How dare she

dtwr434 oh thanks but these 2 loras are not found, even in model manager WAN2.2 lightning I2V14B.. and Wan i2v lownoise pov missionary..the model i downloaded from your post has only 1 missionary model with different name

dtwr434 dtwr434 Thanks i figured out the tabs already and thanks for sharing these loras.However it says sageattention not installed though its installed via pip install sageattention. I have a blackwell GPU though. can i skip sageattention nodes?

dtwr434 I ran after disconnecting sageattention node..I connected others,,,but i get all noise in generation..is this json correct? https://pastebin.com/NZ02guzQ

@dtwr434 How to get the workflow from that vid? i tried to drag to comfyui did not work

dtwr434 dtwr434 it's loaded in the rgthree powerloraloader node, model is plugged into that loader, then the output model connects to the SD3 Sampler node. Each track has its own lora loading node.

GroundDog24 hi where did you get this lora from ? WAN 2.2 lightning I2V A14B 4 steps lora low fp16?

SuitableCar No its just the default 2.2 t2v fp8 high and low wan base models. No idea what one you are talking about.

Are you using this with any other LoRAs? I had the same issue and tracked it down to LoRA conflict, working on resolving it now.

dixienormus935 Nope just High/low Noise Base model > lora loader (1 model) > SD3 thingy > ksampler as per default workflow.

hmmm interesting, you could try looking at using RYS4LYF or another sampler, but I'm not getting that issue when using just this LoRA

dtwr434 Pony is fantastic for general erotic prompt adherence and non-realistic character loras but in my experience it struggles with realistic character loras, which is why I usually train them in WAN and then use T2V. But then prompt adherence becomes an issue. Have you been able to train realistic character loras for use on cyberrealistic pony?

dtwr434 Interestingly I am getting really good results with my WAN 2.1 realistic character lora with the new T2V cowgirl lora
https://civitai.com/models/1865114/cowgirl-reverse-cowgirl-sex?modelVersionId=2111171

Maybe it will work with your T2V as well.

Dang, this guy got thumbsdown bombarded out here... lol

jeyjeysig haters passing up easy cash lmao. stay broke, "AI Enthusiasts".

get a gpt sub, give it your specs, explain you want to use wan i2v in comfyui for nsfw and you know nothing, then have it run deep research on the topic.

@Narzhow did helping him go? 

@EthrealSkull i got no rig, can barely run Flux.1.

go to stablediffusionart.com. I'm just finishing his course now. It's like 15 bucks for a month's access and it's as thorough as you could want. It's very simple to follow, goes step by step and takes you through all versinos of stable diffusion, flux, and comfy UI.

It's not EASY and takes discipline, but it's basically a college course in ai image generation.

@jazgalaxy581 ah thanks, i know how to run it, i'm just on a mac with 18gb ram. my only option is Draw Things or Comfy at considerably slower speeds. not about to learn comfy; ALL these clients are gonna go away soon when someone streamlines genAI for mass consumption. we've had fun but we'll be obsolete as our acquired 'skills' with current GUIs will be useless.

@Narz You're absolutely right and that's a frustrating element. But I think it will be a while before we really get a general purpose consumer grade ai model that does what a trained user can do at the moment. The weird thing about generative image-making, and you can see it here on Civitai, is that it's "wow" factor is high when you are being bombarded with a page of random stuff. But as soon as you start trying to make it do something SPECIFIC, it really starts to show how limited it is. And maybe that gets solved overnight. Ai is moving quickly enough that it may be. But I really start to wonder if maybe we're there's a small amount of smoke and mirrors happening that makes it look like we're further along than we really are.

@EthrealSkull chat gpt does not know the existence of wan2.2 in his last training. Will only say bullshit about wan2.2, doesnt know what it is yet

@atenea_27 You can feed it documents/files then also have it do deep research on the topic, I think files can be included in the deep research as well.

@atenea_27 I think you're out of date, brother. ChatGTP is online now. It pulls it's information from the internet, not it's last update as it used to years ago. It is very current and knowledgeable about Wan2.2.

Still need help? :>

its hard to get pulling panties to the side with base 2.2, kind of a seed lottery deal. you might have better luck trying to do that as one gen, then use the last frame of that to do the second gen

Honestly this is the only one I've needed for 2.2 I2V: https://civitai.com/models/1818841/wan-22-workflow-t2v-i2v-t2i-kijai-wrapper
Just grab the regular I2V one, the others are hit or miss.

@duhsmith84631 Thanks

Try it without lightx2v, in my experience it really messes up movement. 20 steps total (12 on high, 8 on low) has been working well for me recently.

@TwoMoreLurker Ain't nobody got time for that :/...
20 steps means like an hour generation time for 121 frames on my 5070ti....

@haenlesn937 Yeah 131 frames takes me like 35 minutes, it sucks. Especially when it finally finishes and you get some weird AI taffy dick or a crazy look on her face XD

My understanding and experience is telling me these days that if you get bad results, it may be your image and not your prompt. Gen Ai predicts "the next step in a sequence." If it's not obvious to the computer what the next step (frame) should be, it will start to choke and the actors in your scene will start to behave like they don't know what they are supposed to do. You need to try to go for stereotypical scenes where it's obvious what is supposed to happen next. In this case, it will be scenes closest to the training data.

不使用lightx2v，CFG3.5 ,高低都是12步。这个LORA比较耗时间

haven't try one with clothing on for input image, but yellow/skin color clothing may be your generation model's doing. There's a lora to remove clothes if you really have to start from fully clothed, but if the genital is already visible, or the pose match, i think there should be no problem using this lora. i have some that have upper clothing intact, the breasts are still bouncing, and everything else also seems fine (kinda like wan2.1 version better tho)

Yeah I got the yellow clothing thing too.

I'm sorry. It was just a wrong observation. Great job with the lora man.
Btw, can you share your workflow please?

can you help me setup this in comfyui or any workflow rcomendations pls

no that is not normal

how did u get the workflow? i tried to download the vid then drag to comfyui did not work

I solved this problem. Use wan 2.5 on higgsfield to generate a video of the person removing their clothing. Then, export frames and chose the frame you want. Use it in first to last frame and it will work.

You can face swap your character into your starting image.

mind sharing your settings? Are you on comfyui?

@hmoobguyhero523 I use ComfyUI. What settings do you need to know?

cfg, steps etc. Do you also use loraloadermodelonly or go straight from load diffiussion to model samplingsd3? Sorry, I'm kind of new to this whole thing.

proper vae might solve your issue

Put a picture of your workflow in Grok or Chatgtp and it usually can tell you what's wrong

Sometimes this will happen if you are using an image that is the wrong size / wrong ratio for the video you are prompting it to generate. For instance, if you use an image that needs to be 'cropped' and the important bits of the image are cropped away (for instance if you are prompting about vaginal sex and the vagina is cropped away) then the video will pan down to that part of the body (because you sorta told it to in the prompt).

Sometimes the workflow you are using is trying to crop as best as it can (for instance to 640 x 480) and if you give it a 800x800 picture, some of the good bits are going to get cropped away. Try using an image that maintains the same ratio (e.g. 1280 x 960 which is 640x2 by 480x2) and it will not crop anything - it will simply resize it.

Or if you give it a picture of a face and talk about vaginal penetration, then its going to pan to a vagina (and even make one up) as best it can. If the face is too far away (if you are going 480 x 640 and it has to pan too many pixels away, goodbye face).

TL:DR: Use a correct size image.

@Farnuge Thank you. Actually, through continuous attempts, I have also figured out the reason. What you said is correct. This problem is mainly related to the initial image I used. The proportion of the characters in the image needs to be appropriate. As long as I find the right image, this problem won't exist.

Add a penis lora i guess

@rolfwolf820229 how ?