What up coomers <3 It's been awhile since I've posted anything over in the Hunyuan thread. I felt like I'd hit a wall with progress for Lora improvement. Luckily, Wan2.1 (t2v + i2v) was released...and wow. It blows Hunyuan out of the water in pretty much every application. Not only that, but the diffusion-pipe was updated to support training almost immediately and the datasets for Hunyuan/Wan are interchangeable.
I trained my first Lora last night - ga1n3rb0t_wan2.1 version 1.0. I've been testing it out today and I am more than pleased with the results. I'm not exaggerating when I say, 99% of generations are keepers. Wan2.1 has exceptional understanding when it comes to prompt adherence and appears to train much easier than Hunyuan.
My current Lora is trained for the Wan2.1 14B T2V model which can generate 480p or 720p videos. I have no idea if it would work for the I2V or 1.3B models, it's possible, but I haven't tested it out.
JSYK: For local generation, like Hunyuan, Wan2.1 is very resource heavy and generation times can take a few minutes. However, new workflows are quickly being released to accommodate lower-end consumer cards. I've seen some for 6GB of VRAM already. Keep your eyes on CivitAI, I'm sure you could find something that'll work for your setup. Like with Hunyuan, to experience this tech to it's full potential, I recommend renting a runpod.
As always, I plan on releasing this Lora publicly ASAP. Unfortunately, CivitAI has not yet updated their models to include Wan2.1, so I am just waiting on that. As soon as it becomes available, I will upload it there and create a post here with a link to the page. Until then, I am happy to take prompt requests which I will generate and post here.
Finally, here are some examples of what I have generated so far. I'm excited to hear what you think. I'll continue posting throughout the day as generations come in. Thanks for the support.
-drewski
