Very very very simple workflow. All you need to do is upload ref image, ref video and a background image. It has a auto prompt using Florence2Run for both background image and ref image.
You can also add your own prompt ofcourse.
It has 3 different preprocessors.
It is a native workflow, so you can use gguf or safetensors.
You can do 4 steps, or 7 steps, i do not see much difference except 4 steps is faster lol.
Can be used for sfw and nsfw.
Wan FusionX Vace ggufs: https://huggingface.co/QuantStack/Wan2.1_T2V_14B_FusionX_VACE-GGUF/tree/main
Wan all gguf's (vace included): https://huggingface.co/calcuis/wan-gguf/tree/main
Vace gguf im using: https://huggingface.co/calcuis/wan-gguf/blob/main/wan2.1-v2-vace-14b-q8_0.gguf
FusionX Vace safetensors: https://huggingface.co/QuantStack/Wan2.1_T2V_14B_FusionX_VACE/tree/main
Lora - Lightx2v all versions: https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Lightx2v
Lora: CausVid_14B_T2V_lora_rank32_v2: https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan21_CausVid_14B_T2V_lora_rank32_v2.safetensors
Don't forget to press the [Like] button 🙂
Description
FAQ
Comments (8)
Workflow looks good. I downloaded the workflow but not all the models yet, because it would take long time for me. So I want to ask this first: Does the reference image work for 2.1 only or 2.2 as well? Because I thought it would need VACE and I see it only for 2.1.
Reference image, and video work for both wan2.1 and wan2.2.
The current wan2.2 vace gguf models have issues with preprocessors.
Here is a wan2.2 model (safetensors) - 32gb big, Lightx2v V2 Wan2.2 Vace - 3 steps - v1.0 | Wan Video 14B t2v Checkpoint | Civitai
It is not for people with low ram tho. You need 16gb or higher to use this model but it works perfectly with preprocessors.
OmegaWPNÂ Thanks for the reply. I have 64gb ram but only 12 Vram if you meant that maybe. Now that I know it works with 2.2 I will just dl and try. Thanks again.
OmegaWPNÂ Sorry to bother but I got one more question: I replaced the UNET loader wih normal loader to load the checkpoint you did link. Since it is only one file I assume it is merged HIGH + LOW model all in one. Your workflow uses High and low separate. Should I disable the second one of them and add the steps count to the first one then? Or is it a vital thing to have steps split up even for a merged model?
bowiba1265909Â No, you dont need to 2 ksamplers high and low. Disable wan2.2 ksamplers, then enable wan2.1 ksampler and attach the unet loader with the Lightx2v v2 wan2.2 model, connect the unet loader to the sageattion node and you are done.
Sorry, i mean 16gb vram or higher but dont worry, native workflow has blockswap, and teacache so if you get out of memory error then use them.
If you want to use blockswap and teacache then connect them in this order.
WAN Model → SageAttention → BlockSwap → TeaCache → LoRA → Model Shift
OmegaWPNÂ Thanks, that works! I do not use SageAttention or any of the others. But it works and I only have 12 GB VRAM. But one question remains now for me. How to use the 2.2 loras correctly as they come in pairs high+low. Now I only have the node "Power Lora Wan2.1". Should I load both loras version high+low in that node? Or maybe just one of them, if so which one high or low? Or maybe it is better to add more lora loader nodes and load them in a row/chain?
I hope I am not overdoing with my questions. I just transitioned from forge UI to comfy UI and it is still confusing a bit. <3
bowiba1265909Â Low_noise wan2.2 lora models should be used only. Do not use the High_noise with the Lightx2v V2 wan2.2 model or you will get ghostly, weird videos lol.
Note: Using too many lora's can break the preprocessor. For example, if you are trying to create a video with lip sync, it will not work if you have 4 or 5 lora's enabled.
OmegaWPNÂ Good to know! Thanks once again.