V2

V2 is more consistent, has more stable movements, and should get less artifacts. Seems to work very well for 2d inputs as well. All previews were prompted with one prompt for both t2i and i2v, writing separate prompts and picking a good starting image should give even better results.

Use "turbo" lora for high-quality generations in just 4 steps!

The turbo lora is available on huggingface: https://huggingface.co/Kijai/WanVideo_comfy/blob/main/LoRAs/Wan22-Turbo/Wan22_TI2V_5B_Turbo_lora_rank_64_fp16.safetensors

To use it, set steps to 4 and cfg to 1. I'm not sure what the recommended sampler/scheduler is but I've had great results on multiple samplers and schedulers. I personally use euler/euler a + beta scheduler.
Using a slightly lower resolution (but not low enough to reduce quality much) I can generate 80 frames in just 2 minutes on a 3060.

This lora is recommended for i2v, but t2v might work decently as well.

Trained on my new mixed furry/human dataset with detailed captions. Older versions of which were also used for the experimental and semi stable text to video loras.

Prompting

Prompting should use natural language, you need to generate at 720p, so for example 1280x704, 704x1280 or 960x960 will be valid. This might be more important for i2v than for t2v, I've noticed artifacts with i2v.

In a prompt, you can describe "a 3d animation", "a 2d animation" or "a real video", this is most useful for t2v but could help i2v as well.

You can view the prompt on the example videos for info as well.