Support me so I can make more models, faster: https://ko-fi.com/the_cook
i2v POV Cowgirl insertion LoRA, similar to the POV Missionary insertion LoRA. There are two High Noise LoRAs: A is trained normally and has good movement, but might influence the likeness of the person. B is undertrained and has least amount of influence on the person's likeness. So to get the best balance between good movement and good likeness, use both (See workflow attached): For 8 steps total, step 0-1 is run with model A, then steps 1-3 are run with model B, then the rest with the Low Noise model. If you're not bothered by likeness and just want movement, you can use model A on its own.
Use with Lightning/Lightx loras, but set the strength of the High Noise Lightning to 0.5
By default the man will hold and insert the penis, but you can prompt for the woman to handle the penis. The woman will be in an upright position, but you can prompt for her to lean forward, or lean back, or a combination. You can also prompt for up and down movement or just a slow sitting action. Example prompts:
| Immediately cut to a new scene where the exact same person is now completely nude and squatting with their legs in an open position, with her naked vagina visible. Her face looks the same as the first frame, she looks at the camera. At the bottom of the frame a man's groin with an erect penis can be partially seen, as the rest of the man is out of frame and not visible. The man takes his penis with his hand and inserts the penis into the woman's vagina, the woman squats down on top of the penis. She then moves up and down, as she pushes the penis into her vagina repeatedly. The man holds on to her legs with his hands. The scene is a low-angle view.
| Immediately cut to a new scene where the exact same person is now completely nude and squatting with their legs in an open position, with her naked vagina visible. Her face looks the same as the first frame, she looks at the camera. She is leaning forward at first. At the bottom of the frame a man's groin with an erect penis can be partially seen, as the rest of the man is out of frame and not visible. The woman takes the penis with her hand and inserts the penis into her vagina, squatting down on top of the penis. She then leans back, as she sits with the penis in her vagina. The man holds on to her legs with his hands. The scene is a low-angle view.
the camera follows the exact same person as she moves forward and leans forward to squat with her legs in an open position over a man's groin with an erect penis. The woman squats down on top of the penis, and the penis goes into her vagina. She then leans back, as she sits with the penis in her vagina. The man holds on to her legs with his hands. The scene is a low-angle view.
Description
Wan 2.2 i2v High Noise v1 B
FAQ
Comments (78)
With the workflow I sent you, high noise affecting the face at low resolution doesn't affect the final output!
I don't understand what you're trying to say. That sentence is incomprehensible.
@Tompte No, their sentence is quite clear. They sent the creator a workflow. That workflow, unlike the creator's WF, does not have the issue of high noise impacting the face when resolution is set to a lower value thus not negatively impacting the final output.
Do High Noise A and B need to be used together simultaneously, or are they meant to be used separately in sequence?
Separately: A should do the first step, B should take over after that. Because A will establish the initial movement, but you don't want it to affect likeness, so B needs to take over, because B is less strong.
It still changes the face of the original picture, even though it is very similar to the original picture. However, this lora is still excellent. The author provides a new idea, which can be regarded as another way to generate videos based on reference pictures. I look forward to the author's subsequent updated version to optimize the consistency of faces.
Use B version for high str 0.35 to 0.60 for keeping face as close to original and throw in WAN DR34ML4Y - All-In-One NSFW both high and low at str1 and u will get good movement and action. Keep default prompts. U will get acceptable results.
This issue was solved a long time ago; you either crop out the faces in your training dataset or use masked loss. There's really no need to use two high noise LoRA on an already tried and true process for Wan.
Why not simply train loras where you blur out the faces or create them transparent holes with gradient. I heard some used that approach and it works well.
How the model will learn to preserve a face then?
有神快拜❤️
Question: On the examples with a clothed, posing subject in the thumbnail, is that image the actual first frame used for generation?
Yes, the kind of "cut to action" is trained into the model to make transitioning into the action from any image easier. But seems it's also not necessary, if you prompt it to happen more manually.
Edit: it said training data so I didn't get it. The workflow is the training data.
You say the workflow is attached, attached where? Is it some metadata stuff? I am probably not the only one confused by that, I know i can drag images into comfy to get the workflow but I don't know how it works with video
How do I run your workflow without going OOM? Doesn't it need to load in 3 models at the same time instead of 2? I might need more system ram, with the 2 models it barely fits into both my 16 gb vram and 64 gb system ram, adding another kills it
only the lora is changing, you would still have two models (high and low).
@xivami6907440 yeah but I think having to use the high model twice (like it is shown in the workflow) comfy loads it twice because it is using it with different loras applied. Have you done it with the recommended workflow? I tried with that one but it maxed out my vram so i tried the same setup but with the kijai workflow where I could use blockswap and that also didn't help so I don't know what to do
@Radyschen ya i'm on 16/64 just like you and it runs fine, no swap needed using the kj scaled fp8 models.
@xivami6907440 I do use those models, I actually put the exact models that were preselected, the only thing different are the loras where I used the new lightx2v high lora instead (the MoE one) but that shouldn't make a difference in terms of size. Yet it immediately maxes out my vram and takes forever to generate. I actually didn't oom, but it just took so long that I stopped it at some point so that my graphics card wouldn't crash. Where does your vram peak? I don't get it
@Radyschen loaded the same lora and runs fine. hitting 85vram with 480x720 81 frames. Double check your frame count and dimensions
The problem is the size of the image you're trying to use, his workflow is passing the raw image to the "Wanimagetovideo" node and large high res images are crashing and freezing it. Removing the link that's passing the raw image and instead linking the resized or down scaled image should get it working.
@azel1221 thank you, it worked. the workflow takes the dimensions of the original image instead of the dimensions of the resized image
I had good results with my wan 2.1 character models using the workflow I made to go with 2.1 models in wan 2.2. It uses high noise for 25% of the steps and low noise for 75%. In the samples I posted, the high noise is using High Noise A with no lightning acceleration for 2 steps, then 6 steps of low noise with lightning. The faces bodies of the characters are still accurate, and the rest of this model is working well. So that seems to be a pretty good combination.
Are you using the weight 1 with those steps?
@prabhjot5877 Yes. Also the comfyui workflows should be embedded in the videos if you want to look.
I just tried your method and it works really well. I just had to lower the weight to 0.8 to preserve the face fully. Thanks for the tip mate 👍
@darkroast175696 hi, i downloaded one of your videos to use the workflow, but there is this one node which i'm unable to install. i have even raised the issue here: https://github.com/aria1th/ComfyUI-LogicUtils/issues/21
any idea how i can make it work?
@HoneySandhu That question has come up before so I should probably release an update that avoids using that one node. Until I do that, you can look at the comments for my workflow where I explain a workaround (how to replace the problem node). Edit: Forgot to add the link!
https://civitai.com/models/1999304
I've never set up a workflow where different Loras are being using for different steps like that. Does anyone have a link to one that could accommodate that? I'd love to try it out, but don't even know where to begin with that.
This works well. In terms of maintaining the subject's original face, I find the high-noise A model works well enough provided you prompt while maintaining their exact likeness or similar.
Do you mind sharing how you trained the "immediately cut" part of the example prompts? While Wan natively supports cuts and scene transitions, an I2V LoRA focused on generally guaranteeing an immediate cut occurs at the very start of a video would be super useful.
This Lora is excellent, but I think it's a bit inferior to Wan POV Miss-ionary Insertion. I think one reason the face changes is because of the "screen switch." Even though I used the "the camera follows" cue, it might still switch for unknown reasons in the photo. Once the screen switch occurs, there's a chance the face will change. Someone shared the "girl drops her nightgown" cue in Wan POV Miss-ionary Insertion, which worked well in Miss-ionary Insertion and didn't switch screens. But the same pattern in this Lora has a high chance of switching screens.
Like @dravodir504 said, if one uses "While maintaining their exact likeness," at the very beginning of the prompt, the likeness does hold up rather well. I've been getting nice results with just using the high "A" and not using the "B". This is a very solid lora.
I think one of the reasons why the face changes is because of the screen switching. Even if I use the prompt that starts with "follow", there is still a chance that the screen will switch. Do you have a way to avoid the screen switching?
Can confirm. Getting good results just with the A high file + low file.
The LoRA is good, but it consistently changes the source face, im using 0.7 strength, in about 7 out of 10 of my generations, the face ends up different.
Is there any trick to keep the face unchanged?
Using high noise-A (S=0.35) greatly reduces the degree of facial changes. But the face still changes due to switching scenes.
Actually, I don't understand the logic behind making the words "Immediately cut to a new scene". I think this is the key point that destroys this beautiful model. Using high noise-A (S=0.35) greatly reduces the degree of facial changes. But the face still changes due to switching scenes.
If the same model is not trained "immediately cut to a new scene", it would be even better.
The point is that you can feed it any character without a character lora, and it'll do all the work for you to get them to that situation. If you look at the first frame of those example videos, you'll see what I mean.
If you don't want to use it in that way (again, I urge you to look at the first frame), then don't tell it to cut. If you aren't telling it to cut, then yeah, the training is getting in the way, but the training makes this into a single tool to do what used to be multiple steps.
[Latest Test Report and Recommended Data]
I spent an entire afternoon testing every possible pormpt or data. However, I overlooked one possible factor: the KSampler selection! While this is just speculation, after I modified the KSampler, the AI completely adhered to the cue word, and the scene didn't cut to the new scene.
Below are my data and cue words for reference.
High Noise A-S=0.35 CFG=1
Low Noise A-S=1 CFG=1
Total Step 6 Low=3
KSampler for both high and low noise was [euler / simple]
[prompt]
While maintaining their exact likeness,
The camera follows the same girl as she pulls up her skirt with both hands, exposing her vagina. She moves forward, bends down, and squats with her legs spread, straddling a man's crotch—his penis is erect. The girl grabs the penis with her hand and inserts it into her vagina. She then leans back and sits down while the penis remains inside her. The man grips her legs tightly with both hands. This scene is filmed from a low angle.
This Lora is fantastic! Thank you so much for sharing. This is already up there with my favourites. Best in class 🥇
Im still running 2.1 (for various reasons), so im only using the low versions of Loras, but they mostly work pretty well still.
It says there is a workflow attached, but I can't find it. Where is that exactly?
Just download the sample video and drop it into ComfyUI to get the workflow.
Did you train this lora to do the immediate cut, or is that built into Wan 2.2? I saw a lora to do this for Wan 2.1, so I'm thinking it's either trained into this one, or Wan just does it, but I'm assuming it's the former.
Wan 2.2 (and 2.1 too) were learned to do those cuts without loras but this is a big prompting issue. I've found a prompt for sharp transition which works perfectly with any other lora/scene but it took me 3 days :)
@eugskb532 Oh cool! So if this trigger isn't sufficient, that means this lora probably trained it in for a perfect 1 frame switch. What prompt did you find? Congratulations on finally figuring it out.
I found a better solution: using "Wan POV Miss ionary Insertion" lora with the "Wan POV Cowgirl Insertion" prompt. The action is executed perfectly, and the facial changes are minimal (even when the high and low noise values are both 1) (even when the character lowers their head and then raises it again).
Hi, are you using A or B?
@icefox21974
I always use A. But this post is to tell everyone not to use it. Instead, use the miss1onary Lora to execute this prompt. You can still complete the cowgirl.
@fctq9169 谢谢,
Probably the most character-consistent LoRA I've found, period. With the new lightning LoRAs, I recommend using 0.30 strength as they provide too much motion otherwise.
Incredible work.
Anyone happen to have any insight on how to set up the two high noise models for better likeness in the pinokio browser "WanGP v8.2 by DeepBeepMeep" way?
Works great but yeah the A model does lose the likeness quite a bit.
Damn I should really look into ComfyUI.
Either way, thanks for the fantastic Lora's you create!
Love the whole idea of it just kinda cutting into the action from no matter what kind of image.
My results are a lot worse than all the examples, using native/default workflow with suggested settings. Male anatomy is coming out quite mangled. At 720p it pretty much was not working at all, generating at 480 is significantly better but still not as expected.
It works good but i just have a question, sometimes it looks like she grows a penis of her own when she starts to sit down, is that something you can fix through prompting or lora strength?
im having that happen too
Godlike Lora.🙏
can you please help me setup this
Would love a cut to side fucking lora. Please consider!
This does not appear to work anymore.
KSamplerAdvanced
Allocation on device
This error means you ran out of memory on your GPU.
TIPS: If the workflow worked before you might have accidentally set the batch_size to a large number.
This is on a RTX 4090 on the first KSampler no matter which model I use and how much I lower the resolution, even such as 640.
As this error is on a fresh install, and I tested 7 of my other common workflows without this error and haven't seen this error prior to attempting this workflow it seems very unlikely to be something on my end.
After more testing on each of your workflows, including the missionary workflows, I found the issue is a bug with the cropping in your WF. For some reason it does not appear to crop or resize properly contrary to what it claims in the resolution values because the issue only occurs at larger resolution images like 4k and up and randomly (like 70% of the time). For any input images of around 1080x720 or smaller it never ever occurs.
Try the updated workflow in Downloads under "Training Data"
@The_Cook Yup, just tested and issue is solved with the new resize node. Thanks. Is much faster now since it is working with the correct size, too.
Testing with Lightx2v Distill Models + Wan2.1 Light Lora + Wan2.2 Fun Reward Lora yielded excellent results.
High noise model:
W22_POV_Cowgirl_Insertion_i2v_HN_v1A : 0.35
lightx2v_I2V_14B_480p_cfg_step_distill_rank256_bf16 : 1
Wan2.2-Fun-A14B-InP-high-noise-MPS : 0.5
Low noise model:
W22_POV_Cowgirl_Insertion_i2v_LN_v1 : 1
Wan2.2-Fun-A14B-InP-low-noise-HPS2.1 : 0.5
Looks good. They seem to be doing 14 steps, 4 on High Noise, 10 on Low Noise. Also try the High Noise B model at full strength, instead of the High Noise A model at 0.35
these types of lora (insertion/cut to action etc) are genius. please keep em coming.
caqn you pls help me setup this, I am new to comfyui
pls provide the setup guide and download essentials names for new commers
This works amazingly well. 1 issue i have though, is that it changes all eye colors to brown. Can anyone help me mitigate this?
Second issue is I can't create more than 120 frames on my setup. I'm trying to make a workflow for this, but most things haven't worked this far for this particular workflow.
Another winner, I love it. FYI I use model A at 0.2 weight and model B at 0.8 instead of your recommended method and it works like a charm, no issue with face consistency.
Hi! Would you mind to share your workflow please? I can't get this working without completely shifted faces :(
@tanoccb113 Unfortunately, I cannot. But I can give you some pointers. I run 3+4 steps and use the 1030 version of the lightning lora (it still works fine at 2+2 mind you). Nothing special otherwise. But I also generate at 720p using the Q8 model for maximum quality, you may get less satisfactory results with more compressed models?
Any tips on how to prompt it to do something briefly before cutting to insertion? A short transition beforehand would be great.
I've not seen it work for all models, but a few of them I see use + to cut parts of the video in half, so something like
"Dances,
+
Inserts penis,"
This is sorcery.
Does anyone know why the given workflow has loaddiffusionmodel>loraloader>loraloading for each of the three paths? When I try the default workflow with wan2.2>lightning>cowgirl, nothing happens. It runs fine when I remove the middle loraloader.
very nice lora,
I hope it will release in LTX too.
One of the rare loras that worked out of the gate.. VERY great! The results are also insane.
Details
Files
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
Mirrors
w22_i2v_POV-Cowgirl-Insertion-v1-B_H.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
POVCowgirlInsertion_HIGH_B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
POV_Cowgirl_Insertion_i2v_HIGH_v1B.safetensors
POV_Cow_v1_low.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_CowInsertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_HN_B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
女牛仔插入-H.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
POV Cowgirl Insertion HIGH - Wan2.2.safetensors
POV Cowgirl Insertion HIGH - Wan2.2.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors
POVCowgirlInsertion_HIGH_B.safetensors
W22_POV_Cowgirl_Insertion_i2v_HN_v1B.safetensors