Trained for image2video generation, not tested on text2vid
v1.1 UPDATE
Version 1.1 is a significant improvement of v1.0. It produces the doggy style motion consistently. The LoRA was trained on vertical videos - it has been reported that wide aspect videos may have some issues. Please leave some feedback if you have issues with it.
Note: I noticed there can be some image artifacts that appear, usually on the ass. I think this is due to low resolution videos in my data set. I looked into improving the quality with AI upscaling, but ran into a lot off issues that ended up making the videos worse, so this will do for now.
Trigger word is POVdog. "A POV video showing a man having sex doggy style sex with a woman." and "Ass movement and bounce is emphasized" can help too.
I used aipinups69's https://civarchive.com/models/1358184/wan-21-penis-cock-dick-lora-t2v-i2v?modelVersionId=1534254 at strength 0.5 in my testing again and it worked great.
v1 -
This is my first attempt at training anything and is still a work in progress. It seems to work okay - some feedback on how to improve is always appreciated. I trained on 480p 14B, I am not sure if it works for the 720p version or not.
Trigger word is: POVdog. The phrase "A POV video showing man having sex doggy style sex with a woman." Was included in most of the training data as well.
In my testing I found this works better with aipinups69's https://civarchive.com/models/1358184/wan-21-penis-cock-dick-lora-t2v-i2v?modelVersionId=1534254 at low strength (0.4 - 0.5) to keep the man bits not looking too deformed.
I need to take a closer look to my training process and data set and will hopefully improve this in the week. Training took about 10 hours on two 4090s which seems a lot longer than other people have trained their models, so I definitely have some inefficiencies.
Let me know what y'all think!
Description
v1.0 - initial training & data set
FAQ
Comments (18)
Is it possible to have Doggystyle facing cameras in the future?
Possibly. I'd have to see if I can find enough training videos of the facing camera angle. And I wonder since faces are so prominent from that angle if the different faces would impact the training.
There's a keyword that describes that angle besides facing cam - "plow cam". Also, you can blur the faces in the dataset if you like. If you are intereseted in exploring that route and need a tool, here's one I created for bluring faces in img/vid datasets https://github.com/garyfnbusey/face_blur_app. Then when creating captions I just make sure they reflect the faces are censored. The captioning tools usually do a good job describing it.
Once it's trained, when using the lora you'll just need to prompt for what the facial expressions should be, since that part won't be trained in
@Kiefstorm Haha "plow cam" is hilariously accurate - I would never have thought of that! The face blur tool looks perfect and solves the face concern, I appreciate you making that tool and sharing it with us! Looks like I will be attempting a "plow cam" angle LoRA soon.
@lazerblazer Thank you for your hard work
@lazerblazer It works great. I love it.
Can you describe in detail how you installed two 4090 cards? Do they have a common video memory or not? Is this 48G or still 24G?
I have an ASUS W790E-SAGE SE motherboard. Two 4090's fit easily inside, and I went with the water-cooled versions so they take up 2 pci slots each, not 3 like the air cooled. They have 48gb of VRAM combined, which can be utilized during LoRA training with diffusion-pipe, or when generating videos with Wan's gradio UI. When generating videos in Comfy, only 1 of the GPUs is utilized though (24gb). I think there are some nodes that will offload some tasks to a 2nd GPU, but I haven't experimented with that. Lmk if you want any more info, always happy to share.
@lazerblazer thanks for the answer! I will not set the second 3090. The benefit is small, but it will be hot, like in a sauna)) I thought there would be 48 gigabytes for video generation)
@Limonobatono There will be 48GB for generation if you use the GGUF MultiGPU node. But like @lazerblazer said, most of the processing is done on one. It allows you to generate longer videos due to the extra VRAM though. Also you can offload CLIP etc onto the 2nd GPU.
CHAD ALERT
Tested I2V, on first try it was pretty decent but it takes more time than other loras to process the video. Anyway, thanks for your work.
Interesting, I saw that the lora size is larger than most. I will have to look into what is causing that. I appreciate the feedback
I had read some people discussing Wan LoRAs and they seemed to think T2V LoRAs didn't work with I2V and vice versa. They could've been wrong, but in my experience so far trying to use this LoRA for I2V it doesn't feel like it's doing much of anything. Do you know anything about this? Was this trained intended for T2V or I2V? Anyone else have any thoughts on this?
This was trained for i2v, I should make that clear in the title, thanks for pointing it out. I agree this version isn't the best. I am doing a 2nd training run and it is already much much better than v1 at only 20 epochs - much better motion and overall results. Hope to have it up soon. I don't have any insight on the T2V LoRAs working on I2V - I have only played around with I2V.
Thanks for the clarification and looking forward to the new version!
@TheIncredibleOnes Just uploaded v1.1 - if you get the chance to try it out I think it turned out much better.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.