Wan 2.2 I2V workflows that allow longer video generation by passing the last frame of one video as the input into the next. It then combines the videos at the end.
Includes Lora loader node, SageAttn node, FP16 accumulation node in addition to the standard I2V nodes.
If you can run the default Wan workflow, you should be able to run this as well. I noticed minimal, if any, additional overhead using this method.
As with any extended video, context drift is possible. This workflow is best suited for actions where major changes to the scene aren't prompted.
Description
FAQ
Comments (8)
Thanks for this workflow. The latest update works very well!
Thank you for the feedback! Let me know if you experience any bugs. I'll admit I haven't had the time to fully test V2 yet, and only ran it a few times to verify it functioned.
@ChillDesire Will do! Ran Modular V2 a number of times and so far so good :)
Much better than V1.0
Hey there is there going to be perhaps a version of multi stage having multi lora also? the reason is because while vide gen works fine , the prompt does not get registered at all sadly.
I wanted to make a 5 second handjob , 5 second blow job and then last 5 seocnds of cumming in mouth, sadly without multiple lora to be added at each step that might not be doable and even without it there's no way to make sure the 2nd and 3rd prompt are followed after the first one
For the prompt issue: Are you saying you have 3 different prompts, but it's only using the first prompt on all 3? If so, I'll work on debugging that.
For multiple Lora loaders: I'll investigate the feasibility of that. My fear is that it will try to load 6 models into RAM instead of the 2 as it is now. Let me do some testing.
Ok, I did some digging on the prompt issue and may have a fix. It may have to do with caching and the way I used the bypass nodes. I beleive I can correct it.
If you can share any additional information/details on the issue you're having, it will certainly be appreciated!
Regarding the Lora loader for each sampler pass, my suspicion seems to be correct that it will load a whole model for each sampling block, resulting in the need for 6 total models to be loaded instead of the usual 2. While ComfyUI will likely handle the loading and unloading, the performance hit could be seconds to minutes of added time depending on drive speed.
I'm not sure that change is in the scope of what this workflow is trying to achieve, but if the requests are there I may add that feature down the road.
Appreciate your feedback! It's nice to hear the loves/wants/dislikes of these workflows to help make them adaptable for as many people as possible.
@ChillDesire Hi sorry for the late reply , so basically whatever i put in the first green box / prompt is the only action that takes place even if i do add other loras that I have , an example being i wanted to do abit of nipple being squeezed , that worked great no issue but in the last prompt i wanted his hand to just open and fondle the breast but it remained doing the nipple instead.
I did get it to work by increasing cfg but that ruined the entire image's coloring and etc sadly so yes you're right it would require a decent rig but its definitely worth it since you won't have as much of a color degradation issue that usually happens if you manually take the last frame of a video and try to do this.
this workflows nails the perfect transitions between vids, almost 0 color change or motion transitions issues. between vids The only issue left is that whatever is not visible in the last frame will change to something else in the next vid, but that's not something that's faulty with this wf, it's simply the best that can be done with the current models, so overall amazing work here.