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.
We've released a new checkpoint!
This new version was trained on a dataset of 30,000 explicit videos and 20,000 high-quality images simultaneously. This has resulted in dramatically better NSFW anatomy and motion.
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_14b
The dataset was captioned with varied natural language, no complex prompting is needed.
Description
FAQ
Comments (35)
Nice! Was hoping someone would tackle this challenge!
Can you extract a lora from this to apply to vace, etc?
Should be on their huggingface link
@funscripter627 Yep
https://huggingface.co/NSFW-API/NSFW_Wan_14b/blob/main/nsfw_lora_wan_14b_e15.safetensors
TeaCache does not work with this model????
IF the finetune changes enough, the model supporter has to go and update TeaCache. If they're saying they truly tuned it with that much extra content, this model may be significantly different enough to not work with Teacache (however I don't know what exactly you mean by "not work"), like does it throw an exception? or just gibberish?
https://github.com/ali-vilab/TeaCache?tab=readme-ov-file#-instructions-for-supporting-other-models
https://github.com/ali-vilab/TeaCache/issues/18
@makiaevelio543 No error or anything like this, but the time for the steps is always the same. This makes the model significantly slower than others, and creating a video takes significantly longer.
@zyx0815 causevid work?
@zyx0815 I found in my "simple" tests the model could run faster with the fp8_fast setting, but that's only on certain cards. I would say if it does nothing, it's because the author needs to compile their own weights for teacache. Neat. My own tests with the model had slightly less comprehension, like it listened less -- even with WAN loras applied -- but it the results were creative and neat enough on their own. I will still use WAN though, but this is a promising addition to the ecosystem
Can you make the equivalent of wan2.1_t2v_14B_fp8_e4m3fn.safetensors?
@videomaker20211124 Added
no love for img2vid model?
@ifuta Planning it, waiting to see how this one turns out.
Anyone have a good workflow that works with this, tried a few and the results all come out like tv static noise.
I just used it in the default Comfy "Load Diffusion Model" with fp8 selected, and "KSampler" with uni_Pc sampler and simple scheduler at 30 steps. The results were nearly identical to the provided samples. Can get it from the comfyorg tutorials: https://comfyanonymous.github.io/ComfyUI_examples/wan/
Same, I'm getting tv static noise when using the lora on i2v. Excluding the noise, movements itself with a starting image seem much better than any loras I tried, so it's even more of a shame it can't work with i2v.
open up one of the videos, right click and save the video. Drag the video to comfyui and it'll import the workflow for that video.
disable teacache if you have it in your workflow
For those who haven't given up, I recommend for ComfyUI: wan_lcm_r16_fp32_comfy.safetensors (Wan Self Forcing Rank 16 (Accelerator)) from this site - unfortunately, for some reason, I can't insert the link, res-multistep/sgm_uniform, 6+ steps, cfg 1-3.5
i'm a noob. How do I use the prompt.json file?
@delta45424155 It's not something you use, just a reference for how the dataset was captioned for each type of content
Doesn't seem to follow prompting very well. A simple pov of missionary sex type prompt leads to face down ass up doggystyle or cowgirl.
That's good to know. It had a very broad dataset and has most definitely seen missionary, but probably wasn't captioned with position names in mind. Instead, it would probably have been described as "woman on her back, man standing between her legs, they're looking at each other" or something like that.
It is very easily fixable with a position LoRA, in fact I just created one for a spooning position with the leg lifted, you can see those results. https://civitai.com/models/1748186/spooning-leg-lifted-sex-position-for-wan-14b
It used a tiny image-only dataset and trained in a couple hours, turned out great on the first try.
If you're doing any sort of sex content (which is probably going to be most of the usage) I highly suggest using the sex helper LoRA https://civitai.com/models/1748212/sex-helper-for-nsfw-wan-14b
Good job! Waiting for Img2Vid asap :)
What computer configuration is recommended for each version?
Sir, has the LoRA version been tested in I2V yet?"
I think people have reported the LoRA working with the most compatibility, I haven't tried it myself.
this is the best <3
Also get extremely noisy results that otherwise work perfectly well on the standard t2v 480.
@joakimkunzdesign128 Are you using any extra tools, like CausVid/Lightx2v?
Can anyone share a working workflow? Can't get this to run on my own and none of the videos, both from others and OP, have a workflow embedded so downloading them is pointless :/
ComfyUI throws this error: If the clip is from a checkpoint loader node your checkpoint does not contain a valid clip or text encoder model.