I second this, Everyone seems to be using that now instead of Hun

wan wan wan wan, whats going on with wan why is everyone there?

@Devilday666 Use both for their respective strengths. This can be done just fine on HY, no need for WAN in this instance IMO. Unless you have a specific use case in mind?

Wan actually isn't usefull i think. Even hunyuan BF16 models is too much for a 3090 or 4090. Offload to Ram make generation speed very slow when it happend. So 32 GB Wan is even worst if u want accomplish really good video with WAN.

@dayman02400741 thats not true it takes me about a minute per second of good video of Wan on a 4070ti

@ginx hey, friend! could you share your wan workflow and the checkpoint you're using? I have a 4080 and it takes me quite a while to generate a good video. Maybe I'm doing something wrong?

@dayman02400741 I run hunyuan BF16 on a 12GB 4070, without using block swap, at around 432x768x85f. Just wanted to let you know it's possible, if that's helpful. (My setup: linux, sageattentionv2, torch compilation, 0 to 8 loras, enhanced clip model from zer0int, llava llama fp8 scaled model, KJ nodes for main model loading/compilation, not using the KJ wrapper extension [though I was previously], using the iGPU for main OS UI)

@newfangled170 the problem is the memory and it's requirements (the render time too). From my point of view it take too much time. Ofc i can run BF16 models but why ? With SA2, Triton, Torch compile it take billion year to generate a single video because models is heavy asf and is offload to Ram. This process make generation so much slower than only use VRAM. Wan 2.1 weight is 32.1 gb so if u want run it fully on vram you need atleast a 5090. When u want make a longer video u'll have trouble too. Even on Hunyuan with Fp8 weight i got OOM when i generate video near from 200 frames so imagine with Wan 2.1. When i ran a generation for make a LoRa with Wan 2.1 it took 15 mn atleast for a 5 sec video in 640x1024 . With higher res you can double this time. I tested on 1024x1024 with GGUF version too = 25 or 40 mn per generation. This models give good result for sure. But u should admit guys... It takes age to render too.