way to show it :(, but it works !)