Forty 5-second samples were used to train the software for 20 cycles on raw video, and then converted to image-based video for 11 cycles. 512*512
Currently, only oral sex is possible; other sexual intercourse capabilities will be added to this model later.
If you find the time-stop effect too subtle, feel free to increase the High-Noise model's strength. Pushing it to 1.5+ works perfectly fine and won't break the image. My usage habits are: High Noise: 1.1, Low Noise: 1.7, Penis: 0.8，CFG:1
This is the checkpoint I use, and I highly recommend it; it maintains quality very well.
https://civarchive.com/models/2086218/wan-22-10-steps-t2v-and-i2v-fp8-gguf-q80-q4km-models?modelVersionId=2360434
The descriptions are labeled using Chinese, so if you are from a Chinese-speaking world, it will be more effective to use Chinese directly.
Time freeze;
Woman frozen;
Oral sex;
Hands open the woman's mouth;
Remain still;
The mouth bulges and bounces as the penis is inserted;
High noise: 0.8 - 1.5
Low noise: 1.5 - 2
Penis Helper: 0.5 - 1 (Simply add it to the low-noise model)
If the effect is not good, try a different seed. Test effect satisfaction: 50% The test images were randomly taken from the homepage. If they infringe on your rights, I apologize and ask you to contact me to delete them.
Description
Perhaps someone needs this; the cue words can be extracted from my training set.This version does not include NSFW training.You can control the POV hands to perform most actions, such as taking off her shoes (though not every possible movement is supported)
FAQ
Comments (43)
BRILLIANT, trying it now, thanks for making it!
UPDATE: It's great. Oh the fun to be had.
What?
I'm not from the English-speaking world, so the translation might be inaccurate. This is a refined low-noise penis lora that can effectively improve penis shape. Please also add low noise.
@pongjoo IMO you could just rename that lora to Penis Helper, and add description "it improves penis appearance."
@erdelman73267 Oh. Thank you for your suggestion.
Love this concept. What CFG do you suggest?
The example video is available for download; it includes the workflow and checkpoint I used. CFG:1 ，Checkpointing should avoid using hybrid models with dynamic enhancements, as this would break the concept of time freeze.
Great start. Looking forward to the expanded version.
Oh damn I really like this concept! I think it's the first time I've seen this type on as video lora
Tried a wide variety of settings that match your settings, I kept getting detachable penis, or a penis detached exiting her mouth before it could be reinserted...Is this lora better with a penis that already exists in the frame?
Perhaps you could try adding the following to the prompt: "A man's lower body enters the frame in the scene."
优秀的创意与制作
Could this be used to simply keep objects or people still without any additional nsfw thing? Wan tends to make things move even if you tell it to keep them still, I have a use case for keeping subjects still
I actually trained a time-freeze model without NSWF, which you mentioned, using 44 five-second clips. Here's the link. The documentation for the model training set is also there.
https://huggingface.co/PONGjooo/timestop
@pongjoo oh that's awesome, thank you. Is there a trigger word and low noise model? I guess it's not that necessary given that it's motion-based
@dailydoseofaiart I've uploaded the training documents in a package above; you can download them to view the keywords in the documents I've labeled.
@pongjoo oh yeah I saw the prompts but didn't think to take a trigger out of that... dumb me, thank you!
typically the kind of stuff we need. thanks
please include metadata in the clips so we know what youve used
My sample video contains a workflow, which you can download and view.
@pongjoo its ok it works perfectly.. amazing job mate, looking forward to your future work!
@pongjoo I feel blind, sorry - where is the workflow? I checked the training data too, but not finding this.
@kurokuro Drag my sample video into ComfyUI; it has the workflow embedded.
@pongjoo Whoa, I didn't know that was a thing. Makes sense, I always forget about the save_metadata flag for video helper. Thanks.
great lora, thank you, works in conjunction with other loras
Yes, I actually removed the blocks in the model that were prone to conflicting with other models, keeping only the core modules. Therefore, the low-noise model is significantly smaller than the high-noise model.
This is a really neat concept. And whist not perfect, when it does work its awesome. Kudos!
Awesome. Couldn't download this fast enough.
uh... for a friend.
or... for science.
or something
These are quite fun to play around with.
I already luv it and I didn't even have the chance to try it! The concept is such amazing for epic storytelling moments, thank you very much! 🫶
I'd love a 'deep asleep' variant of this. WAN always makes people wiggle and change poses.
A model similar to what you're describing was recently taken down from the site. You could try searching for it. civarchive
@pongjoo The kissing / licking one? I saw it, but it was pretty specific to just that use-case. Anyway, thanks for the suggestion.
我猜您是用sd2产生的素材进行的训练，感谢。另外，做这种时间静止的，可不可以用vam或者实体娃娃来取样呢？
I believe incorporating some doll-like materials will better enhance the concept of 'stillness'. It should help achieve a more static, frozen look in terms of micro-expressions.
thank you for the lora, it's great work!
is there any sure way of preventing orbiting/horizontal camera movement when zooming in? I've tinkered with both chinese (along with translating it) and english prompts, but the camera almost always orbits around the character, when I simply want it getting closer
Please use the camera to zoom in. The corresponding Chinese text is ( 镜头推进).
靠! 中文耶 竟然可以用母語 真是太爽啦!
Can I ask — for your dataset, do you use Qwen to edit and structure it into first / mid / last frames, then use WAN to generate it as a LoRA training dataset?
I didn't manually split it into head, middle, and tail frames. Instead, I used Qwen3-VL-4B for automated batch captioning. I simply prompted the model with a 'time freeze' theme, and it generated the tags accordingly. My workflow then saves the videos and their corresponding text files into the same folder, which I then upload directly for training
The video materials were collected and curated from the web
is it possible to make a lora for LTX 2.3?
It liking Chinese prompts is an understatement, I can barely get it to work completely in other languages.
More general SFW use and being able to better understand other languages would be great if there is ever a new version.
Details
Files
I2V-H-timestop.safetensors
Mirrors
6-timestop.safetensors
I2V-H-timestop.safetensors
I2V-H-timestop.safetensors
I2V-H-timestop.safetensors
I2V-H-timestop.safetensors
I2V-H-timestop.safetensors
wan2.2-i2v-high-timefreeze.safetensors
I2V-H-timestop.safetensors
I2V-H-SFWtimestop.safetensors
I2V-H-timestop.safetensors
I2V-H-timestop.safetensors
I2V-H-timestop.safetensors