Wan 2.2:
example prompt: Very high quality 4k animation, This is a masterpiece of the best quality, the video is showcasing amazing and highly detailed artwork. Man's POV of him laying down with a woman facing away while riding on top and having anal sex with him. She is riding on top quickly bouncing her butt up and down as his penis goes all the way in and out of her asshole.
Wan 2.1:
This model was trained on realistic videos on the Wan T2V 14B model. Version 1.0 was trained on the I2V 720P model, but Version 1.1 should be compatible with both I2v and T2V. I trained locally on an A6000 using 20 2 second videos at 1920x1080 resolution (24 FPS, 48 frames each). Training 2800 steps took around 48 hours.
Training Setup
I trained this model locally using diffusion-pipe: https://github.com/tdrussell/diffusion-pipe.git
Here are the toml files I used to train as well as the command to start training (within train.sh): https://drive.google.com/drive/folders/1Ns6IPQlNp-jYz76LlqpYpOoaQwUx3AN9?usp=sharing
Description
Retrained on the Wan 2.1 T2V model for compatibility with both I2V and T2V
FAQ
Comments (12)
Question about your dataset.config file.
Is there a specific reason that you set frame_buckets = [1, 8, 16, 32, 48]. You mentioned in the model card above that you trained on consistent videos at 48 frames ea. Is this just for your convenience when using this file with other datasets, where clip lengths may not be consistent?
Or am i misunderstanding the purpose of frame buckets?
This was actually the default. From my understanding, you can train on a variety of number of frames within a video, and diffusion pipe will put them in different "buckets". I haven't tried mixing frame counts, but you just have to make sure that each video has one of those amounts of frames (1 being an image of course).
@pikenrover Huh, the defaults i've seen used (from tdrussell's repo) are: frame_buckets = [1, 33, 65, 97].
My understanding of how frame_buckets are used is slightly different: any videos with frames larger than a bucket, get bucketed to the closest smaller bucket. IE: in the case of frame_buckets=[1,33,65], a 48 frame video would be put in bucket 33. I don't know what the implications of bucketing the larger number of frames in a smaller bucket is though. Do they all just get processed as 33 frames even though there are 48 frames present? Need to dig into that more.
@lowcaloriesyrup Oh interesting. Either I changed the buckets and forgot, or the example changed in the repo since I cloned it. If it was the former, I did so because I already had 48 frame videos and wanted them to fit in a bucket. Your understanding of buckets is probably right. I had AI explain what it did.
combine this (0.6) and "SingularUnity - Anal Reaming" (0.9) for good anal doggystyle. works better for me than both on their own
When downloading v1.1 it shows t2v and v1.0 is i2v.
If those are true, perhaps rename the top label according to type 2 vid for less confusion?
Thanks for the share.
I see the confusion. I labeled according to the model it was trained on
@pikenrover If you edit the civitai model page, you can modify the name of the 1.1 and 1.0 to t2v and i2v respectively.
Hello first of all, that good lordas incredible, a question I see that in each of your laces the number of videos you use is very different, you consider that using more videos, a better Lora is created, I am surprised that a training is taking up to 5 days. At the moment I do not have the computational power to try an XDXDXDXD video LORA, great work I admit it. I will make one of your followers, for your visual content XDXDXDXD.
Pretty amazing! Would you consider doing a similar Lora, but vaginal instead of anal?
wan 2.2 version?
Where can we get the models you used.
Details
Files
wan-t2v-anal-reverse-cowgirl-e54.safetensors
Mirrors
wan-t2v-anal-reverse-cowgirl-e54.safetensors
wan-t2v-anal-reverse-cowgirl-e54.safetensors
wan-t2v-anal-reverse-cowgirl-e54.safetensors
wan-t2v-anal-reverse-cowgirl-e54.safetensors
wan-t2v-anal-reverse-cowgirl-e54.safetensors
wan-t2v-anal-reverse-cowgirl-e54.safetensors
wan-anal-reverse.safetensors
wan-t2v-cowgirl.safetensors
wan-t2v-anal-reverse-cowgirl-e54.safetensors
wan-t2v-anal-reverse-cowgirl-e54.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.