about v2: better face accuraccy, but 2 seconds sacrificed for transition.
transform your image input into deepfake, but the accuracy is kinda low in this version yet.
make your prompts like t2v, but with this first, in example.
DEEPFAKE, scene transition.
the girl of the picture eating burger at restaurant, the background changes to a fancy restaurant.Description
FAQ
Comments (23)
Great work and concept!! One question. Do I use this lora only on 'high noise' or on both 'high and low noise'? And what is the ideal strength? Thanks!
just high noise at 1 strenght
@luisa_pinguin Ok, thanks!
A lora to help is good. Using | in the prompt has a similar effect.
i di'nt know about that, thanks for sharing.
Can you explain? Give an example of using “|”?
@ltcdrbroccoli https://civitai.com/models/2048863/wan-pov-cowgirl-insertion
@ShinyLilly Using | in the beginning of prompt (and without helping loras) the scene change is just too slow. It always uses 30-50% of video length to do a fade-style slow transition. Unless there's a prompting fix to that. Using words like "instant" or "immediately" really don't seem to affect this.
@jevzomqc6b187 It works even without "|", as part of the model capabilities. With proper a prompt it can take 0.5s if it's used during the prompt and a given indication for the model at the start. You can even change camera motion/perspective - close up, medium shot, tracking shot etc.
Example I2V - [Photorealistic, High Quality, Multiscene video] A woman sitting at a street café' , people are walking by, she smiles at the viewer. Jump cut, the scene changes into a tracking shot behind her, she's walking inside her luxurious house, in her living room, modern gray sofa with a dog sitting on it, a modern black wood table with whisky bottle on it, Big TV screen.
My own example (using the word "transition"), changing from tracking to static shot
@billywicked This was really useful info. Thank you! Sometimes I hope someone would build an actual wiki of Wan prompting since even Alibabas own documentation doesn't really cover much of the actual capabilities of the model.
Everything works, not perfect, but it works, thank you. You just need to add it to the high lora and that's it.
I'm not 100% sure what this lora does. Is it supposed to take a character image then immediately transition to a t2v scene, like an instant lora / ipadapter kind of thing?
its a great idea only sharing that wan can do this already with typing the "scene transition to X" and loading a lora to what you want. i get the desire results the same way, so not sure what this lora will improve.
Cool idea and thx but native Wan 2.2 already supports that using the | prompt
Only with kijai's wrapper, not native right?
@allamalla Ah, that's possible sorry. I can't remember using anything else then kijai's wrapper so i already call it native :D
Do you guys have a source explaining the "|" in Kijai's? I would like to read into it in detail
@allamalla Thanks, will check it out!
@allamalla Alright so he is talking about Long I2V Multi/InfiniteTalk -node and wrapper text encoder nodes. Which workflow does include this? I guess the Kijai I2V workflow doesn't work since the "|" had no effect on my results. Can you provide a ready to use workflow which I can fiddle around with if you have one available? Thanks in advance
wan2.2 can do transitions already, but they are kind of random. might take several seconds to transition. this lora insures immediate transition. seems to work fine at 0.1-0.2 strength. too high of a strength affects the scene way too much
I found that I needed to pump it up to 0.3 to get consistent prompt adherence but I really appreciated this advice, which convinced me start from the bottom and work my way up. If I had started at 1 strength like I do with most loras and not read your comment first, I would have written this lora off as useless.
ScaleROPE scale_t 2-3 for a fun time. Sub 1 values also work to correct excessive internal FPS (read: Hunyuan)