I don't think there is currently a video model that is great with anime. Wan 2.2 will turn your 2D image into 2.5D.

does wan do transitions? Is it not that perhaps a merger like dasiwa or remix?,
and this is for I2V using T2V LoRAs which tend to use the training faces as reference.
So, this is intended for consistency it gives more freedom when using base wan instead of relying on third party checkpoints that have LoRAs you dont need and affect the result.

@TieFighterPilot Yes, Wan does do transitions with the prompt "the screen goes completely dark and the camera jumpcuts to the next scene". I am using the base model. But after playing with this lora, I understand the use case.

@fenasikerim Oh you are right, I tested it and yes base wan does indeed scene transitions

The real question is why anyone would bother with this at all if they can just an image model to create whatever scene they want instead of playing slots with waiting for wan generations hoping one of them gets it right and even then they have like 2.5 seconds of usable footage. I'm wondering what the use case is with this one. Is it doing some specific NSFW thing that is being hidden from the description?

@RenegadeZebra Oh, I see... Well, in my case I do always 33 frames + prompt scheduling to get faster result before to push it to 81 or 121 frames, the scheduling avoids the long transition phase when increasing frames.
This is for I2V using T2V LoRAs, there are non-compatible T2V for I2V where the face consistency gets totally overrided because of course the training is meant for T2V not for I2V, so this is a helper on that, and it does really a very well job!

@RenegadeZebra After playing around with it, I find it pretty useful actually. Think of this lora as a scissor. You provide an input image and generate the video. The lora basically takes your I2V generation, "cuts" the video where the screen goes black and then continues the video "from scratch" like a T2V but it drags the input image along with it. This doesn't work good with vanilla WAN like it does with this lora. The lora basically says to the model "here is the reference image, use these T2V loras, but use the image to do your stuff instead of generating random T2V visuals". If you for example use a T2V blowjob lora with this, the scene after the dark screen will show your input character doing the action directly, instead of waiting for the male to appear from the side.... etc etc

@fenasikerim Yes, and after playing around with the FFGO and this LoRA I found a trick but i am still working on it to make it perfect but dayum this is impressive, i have tried before without success some banned t2v with couples on i2v until now!

Yeah, just use Cinematic Hard Cut. It's already so good.

@boobkake22 Cinematic hard cut does not do face consistency from i2v using t2v loras, it literally creates another character

Use time scheduling, it work pretty well with me ...
[0.0s → 0.1s]: The screen goes dark and then lights up again to show the new scene.
[0.2s → end]: At this scene {prompt}

@TieFighterPilot What Wan model are you using that actually adheres to milliseconds?

@Jellai Base wan or nsfw_remix, the miliseconds is an example but the scheduling it does really work, in one of my example videos I made a lady to jump, I did as example with just 33 frames I forced the result with the scheduing, try it out you will figure out later because it does work.

@TieFighterPilot i use something similar

beat 1 (0-1s): ...

beat 2 (2-5s): ...

it also works well

Would you be okay sharing your FFGO workflow? I’d love to try it at least once — feels like I might be missing something on my end

@Saxlive hey, please try the workflow here and let me know how it goes, If you want to extend video more, copy and paste the SVI Extension and connect the correct nodes, also you can tweak this however you want, this is still a work in progress and just me testing things
https://civitai.red/articles/30744

@MQ_Lab i hope you don't mind me posting this workflow here, i have gave you credit in the post for the workflow but if you aren't happy i will gladly take down the post/this comment.

What kind of lora is that you are mentioning? Do you have any link for it?

In Citiai (red), click models, and use the research: it will activate filters. Now, in the filters, click lora and then click T2V A14B. You can use any of those, which is really powerfull, since you can with the lora 's scene change:

- teleport the model everywhere

- change it's pose

- fusion of models in the video: You can for example open 2 pictures, side to side. Print screen. With this ugly picture, you can make a video with scene change, both of them will be extracted and teleported in a new environnement. Amazing !

@igotothebeach8973 But how do you prompt it? I tried it without an lora (just the scene change lora at LN1 and HN1. And it didnt work I get like 2 videos side by side lol.. My prompt was "The screen goes dark and then lights up again to show the new scene. both people are now in a hotelroom on a bed, kissing each other." I mean it merged them together but 2 videos side by side and loss of identity.

@teeay25225 I use this workflow, you'll have only one video, not 2: https://civitai.com/models/1474890/wan-i2v-with-upscaling-interpolating-and-smoothing

you prompt could be:

The screen goes dark and then lights up again to show the new scene. The man and the girl are now in a hotelroom on a bed, kissing each other.

Be careful about your source image: it must contain the both caracters, and if possible, try first with a one who contains only the face. The AI focus on the number of pixels, if the face represent only 5% of the pixel, your result will be not good. Cheers