@definitelynotadog Thank you very much for your feedback. I'm very happy with this news!! We will wait anxiously! The 1.3B WAN is great and very fast for generating videos on modest computers, but it needs great quality loras like yours to generate nsfw content. I'm very happy to see producers taking an interest in making loras for version 1.3b!

@definitelynotadog Thanks for your reply.

Thanks for reply, How to read this metadata ? best way ?

Nice. 250ish seems like a sweet spot for efficient movement training. 1280 is high! Maybe makes sense though since the part of the image you care about for this lora is pretty small. I've been doing 720x720 for images, maybe I should try higher.

@definitelynotadog Gotcha. So what if i'm going to train long fights that can be several minutes long, then i guess it's best to split those up to as long as i have video memory to cache? And do you extract frames? Or raw mp4?

@definitelynotadog Gotcha. And if the video clip i want to train is longer than 5 seconds, just repeat the other text file for the split half for it's own text file? Or train as if it's a wan prompt using punctuation to end one sequence before starting the next?

@definitelynotadog Gotcha. So i if i for instance used ffmpeg to export all frames as png's, and named the different videos as "frame000, frames000, framer000" as an example which would allow me to easily cache and process 1 frame at a time, would that work? As i'm 1 epoch away from finishing 25 epocks of 1730 frames from 3 clips, some over a minute each. would that then not work? And wan's requirement is "absolute", whether it's extracted as frames or as a video, each "segment" can't be longer than 5 seconds?
If so, if there's a motion i wish to train being longer than 5 sec, how would i train that? Just repeat the same dataset text file for both of the half split videos? Or are my wan loras going to be limited to only get to use 5 sec max clips regardless if split or not?

Just so i understand the core limitations/restrictions the wan training requires and works with.

@definitelynotadog Ohh, so that's what the frame numbers are for in the config. It's being told how many frames each video is? And how can it tell which is which? Cause if 2 clips are equally long, say 120 seconds, won't that mix up with other clips also being 120 frames long and mess up the sequence of split clips?

And iirc i read that it's faster/easier on the memory to make each frame into png's instead? Or is video files best for consistency and full-er context?

@definitelynotadog Ah ok. Gotcha. I shall experiment with multi second videos by splitting them at half each +1 on each (same with clips longer than 10 sec, split to 81+1 frame clips, and then try first with 1 text file duplicated for each split segment describing the same for all for that split clip, and a second training where i try to describe what happens in clip 1 part 1, then a new with what happens in part 2, and so on.

So now that clip assembly will be tested, for the dataset text files, did you write the text for each clip as if describing everything that happens down to the smallest detail if i'd want it to be as dynamic as possible? Or simply just state

"golden labrador, runs towards a skateboard and jumps on it, street at sunrise"?
Instead of a larger description like this.

"widescreen shot of a golden labrador standing in front of a skateboard. The golden rabrador starts to run towards the skateboard, and jumps gracefully on top. It achievines a perfect landing, and camera follows the dog as the labrador rolls down the street."

thanks for elaborating on your training process @definitelynotadog

definitelynotadog Long random realization from my side i think, so if i want to train something that might be 20 seconds long, all i need to do is either to extract each frame as png for the entire dataset, or split into 1-5 seconds videos if doing video files as far as my vram goes, and let it rip?

@definitelynotadog Ah... you mean this one here?
https://civitai.com/models/1368557/masturbation-cumshot-for-wan-i2v-480p

@definitelynotadog  Thanks. I wil try.

确实下载了，没效果，没有提示词案列

@definitelynotadog The face of a person will change. Even if it is made with the last frame, the face will become more and more different. And if there is a light or an obstruction in front of the face, such as a microphone, the face of the finished product will change during the first production. It will not become ugly, but it will not look like the original character.

@fctq9169 is there any way to keep original face consistent?

@definitelynotadog I2V
Maybe it's because the character has bangs or wears glasses? Or she made a change in head pose (for example, she raised her head) so the face changed

@fctq9169 the original face always change to another specific face

Similarly, I tested 10 images. Only one had legs spread open. But she was wearing underwear.

One image of a person sitting did not successfully spread legs. But she lifted up her pajamas and her vagina was seen.

Then I did the same thing twice. The last one even had a vagina. One of them suddenly turned around with her hips facing me. In the rest of the images, the person did not move. So I also want to know how to trigger the correct opening of legs.

Maybe. The author's description of the posture, "Sitting/Lying with legs spread open", you need to describe it separately.

1. Describing her posture with a subject

2. Legs spread open with a subject

EX: [A nude woman] lying on a bed. [she] legs spread open.

I used to use "A nude woman lying on a bed with her legs spread." didn't work because "legs spread open" has no subject.

You can try my way of saying it.

@definitelynotadog 
Today I tested two images. The same prompt words. The difference is Innie vagina or Outie vagina. As a result, both images have a close-up. I don't know if it's because the proportion of the person in the image is relatively large, that is, the close-up. But it is true that after the other images were changed from Innie vagina to Outie vagina, there was no problem of the camera zooming in.

I absolutely know what you mean.....or a cancerous growth, often in dark reddish or even pink colors. If someone knows what (setting) causes this, let us know

Ive used "LoraLoaderModelonly" for both, and hooked them into the "ModelSamplingSD3"

lol can you share some failed experiments?

No thanks.

hey bro do you mean video memory? RAM doesnt matter much in these, its actually VRAM that comes into play. (the ammount of memory you gpu has 16gb's of vram is more than enough, depending on your workflow and precision 8-10 gb is more than enough already. (Mine has 12 and its reasonably enough for almost everything)

@definitelynotadog needs to upload the same file again as i2v, there is no setting to otherwise add it at the moment.