fp8 for phantom?
It's a little unclear from the names, but the T2VFusoinX_fp16 is the same as T2V_MasterModel?
And i2vFusioniX_fp16 is the same as i2vFusioniX with nothing specified?
But then there's the other T2VFusioniX_Phantom with no fp8.
It's a little unclear from the names, but the T2VFusoinX_fp16 is the same as T2V_MasterModel?
And i2vFusioniX_fp16 is the same as i2vFusioniX with nothing specified?
But then there's the other T2VFusioniX_Phantom with no fp8.
Hey! Just to clarify — "master model" was just a placeholder for the original text-to-video model. I've since updated the name, which is where Fusion X comes in. As for FP16, anything without an FP number listed is actually FP8. QuantStack helped with creating the GGUF files and needed the FP16 versions specifically for that process, so that’s why I included them. Hope that clears things up!
The GGUF's are great but they can't be used with WanWrapper. I actually preferred the native flow due to compatibility with the rest of the ComfyUI ecosystem. But the other day I ended up trying a workflow and got such good quality with the WanWrapper. So I did my best to match it, and couldn't come close. Then I did a deep dive into disecting the code for the WanWrapper to see if I could extract the sampler/scheduler from it to create a custom node for the native flow. It would have been close enough to possible to do that for the scheduler, but there was just no way for the sampler. I ended up learning a lot more about k-diffusion and it's tightly coupled nature with ComfyUI. But basically WanWrapper is using the FlowUniPCMultistepScheduler, but it's stateful, and ComfyUI's sampler is stateless, so it's impossible to use any of ComfyUI's Sampler nodes so it's highly non-trivial.
It was an interesting deep dive, but all that to say, the GUFF's serve a purpose, but a FP8 of all the models is also useful if we want to use the superior scheduler that's tuned to WAN and only available in the WanWrapper.
you can find the text to video GGUF wf here
https://civitai.com/models/1663553
and i'll be posting more soon.
I’ll be releasing a FusionX LoRA today or tomorrow, which will let you use any WAN 2.1 14B model you like in any WF you like. If you join the Discord, you can check out some examples—lots happening over there! :)