This workflow for Wan 2.2 generates video from images in only 1 minutes with the use of the Light2X Lora !
What's in this workflow :
- 4 steps generation : with the use of the Light2X lora for I2V Wan 2.2, you can now generate videos in only 4 steps !
- Multi dimensions : The Wan 2.2 models works for pretty much most dimensions, whether you want to generate a 480p or a 720p video or even customs sizes !
- 32FPS videos : With the interpolations node, you can directly get 32fps videos instead of 16fps !
- FP8 Models : to avoid having to load each models after each generations the FP8 models will allow the generation to go directly to the Sampling.
- Multi Loras loader : the RGTHREE Power Lora Loader allows you to load multiple loras without having the mess of multiple loras nodes !
Generations Times : Each videos presented here were generated with the UItra-Pro option on Comfy, I very much encourage you to use it to ensure the models don't get loaded everytime and to ensure a fast generation, you will save time. Anyway Here's the Gen time I've had using Ultra Pro :
- 5 sec 480x832 : First Gen in 3 min 30 (220 sec), second gen in 1 min 10 (70 sec).
- 5 sec 720x1280 : Second gens in 3 min 30 (220 sec).
So as you can see you can theoretically generate nearly 50 5 seconds videos in 480p in an hour.
Workflow guide :
Loading models : Wan 2.2 uses two main models, High noise and Low noise. High noise is always loaded and used first and Low noise always second, I'd advise to keep the FP8 models.
You can also load your loras : Keep the high_noise_model in the first loader at strength 2 and the low_noise_model on the second loader, at strength 1, to uses other loras you usually load the same lora in the two loader. The first loader usually need twice the strength as the second. For more about Loras I explain a bit more down below.
You can change the shift but it shouldn't really be necessary.
Vae : Always use the wan 2.1 vae
Clip : FP8 should be enough, you can try with FP16 if you want but it'll be longer and the quality gain might not be worth it.
Load image : Just put your image and select the dimensions, 480x832 and 720x1280 are the basic dimensions, but you can choose any really.
Prompt : Write your prompt and select your video length : 81 = 5 sec, 160 = 10 sec. The workflow can go pretty high when it comes to frames, I've tested with up to 250 frames, but it will mainly repeat the actions or starts to hallucinate.
The models use each a KSampler (left = High noise model, right = Low noise model) Light2x allows you to use only 4 steps which are divided between each Sampler with start at step and end at step. Euler / Simple seems to be good, Euler / beta is supposed to be good too. CFG is good at 1, higher CFG will give you some unwanted results, like high contrast or "burning pixels".
Once your sampling is done, the videos is assembled here : The first video combine will output the video at 16fps, it will then be interpolated through FILM VFI and be output at 32fps.
Experimental uses of the workflow :
Given that wan absolutely love its slow motion, you can try to get normal motion by deactivating the high_noise_model of the first lora loader, putting the cfg of the first sampler at 2 and generate, it will generate a bit slower (total time was 100 seconds for 480p when I tried), but it seems to be kind of working, buuuuut it will probably give you a much lower quality. Feel free to try it out tho :)
About which Loras to uses :
Loras will be found on Civitai, you can directly download them to your comfy by uploading the lora link to the Loras folder on mimic. Wan 2.2 Loras will obviously be much better, you'll need both High and low models that are offered on their civitai pages.
Wan 2.1 Lora do work, although some of them seems to have quite a hard time, but some of them works perfectly ! All you need is to put the same Lora file on both Lora loader, with the High noise Lora loader (the first) at double strength.
And you should be all set to generate your videos :D