about SCS(Scene Change System)
This system utilizes the Wan2.2 I2V workflow in conjunction with Scene Change LoRA and T2V LoRA, enabling a seamless transition to the T2V LoRA scene while preserving the person from the input image.
Scene transitions using this system offer exceptional character consistency. You no longer need to train a character LoRA every time; a single portrait image is all it takes.
Files required for SCS and their roles
1. Scene Change LoRA(scene_change_nsfw_v2.0_high_noise.safetensors)
This LoRA serves to seamlessly guide the input image into a T2V scene. It does not use Low Noise.
2. Target T2V LoRA
This is a T2V LoRA designed for the target scene. The "High Noise" setting strongly anchors the scene, while the "Low Noise" setting provides the foundation for rendering the genitals. I recommend using the official T2V created by MQ Lab.
3. DR34ML4Y_I2V_14B_LOW_v2
This renders genitals. High Noise is not used.
https://civarchive.com/models/1811313/dr34ml4y-all-in-one-nsfw-wanltx2?modelVersionId=2553271
Strength setting
From the perspective of character consistency, the lowest strength at which each LoRA can effectively fulfill its role is preferable.
scene_change_nsfw_v2.0_high.safetensors 0.70
target_t2v_high.safetensors 0.70
target_t2v_low.safetensors 0.60
DR34ML4Y_I2V_14B_LOW_V2.safetensors 1.00
Prompting
Please refer to the page for each T2V LoRA.
Input image
I recommend square-resolution, portrait images.
Others
Using the 4-step LoRA and generating with "High Noise 4-step" and "Low Noise 4-step" yields the highest quality.
Please download both the High Noise and Low Noise versions of the target T2V LoRA.
You can get the latest development updates from my Discord server.
I am asking for donations from those who wish to support my work on Fanvue. Please consider supporting me so that I can continue creating.
Description
Please download both the High Noise and Low Noise versions.
The prompts are listed in the options file.
FAQ
Comments (44)
Any trigger word for scene change v2 or is it still the same as prev version??
I'm just trying now, but according to the description, use the text file on each lora for prompting
@SillyNovice I've tested it, and it looks like we don't need the trigger word anymore. I tried it with igoon's loras, removed the "jumpcut" from the prompt, and this v2 lora handled the real time smooth transition automatically no black jumpcut. It's also much better than the previous version, a lot better.
wait, where are the text files?
Great! May I ask if there is any relevant workflow? Thank you very much
The video includes workflow data
No low noise?
DR34ML4Y_I2V_14B_LOW_v2
@Saxlive we already know about this one. The workflows reference a low noise version of each lora that isn't there.
@ProvenFlawless @cnovice They're there. Observe the "2 variants available" in the download menu - it opens up.
@firemanbrakeneck where?
@arbsec Search the words "2 variants available" on the page (any besides main scs). Click the little arrow directly below it, next to the bf16 or whatever file it's showing.
@firemanbrakeneck thank you!! found it!
@firemanbrakeneck Whoa, learned something today... now I worry how many models I missed due to noticing the variants dropdown -_- , thank you!
@Neoxena Probably not that many if at all.
It's only been around for a couple months (see log below), and I don't think it's been used for wan HL before (though it's a good use case, people are fanatical about the crummy version separation they've gotten accustomed to over the years).
Mostly, you'll see that for training data, workflows and ckpt quants, which was already the case before. Furthermore, the api returns all file results (except rarely when it glitches out and fails to return versions at all), so the various hf archives are likely to contain anything that's been hidden.
Incredible model thank you so much!!
Amazing, thank you. will test them tomorrow.
I noticed, there are no anal, I hope you will add some anal soon.
Any guidance on optimal resolutions for portrait or landscape?
today I've just tested the standing... but it's great. Not 100% perfect all the times, but at this point, none of us cares if we need to run it twice or even three times to get the desired results. I'll test the others during the coming days. In any case, man, congratulations for your job. I can see that it is usefull, beautiful and beyond my skills. The only thing I'm affraid off is that implementing this may be a little bit too complex or even confusing for newbies. Don't get me wrong, the information is there and you shared everything that's relevant, but I suggest new people to better start with other stuff. In any case, congratulations, I hope you get the credits you deserve
MQ_Lab woke up and chose violence!
Thank you :)
I'm quite surprised those are still up tbh 😂
@GlowingGuardianGirl What does that mean? Please let me know if you don't mind.
@MQ_Lab With all the I2V loras being removed by the civitai staff I mean, because of the possible misuse against celebs. And they're creators getting banned, like playtime_ai who was banned multiple times already and keeps coming back 😂. I'm surprised they're holding up, because it's mainly against the TOS. Keep up, cheers 🙌
@GlowingGuardianGirl
Ah, I see. Well, it might get taken down eventually. That said, I feel the I2V technology itself is the real issue when it comes to repurposing it for use with celebrities...
@MQ_Lab Yeah I guess at some point everything i2v might be taken down 😂
@GlowingGuardianGirl I'm very sad that playtime was banned two times. Because he was only banned two times, and now he stopped to publish on civitai, he only publish on his discord and huggingface repo.
But, to be honest, after the first time he was banned, because of his anorexic lora, this was the only reason of his ban, the anorexic lora.
When he came back the second time, he was very arrogant against civitai staff. And this explain, why he was banned the second time.
If you remember well, the first lora he published after he came back, was the middle finger lora....
The civitai staff is not stupid, and perfectly understood the message behind his act.
After being banned—regardless of the platform—when you are given a second chance, you need to keep a low profile and stay discreet.
Playtime failed to do this; he behaved in the exact opposite way, which explains his second ban.
great lora, thank you so much for making this series, i knew there are other blink lora , but the fact that you’ve built this entire system—complete with comprehensive usage instructions and setup recommendations—is truly impressive. You’ve put a tremendous amount of effort into this; the community appreciates it and salutes you.
hope some more sexual positions can be added. such as 69, prone bone, full nelson, etc.
Looks much more natural than V1, thanks!
good work! can u make quick black screen transitions like Scene Change 1.0? While "SCS for NSFW v2" is great,but the transition is too slow,the best part ends just as it's getting started.
You are quite right; the sex scenes inevitably end up being short.
The main reason is that character consistency cannot be maintained using methods known as "scene jumps" or "scene cuts." Fading to black operates on the same principle. This is a limitation of Wan; character degradation invariably occurs at the moment of the jump. I consider maintaining character consistency to be the top priority. Here’s hoping a new version of Wan is released in the future..
@MQ_Lab unfortunatly, this is clear now, Alibaba stopped releasing Wan open source model since Wan 2.5.
The futur of open source video generation, is LTX now. That why all creators should focus on that model :p
And with LTX Director, LTX2.3 is just amazing and easy to use, hehehe!!! I hope you will start creating LoRas for LTX2.3 as amazing as your Wan22
@Pat3dx Thanks for the comment. That explains why no new version of Wan has been released. I’ve been keeping an eye on LTX, but the issue is that it doesn't maintain character consistency as well as Wan does.
@MQ_Lab I totally agree with you. Base LTX2.3 model consistency is horrible compared to Wan22. but with lora like IC ingredient, 99% of the consistency issue is solved. You can also use the 10s Node, i don't use any character consistency yet in my futanari NSFW workflow because i don't need it yet. And i only use LTX Director custom node in my LTX2.3 WF.
I switched from Wan22 to LTX2.3 to create my NSFW futanari video for my Patreon/Subscribestars/Affect3D subs and buyers.
Of course, some sex position still required to use my Wan22 WF, as it's not possible yet with LTX2.3. But i could say 80% can be done with LTX2.3. And the really good thing, is that, i'm not limited to 5-6s anymore, now i make 10-50s video in one shot, and my subs and buyers are delighted, hehehe!!!
Quick note: for the best consistency, as LTX2.3 was trained with 720p resolution dataset. You need at least to put 1280x720p for output video, or the consistency will be horrible. I tried 560p and 480p, consistency was horrible, and all fixed when i switched to 720p and 1080p
Combine it with pornmaster bukkake lora high noise, it is really good at keeping face consistency when using jump cuts and works with most lora's.
@CraMaga we were talking about LTX, not Wan22 :p
Wan22 have no problem for consistency, as long as the character face stay in the frame, of course, if you plan to make video where your character turn back camera or go out of frame for a few seconds, you need character lora.
@MQ_Lab hehehe!!! I made my CineFlow app for comfyui yesterday, no need to open comfyui anymore :p
See my X post:
https://x.com/AIInfluencerHQ/status/2076387191106773456
And here the example video made with my CineFlow app uploaded on sendvid:
https://sendvid.com/lxqwf0pa
That's why i love LTX2.3 and LTX Director, and now my own CineFlow app that take advantadge of LTX director v2, but in a more useful, and powerful app, as CineFlow is powered by AI, hehehe!!!
And as you can see, the character consistency is absolutly fantastic. Even after the closeup of the masturbation part, when the camera move back to my Camgirl. I don't used extend or any LTX2.3 character lora for the video, as shown on the screenshot of the CineFlow app UI. The 30s video was genrated in one shot. I also made few 1mn video for my AI Camgirl.
The only image used, is the first segment image, that has a duration of 2s, at the start of the video. All other segments are pure T2V.
And the crazy part of this 30s video, is that i don't used masturbation lora, LOL!!!
The AI generated video sfcript was strong enough to describe all the part of the T2V perfectly, that all the T2V parts, was just amazingly well done, without any consistency issues.
And of course, if the video is so well done, it's due to the generated script by my AI Director. This is the power of the CineFlow app.
Of course, it's not perfect, issue with the feet when the camera zoom in to pussy, and double navel.
But, eh, AI is not perfect, even Seedance 2, HappyHorse 1.1, and Grok imagine 1.5 has issues.
Bravo, Loving it so far and I got to say I really appreciate how you laid everything out. You put a lot of work into this I can see and I hope people will take notes on how it's done. Thank you
I think we're missing a few low pass loras, at least missionary b
nah, they are all there.
Anyone have any tips on how to get the guy to put it in her backdoor on reverse cowgirl? I tried several different loras like edible an*s which usually works and nothing will work, oh I prompted in an*s instead of vag and other terms nothing took. Is the data set just not trained to allow it?
These are awesome, except missionary b is still very rough in terms of quality.
This NEEDS a scissoring version, looking forward to updates
sorry if I am being stupid here but where to i get the target lora
The "target LoRA" is any T2V LoRA you want to use
For anyone struggling to find the prompt data, just click on the individual loras here and then scroll down to "optional files" they've provided the text file with the prompt information for each individual lora! Thanks again the the creator, you're a legend