@definitelynotadog I'll keep an eye in the event you do! thanks

@definitelynotadog wow, that's great news 🥰 When WAN came out, I did a quick test without looking at anything, with Pinokio, and it took 90 minutes to make a horrible 4s video 🤣 5-10 min per video, it's God. And seeing your videos, it's really worth investing in the R&D of learning 😊 By the way, all this is img2video, right?

I run Wan(and train it) on 16GB(4070 TI SUPER), about 28 mins for 960x720@5s I2V with fp8 scaled from Musubi, torch.compile, and fp16 accumulation(I don't like teacache but if you don't mind it you can go even faster). I2V is incredible, definitely give it a try! Make sure to use SageAttention and fp16 accumulation for optimal speed. Avoid fp8 fast as it degrades quality too much with Wan, stick with regular fp8 or fp8 scaled. If you wanna use Comfy, I recommend Kijai's WanVideoWrapper. If you like CLI like me, I recommend Musubi Tuner for both inference and training. I maintain a fork with a branch containing extended features like fp16 accumulation, latent preview, etc, here:

https://github.com/Sarania/musubi-tuner

Kohya and I are working on an extension mechanism so that I can more easily interface them with the official code because I've been working on it pretty prolifically, but currently it remains unofficial! I use the fp16 base models(which we found maintain quality better than the bf16 ones) with fp8 scaled quantization, though the version I use is only available in Musubi. You can also just use pure fp8_e4m3fn. You're other option is native ComfyUI and using GGUF quants like @definitelynotadog mentioned but I don't personally have experience with that.

@blyss wow, thank you! I have a lot to learn. I haven't even used Comfuy for images yet. I've stuck with SDXL and Forge. Thanks to both of you for the work you do! 💖

@ai_isern Even for someone like me who is knee deep in the muck of this stuff, there's just too many new developments every day to fully keep up! Furthermore, unfortunately so much of the knowledge remains "if you know, you know, if not, good luck finding out" so I try to spread around the things I do know as much as possible.

@blyss You're absolutely right. It's so much appreciated that there are people like you. People like you are the lighthouse that illuminates the world 😊😘

I have a RTX360 12gb. I use the rapidWAN22 I2V GGUF Q3 model. It works great. 5min for a 10sec 768x760 video.