FYI worked in t2v at 480p for me.

Just grab the official t2v or i2v workflow from the comfui docs and put the loras before the Lightx2v ones.

@ToxicBot before the light2x ones? I had them after! I'll try that and see if it works. But the problem with the official one is that it resizes to a square ratio instead of resizing to max width or height in the same ratio

@ToxicBot Doesn't work for me. It's crazy blurry for some reason.

@iamgregor I am also getting the blur issue. The video gets super noisy with the stock i2v workflow in ComfyUI

Yeah, prompt stuff like "static camera" and "static camera, camera stays in place" or "camera zooms slowly" or "wide angle shot" or "zoom out". etc.

You should be able to prompt for wetness.

Genius ^^

Without realizing it, you have discovered a way to keep the penis consistent for the I2V. I think you should train your future lora with the same principle "A man appears and she xxx his penis" this will keep the consistency and you will be the boss of the penis ^^

How is this comment relevant to this model?

Clearly it does work if the model gallery is full of hundreds of videos and this model has almost 1k likes.

It works, learn to create workflow by yourself

Look up "smooth wan workflow" on civitai. that one is the only one that worked for me

@Ayylmao1773 I will do that

this is one of the best loras out there. it works so well

As I know, fps matters, 16 fps is standart for wan, so it will be nice.

@forfreelsd368 well i am using 60fps workflow.

Probably because you're trying to run it on WAN 2.1 dumbass.

@heccinrealdude its 2.2...

I don't use Pinokio, and don't quite understand what it is, but WAN2.2 Loras have high and low components that are meant to be used together in tandem. I use ComfyUI. It's part of the way the model is created.

@jazgalaxy581 im trying comfy now, but i can seem to get a good workflow as its spitting out videos that seem like they are 5 fps (even though my settings are in the 20s)

@NUGGZ1616 If you're using the an updated version of Comfy UI, click on the "templates" men on the left side of the screen. It should pull up all of the pre-fab workflows that Comfy UI offers. Select "video" and then one of the Wan2.2 options, either text to video or image to video. Everything in the workflow should be ready for you. You will need to wire in the LoRAs yourself, though, by summoning "load lora" nodes. Path the "high" Lora into the chain going from the "high" model and the "low" Lora in the chain going from the "low" model.

I can help a bit, but I'm low intermediate level knowledge. I'm not a pro like some of the people here.

@jazgalaxy581 I tried your method. Loaded the default, added lora steps after the lightx2v steps for both high and low and put the insertion high and low in. Left everything else alone other than prompt. I'm getting high blur on the video for some reason. Should I change the strength of the models? CFG? Steps? Anything?

iam using currently this workflow https://huggingface.co/datasets/theaidealab/workflows/blob/main/wan22_14B_i2v_gguf.json vcan someone guide for 8gb vram how this lora will work?

You need UnetLoaderGGUFDisTorchMultiGPU — there you can set how many extra GB to allocate and check on which value it will run.
Also, try a smaller gguf quant, it will increase the chance of a successful launch.

That loader isn’t good for LoRAs — you need one with CLIP connection, not to mention it only supports a single slot, and that’s already taken by the speed LoRA. And yes, it goes into the LoRAs section, and you connect high to high, low to low.

@redsnail can you please guide me in detail iam completely new

Which workflow should I download

Looks like the OP means 21 as the END AT step. Typical 2.2 setup is 3 and 3, with the end step way overshot. It's how the model works. There is a lot of play there. This LoRa works fantastic for me (with 128rank bf16 distill of course, which is needed for 6 step generation) with an even 3-3 split, low noise ksampler picks up at 3 and ends at 10,000. Try a number between 21 and one billion.

I have found elsewhere that the so-called 3 steps actually divide high noise into 2 stages and low noise into one stage.

@jinandmrzhang8099204 You could do that, theoretically. Doesn't make much sense to me, especially splitting the high side, maybe in specific cases for specific LoRAs. And the OP is definitely not talking about that. And using the word step in the context of diffusion, without qualification, to mean anything other than denoising steps is not something anyone familiar with 2.2 would do.

@Ponder_Stibbons Could you share your 6 step generation json somewhere? ^^

@spragenhall873 I don't use 6 steps anymore. The latest distillations are 4-step. My 2.2 FLF2V has that setup, with respect to the ksampler section. Distillation can be very restrictive however, a lot of the time you will get very predictable and repetitive output. Good for FLF, boring elsewhere. Lowering the strength and increasing steps and low cfg a bit can help, but when I need more control I just go to full model setup, 30 steps. If you use sageattention, torch compile, etc. it doesn't take that much longer. You should pull the example workflow from the templates page and play around. So much depends on your hardware that it's hard to tell people what to do exactly. I have no problem running full sized models for 145 frames at 720p. If you have to swap out for quants to run that, the output will be different, so not very helpful. Same goes for distiller. You have to keep screwing around until you find what works for you. Don't randomize seeds with video. There's a useful tip. Sometimes magic seeds appear. Keep them. And LoRA strengths can radically change output. You really shouldn't have to screw with the ksampler settings that much.

@spragenhall873 I decided to doublecheck myself, I just ran a bunch of tests comparing 4step lightning lora splits. I did 6 steps at 2/4, 3/3, 5 steps at 2/3, etc. and played a bit with cfg and some new sampler/sched. combos/permutations. I haven't tried before. Not too surprised to find very little improvement at all over the 2/2 4 step split. I do like LCM with beta for both ksamplers though, not what I usually use but I got some really nice fine details, so it wasn't a total wash. And it's never a waste of time to experiment, provided you plan your metadata and naming conventions such that you don't wind up with unreproducible results.

yep its the best of its kind.. its like 99% spot on generations

What's your workflow? I have a wan2.2 workflow and have matched settings and prompts the best I can, but my results are blurry and weird. Full men instead of just the crotch area, weird blurry, hairy, and mutated "equipment"

I've had trouble with this actually. Dudes with wangs coming out of their belly button or mouth. Full men animated in when I only want lower body and offscreen. I don't know how much you can play with the prompt and it would take me a long time to find out. I'll keep trying, but does anyone have any tips for this?

Can you share your workflow with me? I have the same graphics card. How long does it take you to generate a video?

In witch format do you want workflow ?

Use a rope

no the question is when will you come over to team green and buy a 24GB or higher nvidia card?

@guimgoom891 Don't worry, the sky won't fall onto your head because gays exist.

@inneradult_gaze good to know, I'll keep being homophobic though.

@guimgoom891 This thread is 5 months old, please touch grass.

@inneradult_gaze 
I need to touch grass... Because I'm not on this site often enough to answer within 5 months ?
I don't think you thought that one through.
The hepatitis is hitting your brain hard it seems.

I like this LoRA at 0.3 for both H and L to help loras that have trouble with penises. You can check my videos for visual examples.

@boobkake22  I have this one set at 0.3, but what cumshot lora did you use with this one? And what were the weights? I have tried this at the strength you mentioned with the f3cespl34sh one (at 1/1 H/L) but when the penis comes out of her mouth it is distorted. and the cum comes from different places. I've also tried it with wan cumshot lora at 1/1 H/L and get the same results. I'll keep iterating but any other things to try would be appreciated.

@mystical8218204 WAN 2.2 Anime Cumshot Aesthetics – Precision Load I2V, seems to generally work the best for most applications, though it tends to randomly appear various places you don't want, and shoot out of nipples occasionally. Pretty versatile, even though it has a few drawbacks.

@Korros Have you gotten a successful strength for the Anime Cumshot to work with this lora? Happy to run some tests and bring my results here for everyone else to use. I fully realize WAN 2.2 as a whole is brand new so we are all doing a bit of trial and error.

I've done some extensive testing. So what I found that worked to EXTEND a video that has a blowjob with Oral Insertions to get a cumshot is the following:

Wan Lightning (High and Low) both at Str: 1

Oral Insertion (LOW ONLY), Str: 0.3 (as @boobkake22 mentions)

Anime Cumshot Lora (High and Low) both at Str: 1

Hope this helps. I also tried this with the f4c3spl4sh lora but the cum never looked as good as it did with the WAN 2.1 lora.

@mystical8218204 what is the sequence? so Lightening -> Oral -> Anime Cumshot Lora? what about k-sampler?

Ha. I was about to comment the same thing. This is a great LORA for mixing, I almost use it in every generation, because there are so few, that can generate an even slightly natural looking penis, when there is non in the first frame.

Even just to ‚produce‘ a penis into the frame and working my way from there.

Keep them great LORAs coming.

what does that mean 0.3 at H and L?

@g932 It means a weight of 0.3 at High and Low noise.

You can try the ALLlnone model and workflow, which integrates the high and low noise model of 2.2 and the 4-step lora model. Of course, the quality of text-to-video generated by this workflow is very low, but it can still generate image-to-video. 16GB is available and compatible with the LORA model of 2.1

If you have no problem generating it, try adjusting your sd3 sampling parameters.

i second this request

I'm using both high and low at 1.0 with the 4 steps light LORA also on 1.0 and I'm getting great results, shift 5, cfg 3.5 on both samplers

@Corbe you win the greatest person on this day for me bro

I've been doing Shift = 6.87, CFG = 5, Total Steps = 30, Steps = 13 (first KSampler), Lora = 0.3 and not using lightning.

I'm annoyed that futa is even a thing. Seriously, I don't want that shit showing up every time I'm off by 0.01 on one stupid input box.

@manchuwook  thats crazy. Just generates a blur for me.

what platform are you guys using? I havent head of "steps" or "shift" i just use lora loder only on comfy.

@djwashout It's Comfy, just adding more custom nodes that set those values before they hit the ksampler.

can anyone post a sample workflow that helps explain some of these values?

share your workflow please for the 4 steps light

Please :)

how do you use automation?

Hello brother can you please tell me way to use it , please

increase weight of the lora? up your CFG?

three possibilities: resolution, lora weight, and steps.

if your res is too low, things might look funky if a lora was trained at something different. 640px on the long side to test.

if your lora weights are too high (for all total loras used, lightx2v included, things can start getting weird. if your latent preview is enabled (i recommend it), how does your high noise pass look? if it looks like what you're expecting, just blurry, then good. and the problem is probably with the low noise pass. your total lora strength for high noise shouldn't be more than around 1.4ish, and same for low noise. so if you have any other loras including lightx2v, try to keep the total under 1.4ish and preferably around 1.

steps - if you're using any combination of lightx2v at any strengths, i recommend 8 steps minimum. 4 is fine for testing prompts. the low noise pass with plenty of steps is where fine details like this will emerge.

so you want a decent resolution - not too low and not over 720p, a total lora weight of <1.4 for all HN and all LN loras, and 8-20 total steps.

i personally use 26 steps, no high noise lightx2v lora, low noise lightx2v lora at 0.6. high noise style lora (like this oral insertion one) at 0.8-1.0, and low noise style lora at 0.5-0.6.

THATS MY FETISH

@rjox thanks a lot for your guide . i have been looking for this

I 2nd this. I've tried several different ways to make the wanker smaller, thinner and curve another way and have yet to be successful. That being said, this is by far the best BJ lora that pretty much guarantees the action will be performed.

4080 super has 16 gb ram :D 5090 has 32gb

@Diogenes84 he means ram, not vram.

play around with the scheduler, mine is on dpm++_sde/beta and it helped with the rainbow.

CFG tweaks too

i think just describe where it comes from, i usually write out which side of the scene i want the man to come in from.

Extend the video and use a blowjob LoRA

Thanks for your help. I try that but it seems every 81 frames the sequence kind of restarts, so its not continuous.

@roger34341976 Hmm, I'm not sure what exactly you mean. What I'm referring to is you generate a video with this LoRA and take the last frame of that and use it as input for another video that uses a different LoRA for blowjobs. I can't figure out what you mean by the sequence restarts.

@magicballoon the problem im having here is that the penis always comes out of the mouth after maybe 2 or 3 sucks, so at the end of the video the cock is out of her mouth. I've tried putting oral insertion and another blowjob lora together but have the same issue. can you think of a way to keep the guys dong in her mouth at the end of the vid so it can be extended into another with a bj lora?

@rdeanrdean207 You might be using too many steps, namely on the high noise model. Try reducing those, or replacing them with low noise steps.

If i replace this lora with lightx2v, how can I use this lora?


can i have your workflow?

Yep what you said worked, thank you. What was the max res you managed to get without loss of quality? 640X640?

big stuff my guy! By the way could you also share the sampler and scheduler you are using? thanks

for me it generated some video but it appeared blurred.

@agent_agent794 i used a res of 768x512

@famest930 increase the value: 3.5 cfg and steps at least 18 or 22 for better results

@Lapalooza1 i just uploaded the workflow

@asimangunsong998109  i just uploaded the workflow

@ElMilaneso it's a shame that even with your settings it doesn't work exactly like the OP videos, i tried both yours and his suggestions, the setting he listed lead to worse results compared to yours. I was hoping he himself would make a workflow so we can see exactly how he is doing it since we are clearly missing something.

@agent_agent794Did you get good results with my workflow?

@ElMilaneso I would say good enough, just not the same quality as the promo videos uploaded by the OP of the loras (hence why I wonder what exactly is he doing differently). Btw with your settings/workflow i was able to push the res to 896X1024. I can't wait for WAN 2.5 to be open source, this will be a game changer.

@ElMilaneso https://civitai.com/models/2005198/lazy-wan-22-i2v-workflow feel free to try this workflow. It reduced the generation time to less than 5M, while outputting quite decent results.

@ElMilaneso thanks a lot bro. Its really helpful! COuld you let me know how I can efficeintly test like you did here? and arrive at optimum parameters?

worked like a charm running the q4 gguf with only 8gb vram. 5 second at 16fps took just around 7min to generate

I'll point all of you to my workflow. You shouldn't need to fuss with this shit; do not remove lightx2 lightly:
https://civitai.com/models/2008892/yet-another-workflow-easy-t2v-i2v-wan-22?modelVersionId=2315383

(Though it's not optimized for low VRAM, so ymmv.)

worked perfect enough for me. I used same settings. I'm on a RTX5090 with Manual Install version. man, been messing with settings on this for 6 hours now. Thanks

moron

You are correct on the first, and incorrect on the second point.

Q8 works very well, even with LightV2x, so quantization is not the problem.

@BobDF can you tell me sample name and schedular typeplease ,for load clip you are gguf or normal um5 xxl encoder fp8 e4m3fn or what sir?

I pretty sure it will? How are you prompting it?

good

yea nobody wants to see that shit

Simple fix, add Dreamly lora and reduce this one to around 0.3-0.4
Also use other WAN Checkpoints they have BBC included.

higher cfg I imagine

The scale of the girl in your image, the penis always comes in at a similar size so try to scale your images to that first.

i got a actual penis
here are instructions to set it up using Wan2GP

first page

insert image: "resize the image to the same size as resolution budget"

prompts: A man appears and she sucks his penis. Static camera, fixed viewpoint, still shot.

catagory: 480p (if you have low vram)

Resolution budget: 832x480

nubmer of frames: 81

number of inference steps: 21

Advanced mode: toggle this mode on

under advanced mode

now in general mode set:

guidance phases: Two Phases

Guidance (CFG): 3.5

Guidance2 (CFG): 3.5

sampler solver/schedular: unipc

Shift Scale: 5

Negative prompt:camera movement, 色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走

under click on

Loras

active loras: wan2.2-i2v-high-oral-insertion

then wan2.2-i2v-low-oral-insertion

under "Loras Multipliers

1;0

0;1

He used 24 Steps, CFG1 3,5 - CFG2 7 - no info on Sampler

@CITYDRIVER What is the benefit on using different cfg between High and Low noise?

yes, it is only 1 zip file, you unzip them and you get the high and low loras.

Mine are also always sped up

first page

insert image: "resize the image to the same size as resolution budget"

prompts: A man appears and she sucks his penis. Static camera, fixed viewpoint, still shot.

catagory: 480p (if you have low vram)

Resolution budget: 480x832

nubmer of frames: 81

number of inference steps: 21

Advanced mode: toggle this mode on

under advanced mode

now in general mode set:

guidance phases: Two Phases

Guidance (CFG): 3.5

Guidance2 (CFG): 3.5

sampler solver/schedular: unipc

Shift Scale: 5

Negative prompt:camera movement, 色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走

under click on

Loras

active loras: wan2.2-i2v-high-oral-insertion

then wan2.2-i2v-low-oral-insertion

under "Loras Multipliers

1;0

0;1

@grandpagrandpa thanks you're a godsend. Have you tried this with the lightning presets or different settings for quicker gens?

@likebot97 Adding a 4 steps lighting lora takes up like 26GB of Vram, so i have not bothered downloading it

@grandpagrandpa great guide, wonder if you could advise I have 32gb of VRAM when I try to do 720p it gets to 100% VAE decoding then I run into OOM any work around on wangp?

@grandpagrandpa what diffusion are you using

thanks

just use the basic wan 2.2 template with the loras inserted before the 4step loras. I just left the CFGs at 1 and the steps at 4 and it seems like a good starting point

Got you https://www.nextdiffusion.ai/tutorials/how-to-run-wan22-image-to-video-gguf-models-in-comfyui-low-vram

@icebergelishious899 you left the lightx2v loras in?

https://civitai.com/models/1497390?modelVersionId=2152516
this one works good on i2v

Same for me, I just found them here https://huggingface.co/rahul7star/wan2.2Lora/tree/1d1e553d07c1bb0e91765752413c4857e31df299/wan2.2

@roger3435331  Thanks man, appreciate!

Try high noise at 0.3, low at 0.3 and lower accordingly for penis only.

@MiddlingMaker  I will try that thank you, to be honest I default at 1 and work down, lowest I have got so far is 0.6 so will try your advice.

@MiddlingMaker Sorry, but I don't quite understand – should I set both the high and low to 0.30?

download to your local machine, unzip, then open it from RunPod. easy right?

@jagermister0901 no