If you want to give runpod credits to help with training, feel free to send a code in dm.
A version of the furry nsfw lora for the 14b wan model. Also works with humans.
This is an img2vid lora, it's meant for the wan 14b img2vid models (both 480p and 720p should work). Attempting to generate txt2vid can yield unexpected results.
Avoid using teacache, if you do use it, keep the threshold low. Teacache causes more artifacts with this lora, sometimes making it do strange things.
(Up to) 5x speedup with causvid
I recommend using lightx2v causvid, combined with the mps reward lora for more movement. When combined, complicated actions continue to have lots of movement. Even on 6 steps with euler/euler a + beta it yields great results with lots of motion and great physics.
Previous causvid lora info below (Personally I prefer v1 at 50% over v2)
Using the causvid lora you can get a speedup of around 5 times (assuming you used 20 steps beforehand) It even works with the wan gguf models. Just make sure to follow these steps:
Load the causvid lora at 0.5 strength (as well as this lora at 1 strength)
Sample at 4 (minimum) to 8 (6+ seems to be the sweet spot) steps with a low cfg (<=3, I usually use 2), use beta scheduler for best results. If you see ghosting, ensure your steps and sampler are set correctly.
On my 3060, this lets me generate a 60 frame video in under 4 minutes. The quality is usually higher than teacache, and it's much faster. Do not combine with teacache.
Purpose
This lora can keep characters consistent, and can handle many positions from pov or similar perspectives. It was trained on many positions, follow the prompting guide. Avoid using with t2v loras, your characters might warp and transform.
This lora is capable of generating (without the need for other loras): cowgirl(+reverse), missionary, doggystyle, blowjob(+deepthroat)
It is also capable of handjobs, titfuck. v1 might need assistance, v1.1 seems to be good on that.
It is effectively a lora for NSFW motions in i2v without changing character consistency. And with a better understanding of furry characters.
V1.1:
Continued training with a new dataset, entirely new captions, more perspectives. It usually yields more motion, and it's easier to tag. It can do many positions, perspectives and motions without the need for a second lora.
Prompting v1.1 is like prompting a t2i model, aside from motions, you can prompt for "moving up and down" or similar. Although v1.1 will usually have motion anyways.
V1:
Note: if you're not getting enough motion/not the right speed
If you don't prompt for motion, you won't get any.
"deep thrusts, fast thrusts", "medium sucking, slow sucking", etc. will adjust the depth and pace.
"The woman moves up and down as she rides the man", "the woman uses her breasts to stroke the man's cock." Should be self-explanatory. You can even prompt for pulling out, varying degrees of success.
Prompting should be similar to prompting the 1.3b version.
Trained on 400 res, frame buckets of [1, 8, 16, 24, 32, 40, 60, 80], context as "multiple_overlapping".
Compared to 1.3b version
This model performs much better at oral, has less stretching artifacts. But might be a little harder to prompt right for the motion, at least on really short videos. Make sure you include a speed and depth in your prompts in the case it doesn't animate enough, this could help.
Trained on img2vid, not recommended for txt2vid
I cannot give any promises about quality when used on txt2vid. Especially furry content, I have not tested it and cannot guarantee quality. It might be able to generate some human content, maybe a little furry content. But I would recommend using a different lora instead for those situations.
Description
Resumed from epoch 17 lora, with a new dataset, 22 new epochs, so technically epoch 39, although the epochs with the new dataset were slightly shorter, it has more variety.
FAQ
Comments (89)
It's crazy, it kind of almost feels like this being trained on anthro/furry somehow makes human characters change a LOT less in I2V; Magic Lora!
please t2v
I have considered this, I'm still working on a couple of things. I will probably train one for the 1.3b first, if it's a success I'll consider training a 14b t2v lora as well.
I've trained 2 text-to-video loras for now for furry+human nsfw. They are both intended to be all-in-one loras
For wan 14b: https://civitai.com/models/1729672
For hunyuan: https://civitai.com/models/1713241
The wan version has the best results by a long shot, so I'd recommend using that one
I'm just starting with WAN so thank you!! 💖
Just a PSA in general Tea cache and wan enhance a video cause weird artificing and produce bad results its really not worth the marginal speed up when the end result is bad or straight up incoherent because the character just no clipped
"Avoid using teacache, if you do use it, keep the threshold low."
@AbsoluteBussin this is a general message for all wan lora's not just his it makes all my lora's perform worse and all others I've tested as well
This lora is probably the worst offender with teacache. I believe it's just been trained so much that the teacache config isn't accurate enough anymore, causing weird artifacts. I got some funny results with teacache.
Teacache has a new mode or something called retention and I dont think the most popular workflow here with the chart res selection is using it, another workflow I have uses it and the results are not bad.
@pocketpie would you mind sharing that workflow or mode for the teacache?
Yeah, also this lora is compatible with causvid, allowing generation in just 4 steps. Less works as well but 4 seems to be the sweet spot. I use beta scheduler, 3 cfg, 0.5 lora weight on causvid and it works perfectly with my lora. In fact, I just realized the nsfw examples on the causvid civitai page use this lora.
@mylo1337 4 steps?? I've been trying different combinations of workflows and cannot get anything to work close to 4 steps, can you share a workflow?
This looks amazing but how were you able to use the kijai wan wrapper and causvid on a 3060? It doesn't support gguf so I'm completely lost here. Could you share a workflow? I'm also on a 3060
I use native comfy nodes. The kijai wrapper isn't required for causvid. Using euler with beta scheduler works at 4 steps (with causvid at 0.5 strength).
Actually I use swarmui, but swarm uses native comfy nodes, so it ends up working the same.
@mylo1337
Thank you! Finally got it to run! I assumed you needed the kijai wrapper to make that speedup work but with your help it really works now fine native in Comfy. I also did a test with causevid deactivated and oh boy, apparently causevid makes 4steps possible since without it 4 really just isn't enough it seems. Again, thanks for the help and the furry lora that actually makes this really fun and interesting!
After using this Lora more and more I finally can give a proper review.
How does this Lora make a difference with anthros?
While Wan 2.1 can do anthro stuff good, this Lora really brings it life. You will never want to gen furry content without after you try it.
It animates the ears to flop and sway during motion.
Tails sway and move realistically.
This Lora makes WAN get creative with how tails, ears and fur are interacted with.
(Tail grabbing, petting, running fingers through fur. Even when not prompted)
No more bad gens were the anthros get feral features.
Retains canine genitalia and it's color (red, pink, black, blue, multi color.)
Sheaths work better. No more making the penis look much thinner than the sheath.
Motion related:
Adds better control for motion speeds and re-enforces sex acts when using with other motion Loras.
You can use this with humans, monsters, anthro and more to enhance explicit motions.
Works phenomenal with 2D and 3D illustrations.
Honestly this can work as a general use, all around NSFW styled motion Loras.
My recommendation for settings.
Use this Lora at 0.7 ~ 0.9 when with other motion Loras.
1-2 motion Loras(0.9 strength)
3-4 motion Loras(0.8 strength)
4+ motion Loras(0.7 strength)
Videos look great for 5s but when I try I get not so great results and when I press the paint brush button I get a "something went wrong" message
Might be something with the civitai generator. I don't know what they do under the hood. Also make sure you're giving it a starting image, as this is for img2vid, not for txt2vid.
@mylo1337 ah I see it was trying text to video, am on mobile and for whatever reason I can't seem to put an image from my gallery
Sadly,it said "Blocked video Mature content restriction" , can't even see the results :') is there any way to counter it? I spent 425 buzz for it
Civitai has some restrictions on generating nsfw content with user-provided images in the generator to prevent nsfw gens of real people. I don't think you can get that buzz back, the warning message says you won't. It seems to be about metadata in the uploaded images.
Damn it,rip the 425 buzz but how do people generate it then ? is it not possible at all with the generator?
Thank you for your reply anyway :3
@StevenRenn I think it's possible if the image provided has metadata, like an image generated with a1111 or comfyui should work. If there's no warning it probably works
@StevenRenn I had the same situation, I spent only 1k on two videos, and the funny thing is that before that everything was fine, I managed to create 1 video. And now even using the same steps to create an image on civitai, it looks much worse than what I created on another site, and about creating a video, this is just crazy
With framepack, I found that leaving the steps at 25 and the CFG at 10, in my test video setup, works better than the suggested settings for whatever video gen software was used by mylo for his videos. In my test, the load order of THIS lora in relation to an action lora (action first, then this or this firs, then action) didn't make much of a difference in the finished video beyond normal run to run variance with the same seed and prompt. Just fyi for those trying it.
Framepack for wan isn't out, since framepack is based on hunyuan video and this lora is for wan video. Using this lora with framepack has zero effect.
The causvid lora (which allows for lower steps with low cfg) is also only out for wan, not hunyuan. If you're attempting to use them with framepack or hunyuan it's effectively the same as not using them at all.
Wan 14b (this lora is meant for the 14b i2v 480p) can be downloaded here:
Regular models: https://huggingface.co/Kijai/WanVideo_comfy/tree/main
GGUF (for running on less vram): https://huggingface.co/calcuis/wan-gguf/tree/main
I see you've updated your comment. Again, this is not a hunyuan or framepack lora. This is a wan lora. It has zero effect when used with hunyuan or framepack.
Videos I've uploaded so not use any additional loras, just this lora, and in some cases causvid for faster generations.
If you want to use the lora, use it with wan, not framepack. Wan can be used in comfyui or some other tools that support it. Wan and hunyuan are entirely different models and can't simply be swapped out.
I sent links to the wan models in my previous reply.
I'm surprised the tool you were using didn't crash or give an error, it should give an error. Wan's blocks are simply called "blocks.{number}" while hunyuan has "single_blocks.{number}" and "double_blocks.{number}". On top of that, wan's blocks are 5120 size while hunyuan's blocks are 3072. Anyone claiming wan loras work on hunyuan, or vice versa is just experiencing the placebo effect. There is no logical way for the weights to be applied.
I honestly don't know what's going on with Civitai, but now when you try to create a video, there is no WAN option in the tools section. Yesterday, this WAN option was in the tools section, and today it's just not there. Has anyone encountered this problem? And does anyone know what the problem is?
Yeah that's interesting, it might be because wan is the only model that was in the generator that was both able to do img2vid and have nsfw outputs. Maybe they'll post an update on it or they could be reworking something.
@mylo1337 God forbid that this was just an update or some rework. And if it turns out that it was a deliberate removal of WAN from the tools, then the authors of Civitai are real idiots
It's back now I think
Question. Does this turn an image of a regular human into a furry? Or do you have to start with a furry picture?
You have to start with a furry picture. Human starting images stay human and furry starting images stay furry.
Any chance that you might try making an LTX version? I love LTX but it sucks in NSFW. I feel like a lora might help that, but I don't know - no one has made a NSFW lora for LTX yet, so you would be the first! Talking specifically about the new 13B 0.9.7 models. Thanks!
Maybe at some point, probably not soon though, I don't have a lot of runpod credit at the moment
No ltx version yet but I've just released a wan t2v and a hunyuan t2v version.
For wan 14b: https://civitai.com/models/1729672
For hunyuan: https://civitai.com/models/1713241
The wan version has the best results by a long shot, so I'd recommend using that one
Is it possible to see the trainings data? If not, could you give an example prompt?
The preview videos have example prompts attached, I used the prompt for both txt2img and img2vid. The realistic examples use willy's noob realism hd v2 (there's also hd v3 now which is even better imo). The 2d examples were made with noobai-xl vpred v1.0.
The prompts for v1.1 of this lora were created by an LLM (gemma glitter 12b) with a vision mmproj (for base gemma as there isn't one specific to the finetune). Not extremely detailed captions, but it's img2vid, so most of the information should be in the first frame, not the prompt.
@mylo1337 I am curious about I2V. From what I understand you used tags, means that you used pictures instead of videos for the training data?
@AbsoluteBussin only videos were used, but the prompts were comma separated tags.
@mylo1337 I see, thank you!
How do you know what prompts to use to "trigger" your concept and when do you know if your concept wasn't trained properly? I made one that reached epoch 30, but the results are rather meh, but rarely one good video comes out.
@AbsoluteBussin you can try prompts from your training data, or just prompt for the motion you want specifically. Make sure you have a decently sized dataset though. And use "multiple_overlapping" in the model config instead of single_beginning if you're using diffusion pipe.
@mylo1337 I see! it was set to single_middle. Doesnt make much sense to pick one screenshot out of a video... any idea how many videos i should use for a concept?
@AbsoluteBussin single_beginning and single_middle take a video of the longest frame bucket length. multiple_overlapping takes multiple videos so every frame in the video has been included in a clip at least once.
The amount of videos needed depends on the concept.
Lots of love for this lora. will we also get a version for the WAN-VACE 14B model? Right now when I try to load it together with VACE, error saying "lora key not loaded" pops up, and the result seems less satisfying despite it can still finish the job.
Much appreciated!
Really nice,
The motion has become much more natural. one personal issue I've noticed is that when the man is on top of the woman (ex.mating press), an penis often suddenly appears from hers below. It might be caused by a built-in motion like cowgirl or missionary, but I'm not exactly sure of the cause. other than that, everything is good.
Have you found any solutions for that? I'm getting this "penis popping out" pretty often and some negatives I tried doesn't seem to be working
sharpy42 I still haven’t found a solution. I don’t use it anymore unless it’s for missionary or cowgirl positions.
swan haha, that's exactly positions I use often and this thing happens again and again. But if I add some specific position lora at 0.8 strength it reduces quantity of those 'surprise dick'. And maybe negative for 'futa, futanary' works too. But it can affect motion I think. Still, this lora is the best. Default NSFW lora doesn't have this much motion quality as I've tried it
sharpy42 Well, NSFW-general also has a tendency to generate extra object on its own, so it can be tricky to use. But in some poses, it actually produces better motion. So for me, since I generate videos for all kinds of poses, what’s best doesn’t really matter that much.But I don’t mean this LoRA is bad.
Would love a text to vid version since it works better on my system. I get the same "lora key not loaded" error.
Lucky for you I released a text-to-vid version yesterday. Also one for Hunyuan but the wan version is miles better
https://civitai.com/models/1729672/stable-wan-t2v-nsfw-furry
all hell broke loose in that gallery
It seems to not be available when I pick 'Wan Video 14B i2v 720p'. It appears in the search only for 480p
Civitai doesn't consider wan 480p and 720p to be competitive even though they usually are. Same with using t2v lora on i2v and vice versa
How work and where use
i m new to comfy, can anyone share a workflow w me
Any plans to make a version compatible with wan 2.2?
For 5b, I already have: https://civitai.com/models/1827161
Possibly in the future. I'm not sure yet if I want to train the lora based on the high noise model, a merge of high and low (so the lora would be applied to both) or wait for more loras to be made and see what would be optimal. For now you can continue using this lora on 2.2, it should still be a step in the right direction.
Also I usually use runpod community cloud for training. But the servers have been having issues lately so for the 5b I had to use secure cod, which costs twice as much. Training the 14b for 24 hours on secure cloud would cost around 35 dollars, so I'll wait for community cloud to be better now.
@mylo1337 If you plan to make one, feel free to hit me up. I can give you my data for the thigh_sex concept I made and a little bit of funding for the lora training
@AbsoluteBussin honestly I haven't really found a need for it yet. From what I can tell the results with this lora on wan 2.2 are sometimes even better than with it on 2.1.
Additionally, with ltx-2 coming up, I'll have to make a dataset with sound optimally. I'm planning to write a tool similar to my caption tool but instead more directed towards scraping, but captioning isn't out of the picture either since the two often go hand in hand.
I hope ltx-2 won't be a flop at least. Got mixed results over the API but if they release the mid range model it should be capable enough.
Sorry for asking but if i try to make imagetovid it always says it has to be PG13. How can i mane NSFW animations here too. I see a lot of nswf animation posts but no idea how to make
Obviously, all NSFW not made on this resource. These upload here for viewing.
Why does no one share a damn example workflow, god damn I hate this website
Just use a regular workflow or even the template and add a lora load node. I use swarmui but civitai doesn't detect the metadata of those gens.
Real. But that's what I get for being a Drawthings junkie. All those workflows out there for everything else BUT DT.
How much VRAM I need? Will be enough around 12GB
How are people getting such fast motions in some of these examples? I cant get fast strokes on a handjob no matter how many different ways i describe it in prompt. I'm also using WanGP so I dont think I can adjust fps choice. Using this lora at 0.9, the lightx2v lora at 0.5 and the MPS motion lora at 1.0.
'misc.' tab, 2nd option. never tested it myself though
HALLELUJAH
I really liked the loras that you recommend :D
I'm confused about how you are running a 14b model on a 3060. I just got a 5070 ti last month (after running a GTX 1080 since 2017), and I cannot run a 14b in my testing so far.
I've downloaded the new ComfyUI desktop app (wow, it got a lot more user friendly since I last tried ComfyUI), and installed wan2.2 5b IMG2VID workflow.
But... I don't seem to be able to run LORA's for it, and I cannot use the 14b models.
1. Use fp8, gguf (q4_k_m doesn't "feel" worse than fp8) if you can. Gguf uses tricks to reduce knowledge loss while still using less memory than running fp8 for example. The example gens on this model page were generated with a q3_k_m iirc, I don't recommend going lower than that though.
2. I have 48gb ram and a >100gb nvme pagefile so comfy can easily offload the model to real or virtual ram. Pagefile works fine for comfy models, just make sure you use a fast storage drive (nvme m.2 SSD, if you don't have one SATA SSD will work too. But don't put a pagefile on a hard drive, it will often freeze your system on windows.)
I also gave a 5070ti now and it's about 3x as fast as generating videos as my 3060 was.
Also remember that 14b in 8bit uses 14gb of memory for the model. + Memory for the latent, + memory for the text encoder and vae, but those will be moved to ram after use with comfy.
@mylo1337 Thank you for your quick response!
I also have plenty of memory (64GB), and run only Samsung 980/990 Pro SSD's, so fast drives as well. I don't use a pagefile unless I absolutely cannot avoid it :)
I may take advantage of your kindness and also ask you for a ComfyUI workflow if I cannot figure out how to correctly wire the LORA in.
If you can supply a workflow, that'd awesome. ComfyUI will just download what might be missing.. It's the usual with ComfyUI - requires a ton of research just to get it to work.
@Foxdude I'd still recommend using pagefile, I have 48gb actual ram but when running wan2.2 a14B (2 14b models) for example the actual usage including pagefile is over 80gb.
For loading gguf models, you'll need https://github.com/city96/ComfyUI-GGUF, you can use a normal workflow and replace the model loader node with the GGUF node. You'll need a GGUF model though and afaik ComfyUI doesn't have auto-downloading for them. I recommend looking either here on Civitai (in case a model maker already did a conversion) or on huggingface for a compatible GGUF model version. (Most wan merges and finetunes I've seen on here are fp8, some GGUF).
A model I recommend is DaSiWaWan I2V (GGUF linked here, wan 2.2 so 2 models, one for high noise another for low, only needs 4 steps total) https://civitai.com/models/2190659, it works well without any loras but also seems compatible enough with this lora.
Also, I usually use SwarmUI, which uses ComfyUI internally but has a user-friendly interface with a prompt and a lot of settings (which will insert ComfyUI nodes for example).
PS: Wan 2.2 A14B is fairly similar to 2.1 14B, and some loras, including this one will have decent results. I just apply the lora at full strength on both models, haven't seen it break from that.
@mylo1337 "I'd still recommend using pagefile, I have 48gb actual ram but when running wan2.2 a14B (2 14b models) for example the actual usage including pagefile is over 80gb"
- That's crazy. Noted. Haven't used a pagefile since the Windows XP-days.
Downloading models manually is no problem at all, it's all about the workflow, that's where I'm stuck. Thanks anyway, I'll have a look.
@Foxdude With SwarmUI you can just select the models, write your prompt and it makes the workflow for you based on your inputs. I'd recommend it for most people since it still uses comfy so you get all those benefits of good memory management and extensions (Although some will need a swarm extension as well to use extension nodes).
I too was for a long time bound to Wan2.2 5B with my 5060Ti 16GB and every time I used any 14b workflows (even those for 8-12GB cards!), it worked once or twice and then stopped working, no matter how often I restarted ComfyUI, the PC or whatever. Anyway I finally found a workflow that even had auto prompt and it worked a while until it too broke.
But I found that you can add the start parameter "--disable-pinned-memory" to your ComfyUI start batch file, it now always works.
Does this go in the high noise or low noise lora slot?
This was made for wan 2.1 which only had 1 model instead of 2. You can use this lora on both high and low noise though, I haven't had it break.
@mylo1337 Thanks!
@mylo1337 Wait, the lora works on 2.1? I did a bunch of testing and the results turned bad for me. I use the triple ksampler advance setup. How are you doing it?
@AbsoluteBussin If you mean 2.2, yeah kinda. But I would recommend using a capable model for the base, default wan 2.2 probably won't give very good results. But applying it to an nsfw wan model for example can improve movements and looks. I've tried using it on high noise only, low noise only, and both, I prefer high noise only, but on both also works decently.