@c0ur4ge Awesome! Thanks for all your work!

@c0ur4ge really love your work. I have to be the one, but any plans for a T2V v2?

variety is the spice of life. mix and match XD

what is NSFW Catog? a typo?

We did it! Thanks everyone!

Cum/Facial - Wan2.2 - v2.0 - High | Wan Video LoRA | Civitai

this is the best one ive found so far, lower the high strength a bit if theres too much :)

@sirbazzrick No, my point is that the new dr34ml4y produces way more than before. From initial experiments it seems that I need to set i2v v2 weights way lower than for v1.

@WetFeverDream ah right im with you.. sorry didnt read it properly lol.. its very early here :D

The Lora seems to turn any dripping or gathering fluid into cum now. Even rain dripping from a ceiling.... its kind of annoying actually...

@bewunderteradler112 Yea, it was a bit much.

@c0ur4ge I actually use it together with the SmoothMix so I don't know what else 'hides' in the model. However, if I reduce the lora strength to 0.85 and remove the low model, it becomes much better. If I remove the lora I the model performs normally however... I will try with the base model just in case.

@c0ur4ge could you recommend a model setup? like, which lightx2 version, strength, which base model that works best?

@c0ur4ge well I can say that the lora also does this with the base model, although much less "prominent". I still have to try it without the lightx2 lora and see if that makes a difference...

@c0ur4ge Applying version 2 changed the colors of the original image; basically, everything looked much more washed out and muted. Significantly different from the reference image.

@c0ur4ge I ran several tests, and lowering the strengh of this LoRa minimized that unwanted effect.

Then I further minimized the effect by specifying in the prompt that there should be no changes in lighting or color.

But what became clear to me from the tests is that it was this version 2 that was "washing out" the video's colors.

I admire your work with this LoRa; it's very good, and I love v1. I'm bringing this up to see if this problem can be fixed.

https://i.ibb.co/4wfj3QZG/0000test.jpg

Try lowering the strength and better prompting for camera movement. Like at the end of a prompt the last line could be "the camera pans up to her face as she [expression] directly at the viewer". also remeber too that using periods and commas make a big difference in prompting. You can also use ( ) parentheses and strengths to good effect to increase the influence of a certian word or phrase of a prompt. Example... (the camera pans up to her face as she [expression] directly at the viewer:2). If you prompt something and it doesnt want to do it, try doing this. Good luck!

i used to use other lora with V1 but with V2 i don<t need much else. except anal stuff, does'nt seem to work

V1 works great on its own. But I'm having some problems with v2 and certain poses. The prompt is very important, yes. You should do some fast iterations with low resolution a few steps of high noise and just 1 step of low noise to get an idea of what general movements the prompt can bring out.

I think I heard one of the updates was messing with the Lauras and comfy UI that's what you're using. I think maybe it's a bug with the new release because they redid the interface and stuff as well So maybe with enough feedback to get that fixed

@Codeman119 Laura Smith doesn't have a problem with it but Laura Johnson is definitely someone who would redoes her interface every other day so I wouldn't put that past her. But don't be fooled, Mrs. Johnson can do whatever she wants with her makeup, it doesn't change the person she is.

I hope this makes sense to you.

@AlbanianDoorknob Well I guess that is what I get for using voice recognition and not double checking the wording before I post. But I know You know I meant LoRA's.

but I have been using this LoRA with the new CumfyUI and it seems to work fine for me.

I too have been having some issues using this High and Low Lora. I am using the gguf version of Wan2.2-I2V-A14B-HighNoise-Q8_0 and I am using DR34MJOB_BJ_I2V_14b but the movements are just not there for me. I will keep messing around with it.

@DirtyBeagles any luck?

@c0ur4ge Ok, but I mean, what settings would you recommend for your lora ? By the way, I am personaly using wan2GP interface.

Its just like any other lora. Do u not know how to use loras in wangp?

I dont know how to use the right high and low CFGs settings,Loras Multipliers etc....

@thugstyle28322 wan2gp is setop for high pass and low pass and you cont want the high and low lora strengths mixed up. For simple basic setup for the loras. Select both high and low loras, Select the high lora first, then the low lora. Then underneath for multipliers you would write exactly this

1;0 0;1

In this example the ; symbol seperates high phase and low phase . So 1;0 would make the high noise lora only activate in the high phase and not the low phase. And then 0;1 would make the low noise active on the low phase. Each chunk of numbers seperated by a space represents the next lora u have in your list. And you can change that 1 for anything because thats your multiplier. If you wanted to do 0.7 strength on both loras you would write

0.7;0 0;0.7

On a high lora you put the strength multiplier number fist and on the low you put the strength multiplier number second, always. Lastly the seperator is a SEMICOLON ; not a regular colon :

Hope this helps.

@jynkz41 thanks ! and how schould i set the cfg value ?

@c0ur4ge yes

ComfyUI

The issues you're describing are issues I've always had with i2v regardless of anything I attempt or use alongside. My understanding is that these are just key limitations of the process that need to be worked around.

what did you do to make it work

possiby share workflow?

You Ignore those warning messages, they dont effect the output.

@c0ur4ge 我只是猜测，专业还是你

@c0ur4ge lmao holy BASED response

Seems to be stricter with I2V this time around. I use an older 2.1 Lightning LoRA workflow that uses 6 steps, CFG 1. Worked great with all versions previously. Version 1 still works great with zoom outs (+ erect penis and gen NSFW lora). This one seems to dislike my workflow. Minimal pan out at times and you see the penis poking a belly repeatedly.

@c0ur4ge Thanks, I wish it were that easy. Do to technical issues I was unable to overcome, I can't use ComfyUI, so I have been using Pinokio Wan2GP with the Gradio UI with relatively good success, but the workflow is different. I have found through trial and error that the strength settings for ComfortUI don't always comport to Pinokio/Gradio. I'll keep at it.

@c0ur4ge Yes, despite several attempts, and even a new install, I could never get wanvideowrapper to show up (I think that's what it was called) and without that module, I couldn't really move forward. I'll give it another try at some point. I just needed to take a break from it.

@c0ur4ge I dont see any linked workflow, can you please tell me where to find it?

1. Higher resolution video.
2. Lower the LoRA strength (mainly low noise as that affects the details like faces.)

@aurelius thx, would 720p be enough? how much lower? i mean normally i use the 0.7 of the strength on both lora noises, something like 0.2 would be enough?

Higher resolution video just spreads the attention pressure over a wider field so it helps indirectly.
Translated into system terms:

High noise → pose, silhouette, composition

Mid noise → identity scaffold

Low noise → face, skin, micro-expression, texture

If you let a LoRA push hard at low noise, you are literally saying: change that shit bro. keep high noise at a reasonable strength, 0.7ish or experiment with what works for you, and then drop the low noise LoRA to like, 0.2 or 0.3. basically you want the model to decide early and then keep those decision and not rewrite or renegotiate identity in low noise. This is also why I personally use 3 samplers and precision gating (I wrote an article about this recently: https://civitai.com/articles/25076/q2k-greater-fp8-the-precision-trap) but the main thing is 3 samplers, so at that middle ground towards the end of High Noise you can separate that out into its own sampler and adjust it until your first two samplers lock down motion and identity etc and then let low noise take over and polish what is already decided without renegotiating. but the simplest thing you can do right now is definitely drop the low noise strength. I would also advise using LoRA loader nodes that allow you to adjust the model strength AND the CLIP strength of the LoRA independently and drop the CLIP strength of all LoRA in low noise to near zero, CLIP is basically what screams at the model to do things, if the LoRA is still screaming at the model to change things at low noise you are going to have a bad time.

@clzpetn804 NGL, that's too much info for me to handle, i'll test a few things then with what new info i got, thx.

@zereshkealt definitely fair. sometimes my autism explodes on unsuspecting passerbys, apologies XD

@clzpetn804 ok thx, kinda worked, i've had a few generation with the new lora strength, it definitely did something and the faces changed less than before(a lot less), at the same time, idk how to say this, i think the faces become less active, the AI has became limited to do anything(as what i understood from what u said before). now i think i need to test the other custom nodes which has the ability to use strength in lora and clip loader independently, do u have anything specific in mind?

@zereshkealt the nodes that let you adjust the clip and LoRA strength are just called LoraLoader they are part of comfy-core,  when you drag it out onto the workflow it should have a place to select the LoRA, set LoRA strength and also set CLIP strength. they are a tiny bit inconvenient compared to something like LoRA Power Loader that just lets you use one node and just keep hitting add and stack LoRA because you need one node per LoRA but I think its well worth it. and once you get some of this sorted out you will probably need less LoRA overall.

How many steps are you using total and how many of the steps are High Noise and Low Noise? Do you use speed up LoRA like Lightning or lightx2v? if you do what strength are they set to?

@clzpetn804 ok, i've a few custom nodes, which has the option, i'll use them, i am using lightx2v lora with 4 steps and recently i started to use 3 steps on both noises, 6 steps overall and i leave the lightx2v lora strength at 1(don't know if i should change this one?). as for the these built in lora loaders, they are inconvenient, stacking them, might have new problems, right? i've had problems with conditioning nodes for having multiple prompts nodes! i have a question, i've never connected a clip text prompt to the lora loader before, now i have to run it through lora loader? only the positive prompt, right?

@clzpetn804 do you happen to have a workflow for WAn 2.2 with the thing u explain elaborately on several comments ago?

@zereshkealt The LoRA strength necessary depends on what you're doing. It's a case-by-case thing. You need to experiment. Good thing is that you can regenerate without having to do the high noise pass again if you use the same seed and prompt.

@indonesiangirlonly I just made an entire article and workflow for this yesterday as a matter of fact https://civitai.com/articles/25606/updated-wan-22-i2v-workflow

@zereshkealt there are a lot of ways to mess up conditioning but this isn't one of them. you can connect all of them together without issue. its essentially the same thing as using a LoRA loader that allows multiple to stack, the node does the same thing internally. I also just posted an article with a workflow that contains this all set up https://civitai.com/articles/25606/updated-wan-22-i2v-workflow 

@clzpetn804 thx, i'm testing it right now.

@clzpetn804 ok, a few questions, u have 3 models loaded in this workflow, right? 2 high noise gguf models and one base model! does it matter if i use the same gguf model for the high model? second thing is, do i have to use the base "fp16" wan model for the low noise or can i replace the node and use gguf low noise model? (i'm using the q8 models) there are a few nodes which i don't know what they do exactly, namely the unetattention nodes.
edit: well i've done a couple of generation and the result, was the same as always been, the faces changes so much. the funny thing is, i get the exact same faces when i use a couple of people faces for generation, the faces turn exactly to the non existed person!! if i wanted that person's face, it would be GREAT!! it's definitely the lora's doing, cuz it's not happening when i use the base model with lightning inserted in it. but what can u do, without the NSFW lora? thx, anyway.

@zereshkealt I dont know what to tell you, I have never had this LoRA change anyones face in any video I have ever made and that is basically the exact same workflow I use for all of my videos. and yes you can use any models you want to use but those are the exact models I use for my videos.

@clzpetn804 thx, still the 0.7 strength for high noise and 0.2 for the low noise dr34ml4y lora, works best for me, but the faces become so smooth and young!!! with ur workflow and same setting, i used the q8 gguf models, the result nearly the same as the default workflow i use, except i could say that it took about 10-20% less time to generate, but the upscale and frame generation were just too much and crashed a couple of times, not sure what was case. btw, when i said the faces changes, i meant the faces changes slightly, like the nose and lips and eyes... it's turning entirely to someone else, but it ruins the base image, i hope they release the wan 2.6, creating video from video, is much more interesting and could get better consistency.

@zereshkealt oh something I totally forgot to ask but is very important for distinction: are you talking about the face changing in one video suddenly or over multiple videos doing a continuation using ending frames or frames from the first video?

@clzpetn804 well if i could make it through just 1 generation, i'll go for the next one, but it's definitely changing after the first frame, well it's gradually changes with each frame, after 10 frame, the face becomes consistent. but i think it's the face that has been used for training the lora model, it's turning into that face(that's the extent of my knowledge), i've tried to lower the strength of the high noise and low noise lora to 0.2 and 0.05, but it got worse, the face changed a lot more :| ... either way, i think i've given up on that, it's weird, sometimes the lora works with some faces without any problem, no matter what value i choose for the lora strength, but some faces are... IDK! thx. u've helped enough. i might try to use IPAdapter, read a few things about it.

put it in the lora folder.

@rickets_xxx is it really necessary to use high and low? im new

@c0ur4ge thanks. I'm a beginner but the workflow looks too complicated. :(

@c0ur4ge can u eventually share me the workflow.) in now understand it a lil better thanks!!!

@c0ur4ge i dont see any workflow...

@c0ur4ge I gotta give you credit, you've got more patience than I do.

@Poyex1234 save one of the demo images, drag and drop it in comfyUI and you'll get the workflow.

@oliall Holy fuck balls... I've dragged and dropped images into comfyUI before but it didn't occur to me to do it as you suggested.

Amazing 👍

need some information about what you are doing first to know whats wrong. what strength are you using for the LoRA? are you using Lightning or Self Forcing speed LoRA? how many steps are you using? etc

@clzpetn804 8 steps, 0.6 for the lora, lightning yes but this happened without it too, i was just trying to get a woman to bend over for doggystyle, i kept getting a weird blob under her that looks like it wants to be a mans body but not fully formed. i played with the settings but id get what looks like balls on the woman or like a dismembered malformed soft on her crotch

@bapybara theres a lot of reasons this could happen, as a first step empty your negative prompt box completely, make your prompt as minimal and specific as possible, and take one step away from low noise sampler and give it to the high noise sampler. try that. also I just released a new workflow and article, check out my other articles as well for a more deep dive, https://civitai.com/articles/25606/updated-wan-22-i2v-workflow

What is a top or top setup (models, vaes, loras) for I2V for nsfw in your opinion?

@FiberCreations2035 Easiest is Wan 2.2 since it has a year of Loras. But quality wise and with effort LTX2 is already outshining it in a lot of ways. Better to learn Wan2.2 first since LTX2 can only be approached when you're very familiar with how to work i2v.

@c0ur4ge What would too high be? I mean I've only actually gotten good results from the standard low rate, but to use that rate it seems like 150-200 repeats needed for a good solid learning; took 28 hours and 30k steps. Audio is only an issue if you use a Lora at higher strength over 0.7, guaranteed bad audio using loras at 1.0. And that's for any Lora even those not trained with audio at all. I'm aiming for 0.5-0.6 strength usage to get around that because you can still get voice tone and sounds without the loud roomtones and cellphone style sound if it's trained enough. I think that's a problem with the model or implementation though. One new thing to try during testing is the new LTX guider nodes and parameters: skipping layer 29,28,27 seems to bring a lot more true motion in from the lora and removes a lot of base model influence. Destroys the visual quality of the first pass in i2v, but you can restore it with a clean second pass that's based on the start image and a low strength distilled lora and a normal sampler, leaving you with just the improved motion.

As I was in your situation some ago too, fighting with super-complex workflows:

https://civitai.com/models/1823089?modelVersionId=2406909
Have a look at this one, pick the right form (I2V e.g.) and his latest checkpoint pair to use alongside this lora. Should work lovely. This cutie rly knows his jaaab !

Search for workflow "i2v skill"

@c0ur4ge insane work you guys are doing!

@c0ur4ge godspeed, my dudes. Beyond stoked to see what you can do with LTX2. Thing's a speed demon, should be a huge leap beyond what Wan 2.2 can do.

try adding pov after d0ubl3_bj, then follow up with something like 'A top down/high-angle perspective' or 'a close-up view looking down at two girls...'.
Try also including something along the lines of "the man's torso is seen at the bottom of the frame, his left hand holds the blonde girls head'.

Mess around with different phrasing/prompts, but that should get you close.

I'm running this with no issues on 12, skill issue

@c0ur4ge hey as a troglodyte what do you think are the minimum resources to run it, even if slow lol. Just wondering if I need to do a runpod or something. I know how to get everything to run otherwise.

I've used this with less ram there is a version where these arent loras but the full model, alternatively use the Q6/Q5 or less version of WAN2.2 and should work even in 12GB models.

Im running on 8gb vram 16gb ram with other loras, watching streams and browsing.

Wish some people like you would at least get a modicum of knowledge about what they're doing before complaining or spreading technical fallacious nonsense and misinforming others. People put a lot of work in these LORAs and you're using them for free. Have some respect.

@c0ur4ge It was a happy accident to find out with the double_blowjob and using "lick" in the prompt. I wish I could make a LoRA as quality as this! Thank you

@c0ur4ge The lora kicks in as soon as you set it to 0.7 or so unleashing a little the prude WAN2.2 model. No blaming you at all! Its just most men end up having vaginas and fingers being penises :D

While i have added the Custom nodes, i still see them under missing Node packs
primarily - Res4Lyf, ComfyUI Chibi Nodes & Various ComfyUI nodes by Type. Any suggestion on how to fix this?

Use this workflow: https://civitai.com/models/1823089?modelVersionId=2406909

I tested & works greats.

@mpayne1985106 Using this workflow and everything is coming out blurry. Like can't see anything. Tried multiple cfg values and lora strengths

@rumham1899 Blurry is often caused when not enough memory, try on low resolution like 512x512

wan 2.2 is two models, low and high noise. you need two lora aswell then. use low on low and high on high only, dont mix

@goeypants400 Okay, got it. Thanks!

@matai515 I would suggest going to the DaSiWa collection and using their guide for video generation. The i2v workflow is foolproof and has everything you need.

I'm using this workflow, working Ok with DaSiWa Tastysin model + this lora.
Just read the whole description about the workflow.
Also, ask chatGPT how to install nightly version of pytorch and SAGE attention, if you have no clue (like me 😀), best and easiest on Ubuntu linux
https://civitai.com/models/2272369/wan22-i2v-gguf-nsfw-8gb-vram-32gb-ram-workflow

If you have SAGE attention issue, just change it to sdpa

Did you try lowering the lora weight and improving your prompts? lmk

@rkyha639 yes I tried everything, now I tried to add 4 different lora, one that shows the dick, the other that does blowjob and handjob together and the other that does deepthroat, I inserted strength 1.0 to all 4 lora, step 6 for both ksampler1 and ksampler2, while HN shift and LN shift to 7.0, I also activate the triggers, putting more triggers of the different lora, like if one has bl0wj0b and the other has deepthroat, active with bl0wj0b, deepthroat, <prompt> but it doesn't work, it doesn't work, I used up 10$ of runpod and I still have to solve it

@c0ur4ge oh nice its full on finetunes? No loras? This is gonna be wild.

Does the wan2.2 lora work with lightening loras? I can't get it to work when used with lightning.