LTX 2.3 Remaster any Video
This workflow allows you to remaster some existing video, you can specify an driving video and it is used as input for the generation. You can also specify up to 4 keyframes to drive your video generation (easy to extend)
LTX 2.3 Extend any Video
This workflow allows to extend videos and stitch the new part seamless onto the existing video.
LTX 2.3 Sound to Video
Added a basic LTX Sound to video workflow that can be used for music videos. Or other use cases.
LTX 2.3 Bounce V 2.5
Added a basic LTX workflow to use the bounce 2.5 lora.
Better Details due to Detailer lora
Referenced new Distill Lora 1.1
LTX 2.3
Added a basic LTX workflow to use the bounce lora
WAN I2V 2.2 / 2.1 - Resume Video + Upscale + Interpolation
ChangeLog
WAN 2.2: Added quality workflow for liquids. Slow but good quality!
WAN 2.2 Lightning 8 Steps added as reference WAN 2.2 workflow with lightningX!
V1_0_6_2_Quality added, since the causVid version did not work well with my newest "big splash" lora...
V1_0_6_2: Replaced the Int node with comfyui base because it caused some troubles for some users
V1_0_6_1: with the causVid lora generations times of < 1minute possible,
Since a lot of ppl were asking how my personal Workflow looks like here it is! It has some handy features like inbuilt upscaling and continuation of videos with the last generated frames. Don't use more than 1-3 continuations because the Quality loss is quite huge.
It's optimized for a 4090 and 24GB of Memory. But by playing with the params less is possible.
Input Modes
Image -> Use a input images selected by the user
Image Directory -> Set a input directory and load with every run another image as source
Resume Last Frame -> Load the last frame of last generation as input image. Ensure you have done one before
Resume From Video Last Frame -> Select an input video and resume from the last frame. This is helpful if you want to do several continuation attempts of one source video
Quality Options
Face Fix-> Uses ReActor to fix the face. I turn this mostly off for wan because its already good enough
Upscale -> Upscales the output to maximize the quality
Remarks:
The ReActor Node is optional, i have my own modified version with disabled NSFW check, you can add any face swap node there, but wan is really good with faces so its not needed
Description
Replaced the Int node with comfyui because it caused some troubles for some users
FAQ
Comments (42)
is it possibe to use this for t2v?
Great workflow.. nice high res results that are fairly fast. Only issue I couldnt select fp16_fast, had to use fp_16. Am I missing out on extra speed?
Seems the fast version requires pytorch 2.7.0. Not sure how much speed it reduce with the fast version.
I've upgraded to pytorch 2.7.0 and it works well! I think I saved 200secs. Commands under python_embeded folder:
python.exe -m pip uninstall torch torchvision torchaudio -y
python.exe -m pip install torch==2.7.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
@ItsThatTimeAgain what about if you have comfy portable... will this work?
@seductivelyai695 it should work for portable comfy, because it has an embedded python folder. You have to open a command terminal inside that "python_embedded" folder. Just right click inside the folder and select "open command prompt here" then enter the python command he wrote in his post.
@Griphen116 I decided to go crazy updating latest version of cuda, pytorch, nvidia drivers etc... Ouch that was a mistake as I found my sageattn no longer worked with the new version and somehow my some of my libs got corrupted. Like a dummy, I had to reinstall comfy portable and start over. But yea, finally got it working. my pytorch was 2.6 previously, I probably just should have followed the op's advice and just upgrade that only, instead of going to latest versions of everything.
if you are using comfy portable do the following from the comfy home folder.
.\python_embeded\python.exe -m pip uninstall torch torchvision torchaudio -y
.\python_embeded\python.exe -m pip install torch==2.7.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
From my tests, it seems the low_mem_load toggle didn't affect the render time. Caus Vid lora at 0.4 has great color retention from image preview, but seems to take away the other lora strength. Aka, less boob physics. Setting the Caus Vid to 1 drastically made the video high contrast and had some noise.
Also, making the Caus Vid lora strength to 0.55 instead of 0.5 or 0.6 will make the render fully noisy. Doesn't seem to like this amount of precision.
So for my settings, I've made the strength to 0.6, augmented the steps to 8 and added a Color Match Image node between the Color Correct node and the Video Combine just to have the same color as the reference.
On my RTX 3090, the Sampler took 204 secs at 6 steps and 350 secs at 8 steps. 6mins way better than 20mins.
Are you noticing it takes a long time to load the model and loras? If I use 3 loras, its taking about 2.5 minutes to load , before it can even render.
@seductivelyai695 Yeah if you look in the cmd, it compiles the weight of each loras. I think this is due to how the Model Loader node is optimising the number of vram with the BlockSwap. Might be able to take out the BlockSwap for faster render but more vram usage?
The repository reactor node has been disabled... options?
I just bypass it, seems to work fine
Use the one that ships with the NFSW check: https://github.com/Gourieff/ComfyUI-ReActor
As noted in the description of this workflow, the check can be disabled.
Thanks for your workflow. There is a node for color balance that I couldn't get to install so I replaced it with KJ Color Match (in : ref out: pic). But there is a node that is inside normal output where it is called "-int" with a value 57 and it gives as missing or something wrong, I tried replacing it adding another "-int" and connecting again but then I cannot change the value and ends up as N/A in other areas of the flow. Any ideas how to fix it? ComfyUI Manager tells me that ComfyUI-Logic is missing but it is installed without errors.. weird.
check out my latest version have replaced the int node
Super great!!!
It has saved my time by nearly 300%.
I want to give you 1000 likes. LOL
Any chance of a first frame/last frame workflow? I'm too dumb to do it myself lol
I am having trouble installing the reactor face swap. The problem with this is that it just doesn't install the nightly build only older build which seems like it is not compatible with the workflow. When i try to go into the reactor github page manually it seems to have been taken down. Does this mean this workflow is already outdated?
have the same. every time i have to delete directory of this node and then install, an restart
Just replace the node in the worklfow with your version of it. There is no special requirement to the faceswap node
@ai_build_art cant find it in adding note, isn't it better i make the nightly build to work to download in comfyui manager instead? why doesnt it work it says security only allows default channel or something. I think i need that exact node
@ai_build_art maybe comfui manager is not updated, how to update it without updating other shit? i am stuck in loops with this error and no progress, please help ?
@Tartrixroar Well if it causes you troubles just ignore it. Im not using it in 99% of my cases
@ai_build_art how to i make it ignore? just run it as it is you mean?
@Tartrixroar you can ignore nodes in comfyui simply by bypassing them (right click on node and bypass)
Its super great and build well, thank you for sharing!
I am missing the custom node 'int-' and cant seem to fix it, any suggestions?
You can replace the missing node with "Get Image Size & Count". from ComfyUI-KJNodes
Connect the WAN Decoder's Images output to Get Image Size & Count.
Take the "count" output from that node and connect it to the "a" input of a Math Expression node.
Set the expression inside the Math Expression node to: a-1.
Basically, you need to get the last frame from the generated video.
Strange, the nodes before the Math Expression are just a video helper suite video loader and a reroute, there is nothing more
I think i found it. The int node can easily be replaced with a int node of comfyui ill update the workflow with it
Thanks!
I've been using this workflow a lot. Solid workflow, nice results. good job.
FYI the title has a typo - it says"claus"vid
Haha i also like claus
is that workflow able to work with 480p wan model? Im trying to switch the model but it doesnt work for me, generaton stuck on 0 step no matter how long i wait
i have rtx3080 10gb and using other workflow with 480p model and it works fine, but i want to try some of the faster workflows
Got me some error when I try to run it:
WanVideoImageToVideoEncode
WanVideoVAE.encode() got an unexpected keyword argument 'end_'
Any ideas fellas?
"It's optimized for a 4090 and 24MB of Memory."
Amazing! Can it run crysis with the 24552 MB left ?
lol first one reading the text - fixed ;)
Any way to fix this error?
Can't import SageAttention: No module named 'sageattention'
The secret to solving this cryptic and indecipherable error is to... install sageattention. You can paste the error into a site named google and it will bring up lots of guides on how to do just that. This internet place is wild.
If you are using the portable version of ComfyUI make sure you install sage attention to the python virtual environment not to you regular environment. ChatGPT can explain how to do this.
Common place people get mixed up at.