Yes, the I2V-14B-480P model is about 82.2 GB, but both of my training runs have fit into 48gb of vram without issue. To be honest I am not 100% sure why or how that works. I think it is because the training is done at bfloat16 with some transformer stuff at fp8. My training data set videos also are not full 480p/720p videos. For v1.1's data set there were 13 training videos all 3 seconds long and 240x426 and it took up about 32GB of vram during training.

@NoArtifact When training on a single gpu I believe you can enable block swap, which would put some of the model into system memory and reduce the amount of vram required. You can also do the full training run in 'float8', which also reduces the vram required significantly. Other people have posted they successfully trained on a single 4090, so I believe it is possible, I just don't have any experience with it or know the best settings for it.

@NoArtifact I am using diffusion-pipe and have been doing i2v only. I am not sure if there is a difference or not. I wouldn't think there would be more vram needed to train one vs the other since they are both 480p 14B models, but again not sure.

All the videos in the v1.1 training data were vertical. I appreciate the feedback on it not working well on wide aspect ratios, I didn't realize that would make a difference. I will add some wide videos to my data set for the next training run.

@NeuroFunkeR  I haven't tried Topaz - but I saw it was only for Windows/Mac, I am on a Linux machine so it won't work for me. I'm not sure if I feel comfortable sharing the .mp4's I used for training - I don't know if there could be any legal/copyright issues using them for training, so I will just keep the sources unknown. I did see DaVinci Resolve Studio 19 is available for linux and has super scale which sounds like my best upscaling option, but it's $300. Seems kinda dumb to buy it just for better AI ass jiggles lol. We'll see.

@Kiefstorm Awesome, that is good to know!

@NeuroFunkeR Haha yes it should have been. And to answer the question for anyone else stumbling upon this, I looked into upscaling since I had already spent several hours on putting the data set together and didn't want to start over if there was an easy solution to improve the video's compression artifacts. Either way I think it turned out fine as is and I now know the impact lower quality videos will have on training output.

Glad it's working well and good to know that 720p works with it, thanks for the feedback!

Well that is certainly interesting, I haven't encountered that before. What is the cfg level you are generating at?

In a future training run it would be possible to change the trigger word, but I think the next version I will be doing a fine-tune on top of the existing LoRA, so it is kinda baked into the model for at least a little while. If it is a common issue I will make it a priority to change the trigger word.

Not sure if it matters but the trigger word is "POVdog" not POVdoggy. After the "POVdog" you don't really need to reference anything with dog or doggy again. You can use phrases like "The woman is on all fours and the man is thrusting into her from behind" or "woman is lying face down on a couch, her ass up in the air, the man's penis is penetrating the woman's pussy repeatedly.".

Thanks! Yeah sometimes the dick gets all melty and stretchy, its very unsettling haha. I found that using a penis LoRA at a low setting (0.5 works well) along with the sex position LoRA at full strength will really help prevent the dicks from getting all weird. I don't think the base Wan model knows what a penis is or what it's physics should be so it needs some help.

For the training of v1.0 I had 2 images in the data set, but due to other training mistakes I made it didn't turn out great. In v1.1 I didn't include any images at all and that turned out better, but I think that was mainly due to data set improvements on the videos. So at the end of the day I still don't know enough to know if images have a significant impact when training a video LoRA or not. I would be curious if anyone else has experience with this?

Good to know that others are using images successfully. I am putting together a data set for doggy style with the woman facing the camera, I'll probably throw some images in there as well to see how it turns out. I need to do some research on it - I am curious if there is a good ratio for the number of videos vs images. I still have quite a bit to learn and don't really understand what makes an optimal data set yet.

It was trained on just one girl in a video, so I am guessing anything that looks like an woman's ass is going to get the movement. You can try prompting different movements for the other women, or try lowering the strength perhaps.

Just now seeing this, it is updated now.