This LoRA is designed to bring more motion and detail to Hyvid hetero sex scenes. It should be used with other motion LoRAs, using straightforward natural language prompts. Consult the prompting for LoRAs you combine with this one.

This LoRA improves other sex LoRAs in the following ways:

- More motion when combining with character LoRAs.
- Larger range of motions and facial expressions
- Better lighting and environments
- Higher detail of bodies and anatomy
- Crisper outputs at higher resolutions - lower the strength of other LoRAs if you get blurriness

Tips

• If you get jitter or sudden movements, try turning up flow shift. Going from the standard 9 to 17 can improve fluidity, especially when combining character LoRAs.

• Adding the Boreal LoRA at 0.2 - 0.3 strength can improve hands, feet, and other fine edges.

• The simplest way to use this is to start with an existing generation that you want to add more motion or detail to, then add this LoRA at 0.2, 0.3, etc using the same seed until you find the sweet spot. The amount of strength needed is dependent on the other LoRAs, so you will need to experiment a bit.

• When using as a standalone, you'll get very different results in widescreen vs. vertical aspect ratios. The training data is all widescreen and includes things like camera pans and zooms, and also a large diversity of angles. Widescreen seems capable of better results but also takes more work to get the prompts right.

Training Notes

v0.1 and v0.2 were trained on a large diversity of short PiV clips, but the resolution of the originals was not great so there were lots of artifacts.

In v0.3 I started over from scratch with 5 videos, each 4k and about 30 minutes long. I used diffusion-pipe to make frame buckets of 121 frames @ 244px and 33 frames @ 512px, using the overlapping_middle method so that the model saw the entire video twice, but with different samples. The idea was to give it higher res with short term motions and then help it understand long term motions with the 244px samples. I made a mistake having one video feature an actress's face more than the others, and it's overfit to her face a bit. The LR was 5e-5 and took about 7k steps.

v0.4 trained on top of 0.3 using 8 new videos of about 1hr each, also 4k sources, this time focusing on 1st person POV. Since most other sex LoRAs are 1st person I wanted it to improve when combining with them. The sources are all 16:9 aspect ratio so they perform a bit better in widescreen. I reduced the learning rate to 1e-5 to prevent any major forgetting, and the total steps was about 12k. So it has seen about 19k individual samples from 13 videos so far.

An important thing to note if is that there were no captions of any sort, so it does not come with any new text biases. The idea is more to fill Hunyuan's pretraining gap, which seems to have not included any video porn (hence why it can do porn stills but not motion).

I have done one fine-tune on top of this (the anal version) just to see if it can pick up on new concepts quickly. This was also trained without captions using one video, and it definitely picked up on the concept quite well just through the video examples. My hunch is that fine tunes should perform even better with source diversity and captions.

Version Notes

V0.4

This version is trained on 1st person POV sex scenes with a goal of improving the motions and understanding of sex from a male's POV. The model does have some understanding of 3rd person POV as well but it's currently less developed. This version will work best in combination with 1st person POV LoRAs such as Missionary, Cowgirl, Reverse Cowgirl, etc.

Widescreen aspect ratios will yield more fluid and interesting results. The training data is all 16:9 ratio 4k videos, and so using that aspect ratio is going to get you closer to the source material. Using vertical aspect ratio will push you more toward other LoRAs in your stack that use more of those ARs.

This version does still suffer from some sameface from getting overbiased by a long previous video at too high of a learning rate. This version reduces that but it will take more training to completely neutralize that. The next version will focus on face details, facial expressions, and diversity.

There are example workflows in several of the gallery posts. The main strategy is to use an existing sex LoRA at about half its normal strength, then combine with this LoRA at 0.5 to 1.5 strength. You will have to experiment a bit, especially when combining image-trained character LoRAs.

V0.3
This version is trained differently than previous versions, using higher res samples and greater diversity of scenes. It currently has a bias toward skinny blondes but can be prompted away from that.

It works quite well on its own as well, just use descriptive prompts to get the pose right and play with the LoRA strength and guidance levels. I have had success using character LoRAs trained on images only up to 1.3 strength with almost no loss of motion.

It works great with vid2vid, so for those looking for more interesting poses, that is a good option.

V0.3 - Anal Edition
This is an even more experimental version trained on one anal video. This is mostly to serve as a proof of concept that the base model can be guided by specific trainings, which opens the door to more specific variants. There is a workflow embedded in the gallery video to reproduce it in Comfy. Sorry for the lack of prompt guidance, I've been forgetting to set save meta to true when re-using past workflows.