A simple Workflow for WAN 2.2. Uses 4 Step Lightning Lora and generates 101-Frame 720P Videos in about 6 Minutes.
Current Checkpoint and Loras:
Checkpoint: https://civarchive.com/models/1820829?modelVersionId=2060527
WAN 2.2 experimental general nsfw: https://civarchive.com/models/1307155?modelVersionId=2073605
WAN 2.2 Cumshot Aesthetics: https://civarchive.com/models/1869475/wan-22-anime-cumshot-aesthetics-precision-load-i2v-beta-version
WAN 2.2 lightning Lora: https://civarchive.com/models/1585622?modelVersionId=2090344
Description
FAQ
Comments (7)
Standard MISTAKE. You have will have LoRAs that will need to modify Clip, but you have no Clip input/output.! Only LoRAs that do not impact prompt processing can operate without this. Other LoRAs need to be loaded with a node having dual input and output.
Ok,so I have to use the other Lora loader. Would I just run the load CLIP line trough the Lora loaders before going into the Text Encode or does those combine with a special Node somehow? And how would I know which Lora need that? My Creations so far seemed to come out fairly ok.
@CookiePie the lightning loras are fine as you have them, it's for look/style loras. Honestly just use the Power Lora Loader which is a multi-parm, so one node expands for however many loras you intend to load and you don't have to keep stringing them into longer and longer chains.
The other thing is your 24fps saver is just playing back 16fps faster by about 33%. Wan can't generate 24fps and none of the interpolators can do 24fps in Comfy.
@BRai4now thanks for the Tips. Sometimes the Video seems better in just 24fps, so I have just been saving both Versions to see which one comes out better. It's just a way to slightly Controller the length and feel of the Video for me, if that makes sense.
I haven't really looked into Interpolation Yet, but ChatGPT says, it would be possible with the custom FILM node.
Wenn I have some time, I will try it out. If it works easy and simple enough I will update the Workflow, if not I will at least implement the Powerloader.
@CookiePie I suspect that a lot of training data was improperly done using 24fps material without either a lora or Wan's builders properly converting it to 16fps to train their 16fps model. That could at least partially account for why so many people get slow-motion video with Wan. Use of "lightning" loras can alter the speed of clips as well so it's all bandaids on bandaids and it'll be messy until 14B gets a proper update to be able to do native 24fps like the 5B model.
Until then they'll also be plagued by subtle audio sync problems.
Err I haven't checked the WF but this is WAN workflow. WAN uses UMT5 and only DITs are modified by loras - the UMT5 is not adjusted by the lora hence why you can't set a text encoder learning rate when doing WAN training (this isn't SDXL). That's why KJ's wanvideo wrappers don't expose clip strength as there's nothing for clip strength to modulate in WAN.
@jamesabides599 yeah I have non clue what this guy's problem is, he is all caps spamming a bunch of workflows here