I'm curious about the answer to this, but I've seen other prompts accomplish this, and it seems to work as a general concept... loosely? Wan will try very hard to create transitions, but those tens to weird shape fades rather than hard cuts. I do wonder if you could make an "edit" lora somehow...

@boobkake22 This is quite the coincidence, because @The_Cook , who made this very lora, made a very interesting and helpful lora that somehow got erased from Civitai. It was called "I2V Contextual_NSFW_Multiscene_Trigger_i2v_14b_v1". This was an experimental lora that was designed to create several sex scenes with one character.

Wan is very good at keeping a T2V character consistent on a given seed, as opposed to I2V. This resulted in the same character showing up in several scenes, and the user could then grab the first frame of each scene, and make several more videos with the same person. The prompt went like this:

"The same woman starts to take off her clothes, cut to the next scene where the same woman is completely naked and posing nude, the scene then cuts to the same woman opening her legs to reveal her genitals with her vagina exposed."

I was bummed out to see that the lora is gone now, because I felt like that could easily be a very fruitful path to go down with training loras, spitting out a complete character dataset with a single prompt/gen. I still think often about training one that does non-sex scenes, and more of them than what was trained into the Multiscene lora. I wanted to hit specific angles and situations to maximize training data usefulness.

@boobkake22 There was also a Wan 2.2 lora called "Transition From First Frame to Consistent Character_v2" by @yoursmilesweet . This only goes between two scenes, but it dresses the character differently, and seems like a step toward the lora I hope gets made, or that I may make myself. It's honestly a lesser attempt to @The_Cook 's Multiscene lora experiment, but I'm glad the idea is still floating around out there. It's just a shame these loras keep disappearing.

@Jellai Huh. Yeah, it's disappointing that LoRA isn't around anymore. I mostly wonder because Sora does it all the time, and generally unwanted. So it's clearly something that can be captured. (The character consistency there is sus, but.) If you peep my stuff, I've managed to get interesting results with a kind of "multi-cam" prompt for facial stuff. That works surprisingly well. The LoRA's heavily encourage them to be similar, which is a down side, but you could use something like regional conditioning to improve that if you really wanted the fuss.

@Jellai This one, to be exact: https://civitai.com/images/97897599

Though I have other examples with stuff on TV. It's fun to cheat extra videos into the gen! Prompt adherance there is middling at best, but it usually gives you something. I also do a thing with wall posters in some prompts. Again, quality is low, but it's fun to force it to make a lot of characters.

Also this one: https://civitai.com/images/101985733

That one will never be appreciated because landscape videos are so hard to see in the UI on these sites. But the girl filming herself is a really nice detail that I love so much.

Where to download the other lora?

@Jellai i have the I2V Contextual_NSFW_Multiscene_Trigger_i2v_14b_v1 Lora, send me a DM

@Lora_Addict Thank you for the offer! I have it too (I'm a chronic data hoarder).

Yeah, that's something I hate about the other insertion loras. We are so limited in the number of frames we can generate and we have to waste so many of them before the action actually starts.

nothing can fix that other than perhaps model retraining with blurred faces.

I've been testing it and sometimes it does it and sometimes it doesn't. Not sure what influences that.

@Slowmoe what helps is if you put high in range 0.3-0.4 and keep low on 1, also what helps is using only frontal full face portraits.

@erdelman73267 Cool. Will test the lower strength on high.

I can confirm the face changes on almost every workflow I run on ComfyUI. Probably needs to be retrained with no faces.

@meropewantsstls For me it's like 50%.

Even I'm facing the same issue, face changes a lot :(

Hehehe!!! I'm sure, what you want to do is like in my video:
https://sendvid.com/04i3lanz

It won't do the switch if in your input the woman is already in position or close to it.

@Slowmoe It seems not so sure. And I want to make a scene where the woman takes off her clothes and opens her legs. The man inserts into her.

@fctq9169 I probably made around 30 videos with this lora. I can tell you, the ones that do not have that sudden skip were because the woman was already in position.

In your case, you are trying to combine 2 sequences and WAN does not like that. You'll have to make 2 scenes and combine them.

@Slowmoe No. I have achieved the effect I want. But I don't know how to set the prompt words or photos more accurately.

even with florence i2v prompt helper?

for i2v put high on 0.25-0.45 range, test with 0.05 increases, it will preserve the original image and get creative when it comes to making the scene happen on the same background also it will prevent sudden switch from image action, the subject sill fall into the missionary position and it will be visible to you and animated. it should work for t2v as well, test it.

@Ghostguru I never added/used florence2 node. But i don't think it will work on an infinite length WF, where all scenes are generated in one row.

@erdelman73267 Thanks for the advice :). I will test it and notice you the result.

You can provide description for first cut before pipe and it will pause or do whatever else you want (change pose, adjust clothes). Most simple actions take only a second rarely two.

@iheartlumina Thanks I'll try it, your LORA is amazing and I'm only having a problem because I'm using the 4 step version instead of the 20 step workflow. Thanks again

can you share the workflow please

@illumizoldyck808524 I got it from this guy's discord that I joined after watching the YT Episode 49. ComfyUI Tutorial Series Ep 49: Master txt2video, img2video & video2video with Wan 2.1 VACE

@MMOFan thanks a lot for sharing it dude. you are a topp G

Anal too, please please please.

If your i2v workflow supplies first frame image without clip vision of it I strongly suggest you hook it up. I barely experience any face changes with this Lora when I use it.

@iheartlumina yup, just tried it. success rate went from 80% to 100%.

@KonoTheSavage1902 Thank you for the tips. :) A quick question if you don't mind:

When you wrote "recomended weights for loras High: 0.25-0.4" did you mean this for the loras on this page, the lightning loras or both of them? I noticed in my own generations that if I lower the strength for the high lora from this page too much the likeness falls apart as well as the motion. So I guess it's a balance.

Since i'm using a cfg of 1 i'm also using the 'WanVideoNAG'. Not sure if you do too.

@iheartlumina Do you have a preference for a which clip to use for the "Load CLIP Vision" node? 

@Rumaben I use KJ nodes and his clip encoder seems to prefer open-clip-xlm-roberta-large-vit-huge-14_visual_fp16.safetensors

@iheartlumina Thank you i'll try it a little bit later. Right now i'm trying to lower the missionary high lora strength to improve likeness but everything below around 0.80 almost stops motion completely. do you have both your lightning loras at 1.0?

@Rumaben I have both at 1. You can see my workflow on my profile in recent scenario post. All steps use the same basic workflow, just put in this lora and replace prompt with whatever suggested here. I do not start with empty prompt and pipe because I do not like the flicker or randomness. I have a first prompt with basic short action, i.e. cover mouth in surprise, smile, etc. The better you describe your subject, the better result will be. As an added bonus try with multiple subjects :)

@iheartlumina Thank you. :) I'll try your workflow. Maybe Florence can be used to describe the start image. like in this workflow: https://civitai.com/models/1480528/wan-endless-facials

@Rumaben Just for this page high noise lora everything else on 1.

@Rumaben My workflow https://civitai.com/articles/19884

@KonoTheSavage1902 Got it! :) My earlier issues is also fixed now. I think the lora got stuck or something. After restarting comfyui I can now lower the lora strength without wierd things happening. Thank you for your help and have a great weekend.

@KonoTheSavage1902 Cool Thanks. 

You guys keep talking about the clip_vision_output hookup for WAN 2.2, but even Kijai says the model doesn't use image attention layers anymore. You can easily test this by locking the seed and trying with a clip_vision input of your subject, and then an input of something random like kaleidoscope images. You'll get the same output either way.

@DystopianDad  That is interesting, maybe I should follow KJ a bit more often. So I just tested your assertion and indeed it does not make a difference. Thanks to you I can clean up my wf a bit.

@DystopianDad It doesn't work for like 70% man. Face keeps changing.

@KonoTheSavage1902 Did you try cropping the image so it's only the front of the face/head visible. That helps a little I think. Right now I'm using 0.25 for this missionary lora (high) and 0.6 for the lightning (high). Unipc/simple for now and only 2 ksamplers since I had some problems with my output being much too bright and lacking contrast with 3 ksamplers and the beta57 scheduler yesterday. Need to do some more testing. :)

Trying out what this guy wrote I think or close to it:

https://huggingface.co/lightx2v/Wan2.2-Lightning/discussions/20#6897529022ebf7a1e53df7c7

It's pretty much identical to how 'The_Cook' suggested it in his 'Posing Nude' lora description. This same link also discuss 'Skimmed CFG' which could be a replacement for the NAG node.

@Rumaben yup I tried. I basically used yolo8 to get cropped image with only face and it works but what's the point if you can't get the 1:1 body. xd

@Rumaben From all the testing, I concluded that if you use potrait image from head to stomach/waist you get really good results. I'm trying inpaint now to generate same face with exact body proportions then feed the output to image to video but I'm new this AI stuff so trying to learn :)

@Rumaben Also, if you are facing issue wth lighting then you can try this tool "wan22-brkn-prompt-helper" to generate prompts. It will give you options to select what kind of lighting you want, doesn't matter if what brightness/contract your image has. Tried and tested.

@KonoTheSavage1902 Haha. :D You're right about that. :) All this changing of nodes and settings is slowly driving me mad. :/ Loras always messes up something, never is perfect. I wish that someday we'll get the perfect finetuned model that just works but maybe i'm just a dreamer. lol. :D

@KonoTheSavage1902 Thanks, I tried the 'Color Match' node but it just messed up the colors completely.

@KonoTheSavage1902 I tried with 3 ksamplers again without the beta57 sheduler. No overexposure this time and much better motion but the likeness of the subject is much worse. Save me Wan 2.5 lol. :D

@Rumaben 2 kSamplers takes around 5 minutes minimun on my PC with 6 steps. If I add 3rd one I'll go crazy lol.

@KonoTheSavage1902 I know, I'm just doing some tests. :D Higher than cfg 1 is also a lot slower. I used 1 step at the first ksampler and it worked, so then I did 4 but only got white noise. Even though I've played with ai generation for almost 2 years I still have a lot to learn it seems. :)

@Rumaben Yeah, learning. Brutal reality, It's never enough. I'm loosing my mind. :)

BTW, can you share your workflow ? I really wanna try 3 KSamplers xd.

@Rumaben ColorMatch node can definitely be helpful. Make sure you lower the "strength" to ~0.10

@iheartlumina I wish it weren't the case. I was also looking for ways to make the subject more consistent, and hoped I had found the silver bullet with your comment on clip_vision; That's how I ended up finding out it's not longer present in 2.2. 

I appreciate your sharing your workflow, though. Getting some really interesting scenes with the high lora at 0.3 like you suggested!

So I gave up on keeping 100% facial likeness with a scene change using the loras from this page. I tested around 10 missionary loras two of them had the best likeness with scene changes:

T2V_PREVIEW_WAN_21 (Best likeness but very bad looking genetalia and no real motion)

Wan2.2 I2V Highnoise (likeness almost as good but weak/thin genetalia, okay motion)

So I ended up using the latter but my problem is no matter what penis lora I use the they have no effect until I'm using a strength of well above 1.0 and then the output quality degrades and likeness goes away. If I try to lower the missionary lora I then begin loosing to much motion instead. The main missionary lora seem very stubborn and is working against changing.

Maybe some other ppl can find a solution, but for now i'm stuck. :/ :)

@KonoTheSavage1902 There's a 3 ksampler workflow embedded into this video:

https://civitai.com/images/99156583

Only difference in mine is that I use the ComfyUI-MultiGPU nodes ('UnetLoaderGGUFAdvancedDisTorchMultiGPU' & 'CLIPLoaderGGUFDisTorchMultiGPU') for my main models and clip. I use no crop or image resize node and simply just feed the load image into the 'WanImageToVideo' node. Around 8 steps total gives the best quality for me (euler/simple). Shift '5'. only other nodes I use are the 'TorchCompileModelWanVideoV2', 'PatchRadialAttn' (dense_timestep '2'), 'Model Patch Torch Settings' and the 'Rife Tensorrt' node for faster frame interpolation (<-Rife49).

I usually generate at 512x768 and 97 frames. 16 fps interpolated to 32. Had a bit of problems with the fancy RES4LYF shedulers (beta57 & bong_tangent) so now i'm using euler/simple. For the wrapper workflows I like 'dpm++_sde'. :)

My card is pretty budget (4060 ti 16gb). I offload the clip onto my old 1660 6gb and main models to my 64gb of system ram with the multigpu and ComfyUI-GGUF repo's installed using Q6 gguf's for both the main models and the clip. Radial Attention may need sparse attention, at least with the wrapper workflows. Kijai has the wheel for it on his PrecompiledWheels page. I use 'sageattn_qk_int8_pv_fp8_cuda++' in the 'Patch Sage Attention KJ' node.

@Rumaben the only real solution is to recreate the LoRA removing the facial data. There's no point in looking for solutions for a problem caused by the LoRA itself.

@meropewantsstls I agree. But, what if I want it that way, because this improved my understanding of LoRAs, ComfyUI, creating my own custom nodes, etc. Basically I'm on this platform to learn about generative AI and ofcourse gooning xd.

@KonoTheSavage1902 great attitude. Pretty soon you'll be training your own LoRAs :)

@meropewantsstls I just did. :)

@meropewantsstls That's true actually. Never thought of it that way. I guess i'm just so used to forcing a square peg in a round hole with comfyui that I never thought about just making my own. I guess it's also a little ungrateful of me to be using other people's loras and then just criticize them afterward. :)

@Rumaben constructive criticism is encouraged on this platform.

@Rumaben I have same card lol.

@KonoTheSavage1902 My brother in goon!! Hihi :D Yeah no way i'm paying those sky-high prices for a 5090's or the 96 gb rtx 6000. :O Maybe I'll get one of the 50xx super cards next year if Wan 2.5 get it's weights released and it becomes too much for my card.

@Rumaben Bruh I can only dream about 5000 series gpus. I had to go for second hand market for current gpu. :)

@KonoTheSavage1902 Well it's not all bad having a potato pc. :) You learn a lot about squeezing out every last drop of performance and utilizing all you've got. You also learn about patience.

@Rumaben wise words from a wise man.

Hey, mate! Would you be so kind to share the I2V workflow utilising the clip_vision?

Hey, mate! Would you be so kind to share the I2V workflow utilising the clip_vision?

@Undoriel @xjtomb548  In your existing workflow.

Step 1: Search for "Load CLIP Vision" in nodes and select it.
Step 2: Search for "CLIP Vision Encode" in nodes and select it.
Step 3: In "Load CLIP Vision" select your model. I used "CLIP-ViT-H-14-laion2B-s32B-b79K.safetensors".
Step 4: Connect the dots. From "Load CLIP Vision" node connect "CLIP_VISION" to "clip_vision" dot in "CLIP Vision Encode".
Step 5: Same thing for image. from your "Load image" node to "CLIP Vision Encode" node.
Step 6: Connect "CLIP_VISION_OUTPUT" from "CLIP Vision Encode" node to "clip_vision_output" in "WanImageToVideo" node.

Done.

You can use any of the step workflows in scenario on my profile. They are the same wf except prompt/lora

There's some posts on reddit that explains why clip_vision is not necessary anymore, if you're interested. Too technical for me.

From my testing, what affects the face is the input picture. It is better to use a picture that is as close to the face as possible and the face looking straight at the camera.

@Slowmoe TY. It was mentioned in another comment and I confirmed through testing that clip_vision does not make a difference. In my experience with the lora the composition does not really matter much, my gens experienced changes very rarely and sporadically. I am now testing gens with two subjects and that is where consistency breaks a lot. If I transfer learnings from stabilizing these gens I would say describing/mentioning things that are important will affect quality and construct of the gen.

@iheartlumina Oooh, I didn't even think of having 2 subjects.

clipvision is integrated in WanImageToVideo Node (hence there is clipvision output on that node) if you are using it in your workflow u already have the clipvision.

I would suggest two video generations - one with the initial scene without this lora, then use the last frame of that scene as the source image for the generation with this lora. Join the videos afterwards.

@alienuser thanks!

Only works when using the provided Workflow for me, but 8 steps is a bit boring, but this lora is amazing!

@zoom83 i cant see a workflow, could you point me in the direction please. i think i might be going comfy blind

@braddas77655 Just paste the the example video of the model

Try the "lazy" workflows found on Civitai made by Chriqro. They have a face retention element added. It might help.

@meropewantsstls hm, seems to me this "face retention" element (ipadapter wan, which isn't really for wan) doesn't works. -_-

@forfreelsd368 did you ever find a good workaround to help retain the same face? Face varies with LORAs and other settings, but always want to keep it as close as possible.

@heyheyhoho nope, I didn't find a solution. Main rules for face consistency are the same as before, nothing new:
- no jumpcuts (some loras, like iGoons i2v, which are nuked now, did a great consistency);
- face must be in the frame all the time (for longer video, face must be with opened eyes at the end of first clip);

- less loras (at least low noise weights) - better;

- prompts like "same person with same face, same scene" may a bit help;

- filtrating loras for face changing (I try to not use such);

- using old good "reactor" nodes at the end of wf for faceswap, and run again through another just low noised wf with low denoise to blend in face and remove reactor face bugs.

Maybe something I forgot. Main stuff - is choosing loras with minimum face changing, or not using loras at all if it's possible.

Heard that Claude ai can do now some fancy stuff with coding and adaptation (like cuda for amd, I'd tried it myself, but locked for my country), so I think one day we got working ipadapter for wan2.2, or something even better.

And another one thing I'm testing now. I use this high noise at .4 https://civitai.com/models/2316517/wan-multiscene-photoshoot-softcore-edition with idea from, again, wf for this lora (like first frame - concatenated picture of multiple views) and add to it other loras (like iGoon's which are nuked on civi), even with jumpcuts works more consistent. Usually enough just concat starting picture with auto crop faces node, but it may change body or breast size, so sometimes I add nude body picture - a bit helps.

agreed

I still love how well it transitions into position without a cut. But yeah the face is superbly consistent. I don't know how a rare few Loras can do this while others can't?

can you please share your workflow. It's not maintainig the face for me.

use wan 2.2 i2v workflow and add lora intro it

i am not getting consistent face. can you tell me how to get consistent face if you are getting it ?

I often hit the same issue. I think --just like when you write about how a girl's facial features look > hard to get from-behind photo's-- If you describe the penis, it wants to show the penis.
How to propperly pull it off though, I also struggle still.

The main issue here is that 10 inch porn penises can't realistically fit inside all but the the most special princesses.
There are two words that will help deepen penetration and increase realism and immersion across every Missionary POV LoRA in the meantime: "small penis"

maybe they say it helps with a cut

Starting with a high quality photo and rendering at a very high resolution are pretty much necessary to prevent identity drift. Describing the face in the prompt can also help. The closer your starting image is to the scene you want to cut to help preserve identity, too.

@petercaldwell1505833 will try it out. thank you for helping out :)