This model serves as an enhancement LoRA to stabilize the smaller WAN model. This should make better videos more achievable for people with lower-end machinery.
This model does not have any trigger words, however "penis" is typically required to see male genitals. It's a quality/stability enhancement for the WAN2.1 1.3B model. The model is trained on SFW and NSFW videos.
This LoRA WILL NOT WORK for the 14B model.
Training data was generated using other AI models and some of my own content found on my other socials. Training data is included as a ZIP if you are curious.
Description
FAQ
Comments (17)
Thank you very much for the initiative! The wan 1.3b model is VERY GOOD AND FAST, you just need to be able to generate PUSSY and DICK/PENIS! I still haven't managed to generate either pussy or dick/penis with this version of your lora. Please, we will be very grateful if you focus on this and manage to generate PUSSY and DICK/PENIS in the next update. If you succeed it will revolutionize wan 1.3b. wan 1.3b really needs a consistent nsfw model. I eagerly await a new version that makes it possible to do this. THANK YOU VERY MUCH FOR YOUR COMMITMENT!
I second this. You would be our hero 🙏
The current version of the LoRA doesn't have many genitals in the training data. All of my training data comes from my previous AI generations and music videos I've made in the past, so I'll take some time to automate that and pick a curated list to train on.
@ComfyTinker OK, I get it. Let's wait patiently. Thank you very much for your time and dedication!
If you can make a version that generates consistent pussy and penis and some basic nsfw poses too, such as: (blowjob, dogstyle and cowgirl) that would be AMAZING. I've looked everywhere and no one has managed to do this yet for the Wan1.3B version. The production of nsfw videos for Wan would go to another level because it will be much more accessible and faster.
@rafaaguiar990170 I updated with a version that should show better pussy/vagina. Penis is much harder as it looks like the WAN model was trained on Ken-like men, so I need to update my training a bit to get the penis to attach appropriately.
I generated all of the training data myself, if you want to review it it's included in this as well.
@ComfyTinker This version is much better! Thank you very much for your commitment and congratulations on your progress, great work! Now I can generate pussy with more fidelity and consistency more than 50% of the time even though I still have some problems/anomalies. I can also generate a penis even with anomalies/deformations, something I was never able to do before. The penis is still the biggest problem but it is already a big step forward. If I can help with any suggestions for the next update, it would be training by adding cleaner and clearer images/videos of the genitals, I found the videos to have dark shadows, perhaps this could be causing harm. I think that if I trained with cleaner images and videos without many shadows in the Realistic AI/3D style I would improve a lot. I also think that adding more images of naked men and women in various positions, not only in close-up but also full body would help a lot. I'm sorry if what I said is unfounded, I've never trained a Lora, just a suggestion. Keep up the great work! ;)
Sir pussy and vegane plz, do not redeeem!
@rafaaguiar990170 I'm working on it! LoRA training differs a LOT between different models, in some cases. I'm glad to hear you can generate consistent vagina in more than 50% of generations. Hopefully I can improve that a little more, but thankfully the generation time for these videos is WAY faster than 14B or other video models
@ComfyTinker Yes! This version is much better, you got a great initial result! I feel like it still has the potential to evolve much further and become the Lora nsfw reference model of 1.3B! This is being very important for most people who have low-cost computers. Again, thank you very much for your effort and dedication, great work! We will wait patiently for new versions!
@ComfyTinker 他可以应用在最新的vace或者fun模型上吗
@pushpibl Yes this will work with the FUN model. It should work with the VACE model as well, but untested.
@rafaaguiar990170 You should now be able to prompt with "penis", "large penis", etc and get a more-realistic penis.
@ComfyTinker You are becoming the hero of wan 1.3B! You have made a great improvement in this version, great work! Now I can generate penises with much more accuracy and consistency, in addition to various sizes. This is already a great advance for nsfw production in wan 1.3b, those with modest computers will be grateful! ;) A question: Do you train lora only with videos? I have seen several big producers saying that in addition to videos it is also essential to add many high-quality images in the training, they use more images than videos. They say that it makes the results much more accurate and of higher quality. I think it would be possible to get great images of naked men and women/pussy and penis in various poses and styles from civitai itself to use in the data set. I think your lora would go to a NEW LEVEL if you did this type of training by adding images for a future version. I know it is a lot of work, but it would be very worthwhile. Once again, congratulations on the great work! We await news! ;)
@rafaaguiar990170 I trained this dataset with only videos. Using too many images does end up muting movement. I included the full dataset with this, so someone else can continue training/retrain further if desired.
重要的是，1.3b的模型非常容易崩溃脸部，在control_fun的作用下，1.3b的动作控制不需要担忧了，只需要解决脸部崩坏就足够了
It need to be more 1.4b lora like this.
This is why I've made this is an enhancement LoRA and not a trigger-based one. This can be stacked with other LoRAs!