Join the NSFW AI community on Discord! With over 6,000 members, it's the place for developers, creators, and enthusiasts to share tips, discuss the latest models, and collaborate on the future of NSFW AI.
Major Update: The NSFW Wan 1.3B T2V model has been improved!
Based on community feedback, we've released a new experimental set of checkpoints (exp_e1 to exp_e14) to fix the quality degradation and "body horror" issues present in the original model.
This new version was trained with a revised, more stable process on a dataset of 30,000 explicit videos and 20,000 high-quality images simultaneously, after an initial "priming" epoch of 400k images. This has resulted in dramatically better anatomy and motion.
We strongly recommend all users switch to wan_1.3B_exp_e8.safetensors (v8 experimental) for the best results.
This new model can represent concepts across the full NSFW spectrum, generates quality video on its own, and is now an even more powerful and reliable base for training your own specialized LoRAs.
Find the new models and full details on the Hugging Face page: https://huggingface.co/NSFW-API/NSFW_Wan_1.3b
No complex prompting is needed—just describe what you want. We're excited to see what you create with the new and improved model
Description
Retrained from an early checkpoint, adjusting dataset and training configs based on observed results and user feedback from previous checkpoints (e1-e20).
FAQ
Comments (16)
The exp_e14 is much better in every way.
However, it has the problem of being fixed to one type of face. Especially on Asian faces. For a wide variety of random faces, you need to train on faces of different races and ages.
I would like to thank everyone involved in this project.
Thanks, glad you're enjoying it.
Good to know about the faces, would love to see any example outputs you're willing to share
I have only tried exp_e9 so far, but I think so far the previous e10 (non experimental) has insane quality, like i cant believe how good it can do as a 1.3b model, combined with self forcing and causvid lora
@ginx That's really interesting to hear! Did you use it with the motion helper, or by itself? That one was trained only on images
@BetterPorn I think the model linked below is the most beautiful representation of an Asian face, please check out the example outputs.
https://civitai.com/models/25494/beautiful-realistic-asians
@BetterPorn Here's another model that works well for Asian faces. You can use the models I've linked below with the prompts asian, korean, chinese, and japanese for different ages.
https://civitai.com/models/118913/sdxl-10-arienmixxl-asian-portrait
@BetterPorn yes with the motion helper
Is it possible to make a self-forcing version of this? It's such a game changer.
Really a nice one!
Godspeed with the training!
Can we get a LoRA extraction from this model?
I extracted a LoRA for it here https://huggingface.co/NSFW-API/NSFW_Wan_1.3b/blob/main/nsfw_lora_wan_1.3b.safetensors
Do you plan to create a VACE version of this model?
I need it very much for I2V..!!
Sorry for asking i am new to pinokio/wan, but where to place the promting-guide.json?
Hi @a400military333 . That JSON file doesn't actually get used like that, it's only for reference to see how the model's training dataset was captioned
请问这个大模型是否支持I2V?
A late answer's better than no answer. There's a LoRA included on the huggingface page for this model that works with WAN 2.1 VACE 1.3B which you can use for I2V. Also, the "helper" LoRA (here: https://civitai.com/models/1709745/sex-helper-for-nsfw-wan-13b ) works with VACE too.
Very well done, I’ll likely upload a generation but prompt adherence is great. Once I locked it in, I used a cfg of 3 and shift of 7 and since I had to use wan2gp, great work and really enjoy it, thanks.