I believe I can fly

I only post the LoRA when I get good results in at least 10 different generations, with completely random seeds. When it's not working properly, I put a warning in the post. I don't know if it was a coincidence that mine worked or yours didn't, but it would be really more helpful if you told me where it's generating and what settings you're using.

@BRai4now And what did you expect? A close-up in 4K? If you thought that was bad, you can try to do something better by creating your own LoRA, spending your own money and resources 🤠

Or you can give me your money and I'll try to improve the quality 😇

@TUOYIMLU88 If you want to maintain overall consistency, simply write down the features. And if you want to keep the same face, try assigning an expression. Wan 2.2 uses an LLM as a text encoder, so it actually interprets text and follows commands. So saying something like "same woman" will usually keep exactly the same woman, but when you get a video that didn't have a reference cut of the same person before and after, it thinks it should cut to another scene with another person, usually more similar to the one in the training video.

You can even test variables, like calling it woman_1 and citing it later. I made a LoRA with 2 women and 1 man, and describing the features of woman_1 and woman_2 actually made it possible to put both in the same scene after a cut.

So if you want to maintain body or facial features, describe them.

Start with 0.8 on High Noise and rise if you need

@TieFighterPilot what the fuck you guys are talking about?! I just redo this LoRA, see my last post on LoRA's gallery to see the consistency of face in the prompt

@TieFighterPilot I train with HLR (High Learning Rate), which causes the LoRA to learn movement patterns very quickly, sometimes even identically. This is called overfitting, and most of the time it's not a good thing. This means that when using 1.0, the LoRA will force a replication of the training, as if you were taking a LoRA with a normal learning rate and setting the scale to 1.5 or 2.0. That's why even with 0.2/0.5/0.8 it continues to work.

Remember that I'm only training the High Noise; it's only responsible for the initial "blur" of body movement and position. The fine details are handled by the Low Noise. I'm creating an overfit in the High Noise, while the Low Noise is normal, and that's why the result ends up being satisfactory, allowing the result to be similar but not identical to the training video. The Low Noise LoRA never got to see my training video, and that's why it prevents it from being identical to the video.

I agree, and the training video was really intense, but LoRA couldn't learn the intensity. But that's also relative, because for example, if you increase the FPS, it will seem faster. I try to work with the WAN 2.2 standard, which is 16 fps with a total of 81 frames. Videos with a lot of movement end up looking very bad with such a low framerate (like 16). If you keep the frames high to have motion quality, it will look slow, and if you keep the framerate according to the duration (keeping the movement abrupt), LoRA ends up interpreting it as "teleporting" frames, and WAN 2.2 tries to fix this to maintain the smoothness of the movements. If you increase the LoRA scale, you will get more abrupt results, but you will also get some pretty crazy bugs.

With her legs closed, you wouldn't be able to see anything 🤪

@Justing1234 If you have a video, it's easier to understand

Just send me the video, and I'll do it as soon as I can

There's no need to reinvent the wheel; if a low-noise already exists that can create the textures, all we need is a high-noise to define position and movement

@zakakki I know a few tricks, but I don't intend to reveal them, at least not here. I spend more time searching for videos than actually doing the training. I usually only use 1 or 2 videos, 200 to 400 steps, and the LoRA is ready. Many people complain, but they're just using the Workflow incorrectly, which usually doesn't bother me since I create these LoRAs for Aisha-AI, and if it works there, I don't care if some random person isn't getting good results.