🚨🚨🚨MISSION CRITICAL: PLEASE READ🚨🚨🚨

download the lora, find comfyui workflow, download all files requirement then create it in there

This is probably the best tutorial video if you have no idea what you're doing:

https://www.youtube.com/watch?v=gLigp7kimLg

Start with lightx2v like in the video, but you should remove it later and check the results, it tends to degrade 2.2 a bit. It does make the videos generate faster though.

@magicballoon followed allat but my results seem to have a very quick and steep drop off in quality, very quickly turning into a blurry mess, am I missing something? I removed the lightx2v LORA
seems to just kind of vibrate no matter the prompt, with everywhere it goes turning into a blurry mess like fireworks or smoke on youtube where wherever it "touches" becomes blurry and blocky

@ThatRedDot I forgot to mention, one more thing you can try is setting the high noise sampler CFG to 3.5 and the low noise sampler CFG to between 5 and 7. Those are the values I use when I generate. It's hard for me to say what your issues might be without seeing your workflow and hardware though

@magicballoon RTX 4060TI 8gb, better than what the guy in the video had, tried WAN, same prompt and image here in civit and it works fine, workflow is the same as the guy in the video just changed the LORA to the NSFW one, I'll probably just stick with photos for now, it's annoying AI is so powerful but yet nobody decides to take a break to make it more user friendly, I mean stability matrix is good but it took me ages to install sage attention, sucks that these guys spend years making this stuff but not a week wraping it all up into a working docker image or smth for idiots like me lol

@ThatRedDot Yeah, it's because most of the people who generate stuff for it are other software devs and researchers. I would say take a break and come back to it fresh later, check out a completely different tutorial/workflow or search for issues online. Video gen was not easy for me to get started either and I have a 5090 now.

@ThatRedDot if you dont mind spending a few bucks for patreon subs, secourse on yt have very nice 1 click installation comfyui bat, including sage attention and others.

@LewdAi sadly I am not in a financial position to be able to afford that, I would if I could though

@magicballoon with how popular AI is, I feel like there is defiantly a gap in the market for something that makes it a 1 click install, maybe using something like docker in the back end, something like stability matrix but that handles dependencies and stuff for you, then sell it for like 5 bucks and I feel like it would sell really well, maybe not though

update: I got it running with pinokio, thanks for the help everyone!

@ThatRedDot If it worked for you, send me a tutorial on how you did it. I'm new to video generation and I'm interested.

@hildesempai there is an app for ai called pinokio, I just downloaded that and installed the wan 2.1 package and it just worked (called wan 2.1 but it uses 2.2)

comfy ui, models, loras drop it in there

wack them all in Power Lora Loader's change order up and down if not getting desired effects + try diferent weight - until you get beautiful balance - works for me - easy control, no daisy chaining, switch them on/off, and unlimited lora slot! PLL GOAT!

@ronikush I basically have been doing the same thing with the daisy chaining...putting all the LoRA's in different orders and playing with the strengths. With the other NSFW models, I put the character LoRA first, the NSFW model next, and then any accelerator like Lightning LoRA. Just seems like there is no rhyme-or-reason with this particular NSFW library.

It sounds like you're doing T2V, is that the case? You'll make your life a lot easier if you just gen an image with the character you want as a starting point.

@magicballoon I thought of doing that, but wouldn't having the full character LoRA produce a more-accurate rendering at all angles? Like I said, don't have this issue with at all with using these character LoRA's in T2V SFW (and NSFW with other models) videos without this model, so I'm scratching my head at why this particular addition (at least the way I'm trying to use it) just mucks everything up.

@TowerT I couldn't tell you why you wouldn't have issues in other LoRAs in T2V, I just know that this LoRA works fine when provided an image with the character in it from the beginning.

probably you forgot the nsfwsks. I made that miastake too

I'm here to say that 🤣 all characters shake all the time, no way to Suppress it

@emadgh true dude

i'm not sure, but when is set weghts to the high loras 2.0 and low 1.0 - they stoped shaking so hard. But its generating a lot of shit anyway. I have a window on art with blue sky, but AI said there is running TRAIN. chu chuuuuuuuuu

I heard elsewhere that you should use 2.00 for high and then 1.00 for low but I am not sure if this is correct. I have been doing that with this lora and it works fine but I am going to try both 1.00 now.

for me the more weights on both symmetricaly the more intense action/shaking so i just use 1 because lightning lora can't show shaking THAT intense

im using 0.9 and its working perfectly fine

@Linklaine Thanks for the answer. I tried 1 and 0.9 and 1.5, it works great with anime characters, but as soon as I try to generate a video with a real photo of a person, artifacts appear. I thought maybe it would be worth increasing the weight of the low lora, but I still haven't found the ideal values

@st3rb3n Combine it with other loras at lower strength, https://civitai.com/models/1811313/wan-dr34ml4y-all-in-one-nsfw is perfect for that. I use both at 0.6 and it's working out well enough.

I did various tests and found that if you want to add simple movements such as basic sex motions (moving your hips back and forth in a constant motion) like with LORA in wan2.1, using "High LORA only" gives more stable results.

If you use both High LORA and Low LORA at a high value (0.4～1.0) at the same time, you will end up adding unnecessary movements such as small vibrations that make it seem like you're constantly vibrating. (※You'll get a deeper understanding of these things if you generate the animation by adding only "Low LORA".)

So, the steps for using it are:

1. First, test whether the prompt behavior can be reproduced without using LORA.

2. Test using only ”High LORA”.

3. Add ”Low LORA” to increase the behavior variation. (※This is optional use and generally unnecessary.)

@misora Thanks a lot! I experimented with your approach and managed to find the best setup for me — it really helped to get rid of the issues and artifacts I was running into.

I still have one question though, maybe you can give me some advice: I’m using the GUFF 480p model together with a Lightning LoRA at 4 steps to speed up generation. A 6-second video takes about 60–70 seconds to render on an Nvidia H100, but I notice that the facial features — especially the eyes — tend to “melt” or drift quite a lot during generation.

Do you think there’s a good way to fix this? I was even considering training a dedicated Low LoRA focused specifically on faces and eyes to make them more stable and sharper.

So, if I'm understanding you right, the base (supposedly NOT trained on NSFW content) WAN 2.2 is potentially better at NSFW content generation, if it is given the explicit detailed prompting, versus what is covered by trigger words in this LoRA?

Did you forget the LoRA keyword in your prompt or something? Like, I just can't even begin to relate to this post.

@TowerT im certain base wan2.2 has SOME NSFW knowledge because i prompted it for a subject to massage their body then hold their penis and it understood without any LoRAs...plus it animates penises quite well (just don't look too closely) where wan2.1 would want to disappear them

Wait Wait! What? .. Wan 2.2 does NFW without Lora? ... Errr oh shit never actually tried without loras hah .. time to peer down this rabbit hole! ...

@ronikush this is I2V NOT T2V

How did you make it run? I have 12GB of VRAM and it absolutely wont run. The WanVideoSampler runs out of memory.

The tip no one wants to hear: Don't use Kijai's nodes, which are the ones you are referring to that require manual block swapping. Kijai himself has admitted that his nodes are not as efficient at memory management as the native nodes on GitHub issues. The ComfyUI core dev team are among the most competent, if not the most competent devs in the community. Perhaps his nodes can work better if tweaked properly, but a workflow that uses native samplers and loaders will handle the difficult parts for you. My videos use native nodes, there are also others on this site you can find by searching "Wan native workflow" I believe. I have the same GPU. You can still use a custom LoRA loader to make the LoRA loading process less tedious.

@magicballoon Thank you! I don't have the block selection node, but a percentile slider instead so not sure if those nodes are different from what you mentioned/ Any links to a specific workflow though? I usually find Wan 2.1 ones via that search but I don't have any 2.1 diffusers and I'm not sure about this lora's compatibility with those anyway

@MizarAiArt I don't use diffusers. You can drag most of my videos into your workspace to see their workflow. It is a bit cluttered although it is separated into sub graphs so you only need to focus on one part at a time. If you want to actually learn the nodes, however, which I would recommend, you should try this workflow:
https://civitai.com/models/1389968?modelVersionId=2081318

It's basically the same workflow I started with. You can sub the native LoRA loader with a more convenient one, I use the Power Lora Loader personally. I also did not ever use a VRAM Debug node between the high and low samplers. Seeing as you have a 5090, you probably do not need it and it probably will make the generation slightly slower. Feel free to try it though.

I would also try not to use lightx2v personally. It makes generating faster but it also degrades the movement of the video. You can do 12 steps like in that workflow, I've done that with decent success, but recently I've moved to 20 steps. It takes longer but it is usually worth it.

Oh also, one more thing is that I've recently found a custom node that uses an algorithm (based on the Wan 2.2 whitepaper) to switch from high to low noise sampler automatically. You can find that here: https://github.com/stduhpf/ComfyUI-WanMoeKSampler/tree/master

This makes it so that you don't have to guess how many steps should be spent in the high noise sampler and how many should be in the low noise. The ratio should be set to 0.9 for image2video

@magicballoon they are way more efficient. he said they are not as consistent, ie you constantly fuck with the setting manually to use as much vram as possible. that it its frequently updated and stuff breaks. they are way faster (about 50%). i have had way more issues with the default memory management, it really likes filling vram and ignoring the reserve vram setting. maybe if you have less vram the default works better, but im not going to pass on a huge speed up because i have to set a single number if i change resolution. even then, with the new pre fetching block swap has like a 3s difference between 0 and the max. you can just set it for 720p and forget about it.

@filiv34543152 I'd be way more inclined to use them if I could get usable instructions. All I get is basically "fiddle with block swap until it works."

There's a lot to learn. A useful tip is that for any image or even video made in ComfyUI, you should be able to drag it into your ComfyUI workspace and it will show you the workflow -- unless the uploader scrubbed the metadata. You can do this with the videos I've posted. However, the workflow I use is huge because it has multiple sub-graphs.

There are plenty of Wan 2.2 workflows on this site if you search for them that are designed to look pretty or simple.

@magicballoon does not matter how i download the videos of yours, "unable to find a workflow" :(

@HBLearning try some from the gallery, most of them there will usually work

@HBLearning Not sure which ones you're downloading but I've tested my own before and was able to see the workflow. Wouldn't surprise me if some fuckery was going on with this website though.

I'll keep trying. thanks for the replies all. fingers crossed

I'd REALLY love to see the training set and all the associated captions/prompts trained with it.

@TowerT sorry i don't understand. I tried the lora. I did not make the lora, so i don't have the training set/prompts.

@Zachman18 I was saying this in general, not to you specifically.

Also, I am using lcm+sgm uniform pair with 10 steps to produce 17 length in comfyUI, the animation is quite perfect but only the cross section did not animated

Have you tried to copy/paste the noise seed in the sampler? If you want the same way of generating, just paste the seed and generate

我的没这问题诶 我还希望它能多生成个人，不然只有女的图没办法生成另一个男的

@illusionGame 我这里贼抽象 生成不生成 消失不消失 甚至还会屁股长前面 小洞长鸡鸡 总之就是不听话加畸形儿

I have the same issue

ive tried and both work well

Bet ur a fat Indian dude with a giant mustache

@x9741074 aye is he good giving head tho? that's the question

REAL

@x9741074 ok but does it matter thooo i mean head is head

Seems like it basically just makes any penis go in a hole or any boob or ass bounce on a sine wave with the amplitude proportional to the Lora strength.

@CAldonni It seems like nobody can give me a straight answer so I will continue to just use the normal WAN 2.2 model. It animates bodies just fine- in fact I have gotten it to animate penis thrusting and movement, breast movement, and sexual motions without any issues. Again, waiting to hear what this actually does... from what I read it just makes things worse.

It helps understanding sex prompts better. I generate almost daily, and the difference with or without it is noticeable

Where did you download the model

@LewdAi is there some kind of documentation somewhere to better understand sex prompts on CivitAI?

@Julzjuice there is none, you understand it by generating more, for example take prompt from one of the video generated here, use the prompt with and without the lora the difference should be noticeable

@Julzjuice What @Julzjuice said. In general, the more coherent and detailed your prompt is, the better your result will be, whether you're using naughty training data or not. Set up a constant seed and see what changes as you add detail. Don't think in abstract terms, and don't use analogies. Concrete, direct and clear phrases. Set up an empty Note node and save everything you write, or use a spreadsheet. Who, what, where, and how hard. There is no why. Lower your weights if a strong LoRA is ignoring you, and heavily weight your prompts.

Yeah this happen a lot with Wan 2.2 and is like the Flux of video creation having a prudish nun built into the design to make everything lame and stop you from seeing what one can imagine because XXX thoughts are impure. I really hope people can see the implications of what thought police means and finally start to understand how religion is the philosophy of control and happened not by accident as the Greeks had philosophy of knowledge and understanding then maintained order with law. The Romans bastardized everything under their rule all the way into the formation of the church post Roman days taking other peoples ideas and merging them into a way of controlling the masses using the mind instead of the sword and if the sword is needed use the mind of someone else on your behalf. It is the biggest Psyop ever pulled off and we still live in it because of the residual culture, existing like artifacts in the mind of people, passed on from simple minds that have no critical thinking capability naturally. Thank the Loyd for the Muslims as they were the ones that saved the Greek works or we would otherwise still be living like idiots of Europe of the time who had no idea the Greeks even existed but became woke when they discovered Aristotle and the Protestant was born eventually leading to our attempt at freedom that failed when the founders died and Martin Van Buren became the first American born president and the rise of the demagogues happened and is where we find ourselves now.

@FemBro Wilhelm Reich and Michel Foucault. Nuff said.

I freshly learned how to use ComfyUI and Wan 2.2 I2V. DM me and I'll help as much as I can !

@thedrekt00 i also could use some help if you would be able to haha. I really strugle with understanding the workflows and loras, I am getting mediocre results

https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/diffusion_models/wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors

https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/diffusion_models/wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors

https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/diffusion_models

U need use both

@SupremeEvil U r so cute ^.^ I hope you have a wonderful day!

i am a newbie so i don't know much,but i guess high affects the the general overall action,while low affects the details

@SmithRose1996 A model with high noise generates a video image where everything is blurry but the essence is clear.

A model with low noise uses this blurry video to create a normal image.

@SupremeEvil Thank you for the explanation!

.9 and .9

For I2V I use 0.4 and 0.4 with using other LORAs

I noticed this too, this lora is really good with simulating sexual motion for hentai I2V, but for normal undressing and others I guess you could just click on bypass nodes on your workflow to bypass trhe lora loaders with this lora, and then load anothers

I use weights 0.4, this calms down the LORA

@uqw Can the LoRA handle both fast and slow actions?

@TomLucidor yes, but make sure to use high noise model on high node and low noise model on low node

if you use lora + lightx2v , u need to set 4 step if u have start step end step a 0 2 / 2 4 , if u set 6 , u need to adjust 0 3 / 3 6, if 8 : 0 4 / 4 8 , and the artifacts will disapear.

if u set 0 2 / 2 4 with 8 step for exemple, it will be blurry and weird

By the way, these kind of lora insert are train at 16 fps usually, so try 81/160 frames (5 or 10sec) 16 fps max , or it will be too fast above 16 fps.

Took them 5 months to advance a version, so January next year for the next version?

This is late, but in the DeepBeepMeep you put Wan 2.2 Lora's in the "loras" folder, instead of the "Lorasi2v" (where Wan 2.1 Loras go).

Then since they have the high and low weight stages separately, you have to mark their weights differently.. so for Wan 2.1 you'd only have 1 Lora per action/function and you'd mark it's weight, for example just "0.8"... here if you first put the High noise NSFW Lora, then Low noise NSFW Lora, you mark their weights 0.8;0 and 0;0.8

That means that the High Noise Lora will use 0.8 weight in first half of processing and 0 in the second, whereas the Low Noise Lora will use a 0 in the first half of processing and 0.8 in the second half.

Or with the full weight of one, it'd be High noise 1;0 Low Noise 0;1

So for example when I have Loras "lightning-high-noise-model" "lightning-low-noise-model" "NSFW-22-H-e8" "NSFW-22-L-e8"
My "Loras Multipliers" section reads 1.5;0 0;1.2 0.8;0 0;0.8"

Other settings (For Pinokio browser/DeepBeepMeep) are:

Guidance phases: Two phases
Model / Guidance Switch Threshold: 900
Guidance (CFG): 1, Guidance2 (CFG): 1

I use sampler dpm++, Shift Scale 5 ..and 6 inference steps.. 4 can sometimes do ok, 8 can sometimes be better, but 6 mostly.

But that is specifically with the "lightning" loras in use that speed up generation times on lower memory rigs. If no lightning loras, then you need to bump your inference steps much higher aand the guidance phase values are probably wrong...

Anyway, hopefully that helps some DeepBeepMeep users, not many of us I guess :D

It can do quite a few NSFW things on it's own with good prompts, can enhance some NSFW scenarios if adding it to compliment another lora (often with a bit lower weight)... I'd say it's most useful for the kind of bouncy or back and forthy motions that NSFW activities often have.
I say that because quite often if using this to strip 2D characters down or make them flash their breasts or so, even the top pull/removal motion ends up with this back and forthy sway.

Still, I'd say it's quite versatile and can recommend you try it out. Oh and it does seem to work pretty well for both real and animated.

@sarcasm83334 Ohh I see. Thank you very much for taking time to explain me this. Really helpful! =)

u need to search on github

You need to go on github. There are links to the huggingface repos under the "Model Download" section.
Here is a list after googling it:
https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B
https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B
https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B
https://huggingface.co/Wan-AI/Wan2.2-S2V-14B
https://huggingface.co/Wan-AI/Wan2.2-Animate-14B

Here's a good source for the I2V GGUF on huggingface as well! The Q_4 and Q_5 are good!

https://huggingface.co/QuantStack/Wan2.2-I2V-A14B-GGUF

@Buttchuckjones Thank you. Are there any instructions so I can create NSFW content later? I understand correctly that I'll download the standard version. Maybe I need to download something else, like from Stable Diffusion, beyond the checkpoint?

@dimxmix668 I made a ComfyUI workflow for just this case, cutie! 💖 https://civitai.com/models/2237688/premium-wan-22-i2v-workflow-by-freyja-pixel

It has very detailed documentation (readme and prompt guide), checkpoint and GGUF loaders, LoRA loaders, VRAM and RAM optimization, web resource lists for custom nodes and NSFW LoRAs/checkpoints, and recommended generation parameters! It comes pre-zipped with all of my test videos for you to drag and drop into ComfyUI to see what settings looked good (or terrible 😭) during my testing! I hope this helps! 😽

This is for Wan 2.2 a video generator

A couple methods I'm aware of:

1) Try to get the input image as close to the action you're trying to achieve.

2) Use an extended video workflow (where it takes last frame from first video, then inputs that last frame into the next video). This will give you another 5 seconds, right as the action is starting.

@ChillDesire I've yet to learn how to do that. I'm definitely interested though. Is it complicated?

@takearushfan374 I have a workflow that uses the FP8/FP16 model that makes 15 second videos. It's 3 KSampler pairs one after the other, with a combine node at the end.

What models are you using? The FP8/16 models or a GGUF model?

If you're interested in it, or interested in a 10 second version, let me know and I can upload it.

@ChillDesire I'm using fp8 at the moment. I'd definitely be interested. I hate how the clip ends before any eh hem action actually takes place lol

@ChillDesire I'm a complete beginner, and I found your answer very helpful. I would be extremely grateful if you could share how to set up a workflow. (This text was translated using Google Translate; please excuse any errors.)

@takearushfan374 @lmpx68 Here's what I used for a simple "Longer Video": https://civitai.com/models/2188676?modelVersionId=2464349

I may eventually add extra Positive prompts to guide each section.

@ChillDesire Thanks. I'm gonna try it in runpod right now and see what happens.

I made some ComfyUI resources that are pretty simple, since I'm trying to make it as drag and drop as possible. This one is fully uncensored and NSFW! There's two versions with a detailed Readme and prompt tips! It's designed for ComfyUI, Wai-Illustrious, and Wan 2.2! https://civitai.com/models/2192188?modelVersionId=2468337 😘

You'll need to set up ComfyUI first: https://docs.comfy.org/installation/desktop/windows
The downloadable executable install is pretty stable and easy to set up, comparatively speaking. The source git install is apparently better, if you're tech savvy! I wouldn't use the portable version! It's pretty unstable and glitchy!

Also you'll need the Manager (for custom nodes, models, and workflows): https://github.com/Comfy-Org/ComfyUI-Manager

I'm working on other drag and drop workflows! I already have some for text to image, and I'm developing some more fore image to video and text to video! 💖

I have one just like that with text to image to video, and it's fully uncensored and NSFW! There's two versions with a detailed Readme and prompt tips! It's designed for ComfyUI, Wai-Illustrious, and Wan 2.2! https://civitai.com/models/2192188?modelVersionId=2468337 😘

I'm working on a standalone version for i2v and t2v for Wan 2.2! Follow for more, and I'll try to post when I release it! 💖

Better late than never. "General NSFW" means general NSFW actions, such as handjob, blowjob and other NSFW actions. This lora is made to make it easier to get some NSFW actions with img2videos generations. No need for u to get dataset for understanding what it does, u can just look at the gallery.

@CubeyAI thanks for the explanation. Makes sense.

@CubeyAI I don't mean to fan-boy at your work, but this LoRA is the foundation of every single NSFW video I've ever made. Even SmoothMix that just dropped still significantly benefits from .. sliding .. this LoRA into High/Low with 0.5/1.5

I don't think people realize they can mix the strengths on High & Low, as they perform different tasks. Either way, at " Overwhelmingly Positive (1,774) ", you can assume its the User and not the LoRA if ppl are complaining.

@NeonScreams In fact in did experiment with that not this way tho but i guess that the point of sharing experiences with people via comment. Again i 100% accept the fact that i may be doing something wrong like i already said. Time to try what you mention.On a side notye to can't expect people new to generating stuff to outright know everything :D

Seems u're not the only one. After more than 100 clips using only smoothmix without loras, i thought i would start trying some. I actually get better results without that one. Maybe as u said for yourself "im doing it bad", but in the end i don't like the outputs. If u can't get good results, after multiples try, just go next. it's AI world, everything changes with updates, settings and workflows. We can't all manage to get good results with same resources. Since it is a popular one, might give it more try later and will let know if i got something to share.

@CubeyAI dated or not, i see many good videos made with it.

@JustCiv i would say try for yourself but it really depend sometime i get really good results but most of the time especially when combine with other lora it doesn't work as good. Still trying to figure out stuff myself.

Edit: After trying some other checkpoint i get way better results some it was indeed something i was doing wrong.

@CubeyAI I found the perfect little strange comment thread where to thanks you !!!!
As some other users said, your lora is mandatory, not only on civit, on sooo many platforms, like seaart etc ...
While it's to your credit that you humbly respond to negative comments and question your own work, the fact remains : Sooo many people are continuing to use your lora cause, whatever the way it had been caption or the the flaws you find in it, it's the one with the the most pleasant results, that respect the more the original composition, giving it that so cool natural feeling and fluidity of movement.
Many dedicated pr0n movement enhancer lora are, for example, breaking a futanari crotch, cause they are over trained on woman anatomy, or degrade faces.
Your lora remains a banger, thanks for your work and without being pretentious, you can truly believe in yourself !

You need to use booth. Wan 2.2 need two checkpoints. Two Loras...

Thanks. And how do people generate 4K videos? I can't see any models specifically for that. Are they using an upscaler after generating the video?

@seanhalley they use a high res picture.

high noise ✅ Best for:

Poses and body structure

Art styles / vibe

Dramatic transformations....Low noise ✅ Best for:

Character faces

Specific outfits or styles

Realism and clean renders

one observation i've had with this LoRA is it really likes videos with two subjects...if your video has only one subject it's as if all of the motion that would be going into a second subject gets concentrated into the single subject, and you get all kinds of jitteryness that you won't see as much with duo

were you able to fix the jitteriness when two people are in the image?

The examples above are mainly just the default workflow with a lora node added. prompts should include nsfwsks and description of the action "penis goes in and out of vagina"

https://civitai.com/models/1981116 you will find the models, and the file workflow. to understand better use youtube

@CubeyAI I seeee, makes sense, thanks for the explenation!!

you should put a Rife VFI node into to your setup (right before video combine), set the multiplier to 2, set video combine frame_rate to 24 (or 60 if your setup can handle it). That'll smooth things out (video length 81 frames). Keep CFG at 1, total steps 4 (change on both samplers), 0-2 (high sampler) and 2-10000 (low sampler). if using lightning loras, i normally keep them around 0.75 (high) and around 0.30 (low) depends on what im doing and how many loras im using.

@oatis81259 i have a really bad experience with RIFE VFI somehow it always breaks my comfyui.

wan2.2 i2v high and low are you'll using that for diffusion models or something different, and thanx for the help too

@KDwright42 i use GGUF since I only have 12GB of VRAM. DasiwaWAN22I2V14BTastysin (High/Low search on civitai) is good. or 'wan22RemixT2VI2V_i2v' both are good.

@oatis81259 thx

Just download the workflow from this person. https://civitai.com/models/1981116 all is already set up

@CubeyAI if you got a dataset for me then I can do it. you need at least 60gb VRAM do train a lora

For the purpose you are mentioning is better you use a model that is already been trained under loras like this. I recomend you to use this checkpoint, this channel also has their custom worflows. Then you can add the image as you mention and specify in the prompt you want her get nude.

https://civitai.com/models/1981116

@Megaarmed oh man you saved my life, that's a time saver! bless you!

it takes 12 minutes per video on RTX 5060 ti 16GB for 5 seconds at 720p and 16fps.

@FranciscoM Hi Bro which checkpoint are you using with this lora, can you share your workflow

yes, runs extremely well with any of the gguf formats as well!

@tsunadehusband26 I used default wan2.2 workflow, with default fp8 models, I replaced the lighting loras with this ones. I guess thats why is slow, maybe there are better workflows out there.

@FranciscoM hi, i recently got a pc with a 5060 ti 16gb as well. i can't get any workflow to output results without ghosting or blurriness, you said you were using the default wan 2.2 workflow? can you link that please?

@Itsuki69 The default workflow is available in the templates menu and also here: Wan2.2 Video Generation ComfyUI Official Native Workflow Example - ComfyUI
You can find links to all the required models and LoRAs there. However, I haven't quite figured out how to properly chain the Lighting LoRA with NSFW LoRAs yet, there must be a way. I’ve also had great results with LTXV2; while it’s less accurate with complex prompts, it looks excellent for simple animations and it takes only 3 minutes for 1080p @ 32fps upscaled:
LTX2 NSFW Upscale + Interpolation - v3.0 | LTXV2 Workflows | Civitai

yeah, i tried using them on my comfyui workflow and it became a blurry mess

@oscarstg1455 If you're so damn horny that you can't even bother to look for the shitton of resources out there then you're a lost cause. You could even ask a damn AI to help you, lmao.

Nobody is going to hold your hand through this. If you're expecting plug & play you are at the wrong place. You're going to have to learn how to integrate Loras on your own. Gemini, Claude, Grok, etc all have internal knowledge on how to do it. Consult them or go pay one of the folks on patreon who make side income helping folks out and/or creating "working" flows.

Besides, how the heck do you expect an auto-install anyhow? Not everyone uses ComfyUI. Some people use Wan2GP. Others use other things. The Lora will often be platform independent.

Make sure your file have .safetensore file type at the end of the name

This lora is designed for use with Wan Video 2.2 I2V-A14B

жаль конечно

Hi, what means damgood? thanks

Still waiting

@Kingp0dd529t ok...

ComfyUI\models\diffusion_models

light models go in regular loras folder

if you look in the top right in the stats it says it's a lora.

@IHateThisSite but what is it used for? i dont understand from the description what this lora actually does....

@TheSemenDemon my bad, yes it's a general nsfw but in my experience is generally best paired with what you're trying to use, like a bj lora reinforced by this one, feels like it helps balance things out.

@IHateThisSite so this is kinda like dr34ml4y? or is it like smoothmix? meaning what kind of Lora like which slot does it go into

@TheSemenDemon No one EVER actually explains what these do. Its the most annoying thing about this site.

@TheSemenDemon I tend to use this and DR34Ml4Y together

@darthmobian57825 what is this DR34Ml4Y?

@RussianMan_BRUTAL its an all in one nsfw lora. really versatile, really good. if you are making nsfw content you can use that for 80 percent of your stuff and only that and it works great.

Its like an nsfw enhancer. So you'd but the "NSFWsks" in your prompt to activate it, and it gives WAN a little more context/understanding for the things your asking about. The difference between this and a lora trained specifically on doggystyle is that 1:This would be better for anime/hentai styles and 2:This isn't just limited to doggystyle. You can use it with other lora too.

It sounds like you have it. What are you having trouble with?