This workflow allows you to generate up to four clips with WAN 2.2 and seamlessly combine them into a longer video of up to about 20 seconds.
The workflow uses the SVI 2.0 Pro nodes and the SVI LoRA to generate smooth transitions by using the last few frames of the previous clip, as well as the initial start image, as character and environment references.
I am using an RTX 5080, but with lower resolutions the workflow should also work well with GPUs that have less VRAM.
Additional features
Interpolation and upscaling
Switching between multiple WAN base models
Color matching to the reference image
Trimming the sides of the video to remove artifacts
Trimming a set number of frames from the first clip, so the video can begin with movement
Viewing and saving the last frame of each clip
Viewing all frames in a clip
SageAttention integration
Setting specific LoRAs as global for all clips
(The SVI and lightx LoRAs are mandatory for the workflow to function.)
Recommended use
Upload an image, set a lower resolution in the CustomResolution node, and select only one clip to be generated. The CustomResolution node was made by me and works with ComfyUI Manager.
Turn off interpolation for now. Upscaling can remain enabled, as it will not have any effect unless upscaling itself is activated.
Then iterate by generating a new seed or changing the prompt and LoRAs until you are happy with the result. After that, consecutively add clip 2-4. Remeber to activate “save output” for clips you want saved.
Once you are happy with the full sequence, you can move to higher resolutions and then interpolate or upscale the final result.
Note:
You can disregard the aspect ratio setting in the CustomResolution node, as the workflow automatically adapts to the aspect ratio of the start image.
Also, only the resolution setting for clip 1 matters. This setting is automatically carried over to the subsequent clips.
Description
FAQ
Comments (17)
I made a few changes to some of the nodes to better suit my AMD card and the workflow worked great with Q8 with and without any lora's. I will tweak it some more to optimize for my use, thanks for for sharing your work!
I forgot one question: Enable Color Matching Off - switching it to ON turns off color matching or turns it on?
@bigconcat Awesome, happy to hear that and thanks for the feedback!
Yes exactly, it replaces the color matching strength with 0 when the switch is on. Had to use that as a workaround unfortunately.
Btw. can you see and adjust the small mxSlider nodes? They had an issue in the past were they didnt display porperly that I had to fix manually.
@AugustusLXIII ok, got it. Yes, I can see the sliders to control the steps. I changed them to 8/3 sigmaSplit and they worked fine., thanks.
Immediately breaks, missing node connections, i am probably too new for this one, i can't figure out what connections to make.
I only get "Required input is missing" when generating with less than four parts. But that doesn`t break the workflow. Have you made sure that all custom nodes are installed?
Firstly, thank you for your work, very well!
But after installing all Lora, i see in the ComfyUI log lots of messages like "[WARNING] lora key not loaded: diffusion_model.blocks.8.cross_attn.norm_k_img.diff"". In GitHub, some people also have the same problem ("Wan2.2 I2V lora key format standardization not fully working"). Is it ok?
You´re welcome!
Unfortunately I cant help with that issue. Have not had it myself.
@AugustusLXIII , Do you know what I need to make a full video in .mp4? There is a problem, because after generation I have 4 separate mp4 (5 seconds each), and one big (20 seconds) .gif, not mp4. Tell me, please, how to download the one big full mp4 file?
@ark777 Check if you have changed something under "format" in the green "Output: Full" node. Should be set to an mp4 format.
@AugustusLXIII Yes, you are right - in the green "Output: Full" was "gif", not "mp4". Thank you very much!
Works well, but how to turn on the upscaling? I see the outputs of my videos are not upscaled.
And the second question: can I extend my own short videos with that workflow?
And one more :) - How to use separate LoRAs not only for the 2nd, 3rd, and 4th videos, but for the first one also?
The upscaling node only works with Nvidia GPUs i believe. Do you have one? And even then it doesn´t work out of the box. I did that with the help of AI then.
No unfortunately that will not work. At least I don´t know how. Guess you´ll have to try and recreate you own video in the first clip and go from there.
For the first clip the LoRAs go in the LoRA-loader on the left of the video output for part 1. Or right above the prompt for part 1.
i tried this but i get all noise coming of the output or white if using gguf. is there somehting im missing? i htink its sage attention or torch but i bypass those and its the same so maybe the sigmas not sure. idont have any missing nodes
Hey, have do you have the lightX and SVI LoRA installed and active? It won´t work without those two. For reference you could download my video and load it into ComfyUI.
@AugustusLXIII Ty for the reply. i do get noise redonwload all loras just in case same. i htink its sage attention node i can use the yet another workflow and K3nk ones and others just thisone with the kijat patch . I selected auto to run it and it runs but if i leave it as default i get
ImportError: cannot import name 'sageattn_qk_int8_pv_fp8_cuda' from 'sageattention' (ComfyUI_windows_portable\python_embeded\Lib\site-packages\sageattention\__init__.py)
@sesaxe1993920
My only idea is that you maybe delete the sageattention nodes or take it out of the chain.