The less than ideal prompt I used to generate the video examples
positive:
Video of sexual intercourse involving penetration of the pussy. In a detailed, high-resolution CGI style, a young woman is filled with pleasure as she feels the unexpected penetration of a man's body pressing against her. Naked and toned as the man's penis smoothly enters her wide-spread legs. Her mouth is open in pleasure as the man thrusts in and out of her body. Her body rises and falls in harmony with the man's penetration, her breasts bounce gently, her nipples tense. Live2D. Emotionless. The face does not change its expression throughout the video.
negative:
The video is not of a high quality, it has a low resolution. Watermark present in each frame. Strange motion trajectory. Static image. Static video. Camera movement. Blur. Change of frame.
Description
FAQ
Comments (33)
Sweet!
Such a jump in quality with the extra steps! Amazing! Thank you for being the trailblazer for these video finetunes! The only recommendation I have for you is to share the seeds you used for the images, because it'll play a big part in the overall motion!
Okay, this has my interest, I've been neglecting COGVIDEO for a while so I don't have idea how to start with it, do you have a guide or maybe the exact workflow you've used/made?
Quite a few people ask about this... Probably it's time to make a separate instruction with an attached workflow.
In the meantime, you can look at the node for ComfyUI https://github.com/kijai/ComfyUI-CogVideoXWrapper
And yes, for generation you need at least ~10 GB VRAM
LoRa files should be placed in the path ComfyUI\models\CogVideo\loras
if you want to try something with more control then try my cog workflow, i've been playing arounf with cog a lot and made a workflow that include anything you may need and more + there is already a lora section. i was waiting someone to release loras and here we go.
@LatentDream Awesome job on that workflow. Only issue is when i try to use this lora, I get this error: DownloadAndLoadCogVideoModel - pop from empty list
I currently have this lora within models/CogVideo/loras
@Tsterbta do you loaded the lora in the cog lora node?
If I use in forge,can someone tell how to use this lora in forge or A1111?Thanks a lot!!
From what I asked bing and chatgpt it's not compatible with automatic1111, you need to use comfyui.
Im in the same boat, but Im going to try to learn comfyui as I think it'll be worth it, if not then at least i've had some experience with comfy.
@skullx198x980 THKS FOR REPLY,I kinda want learn comfy,but my disk capacity is full.BTW,is this lora can work with illsutrious model?
@zczcg I got it working and it's pretty neat, tho it was a huge headache to set it all up but maybe that's just me and the tuts I watched.
I'd say be ready to have around 50gb+ free in case you download wrong models and such, maybe a minimum of 30gb.
Some knowledge carries tho some was new like clip models.
I dont think so, I think illsutrious is a branch of sdxl, I had to have sd3 enabled to use this lora but didn't try the other options.
@skullx198x980 it`s pity,Cuz the illustrious model can make the realistic nsfw model I tried,may someday some extension can support that.
@skullx198x980 I fully change to comfyui and it works!Thanks for sharing this lora
why not training on the new 2B or the Xfun version? those are more vesatile/fast!
anyway..
finally someone noticed cog loras worth the effort at least. thanks for sharing.
may i ask you if it was difficult to make this loras? i tryed multiple times to get into training and got depression
Here are a couple of stages and approximately how much time they took.
1. (10 hours) Dataset of 800~900 video fragments from rule34 sites (More precisely, I took about 100~200 videos and cut them into 6-second videos with a frame rate of 8 and then manually sorted out the normal options)
2. (I think about 2~3 hours of computer work) Automatic prompting using the first frame of the video (There are no VL models yet that would correctly describe nsfw videos) and joy-caption. Then, using LLM, I brought all the prompts to one template.
3. I trained LoRa for about 10~12 hours for 5000 steps on H100
I used the default settings https://github.com/a-r-r-o-w/cogvideox-factory (I just removed the test videos)
In general, the preparation took about 12 hours and the training itself took about the same
I also want to say that I myself would like to try to train LoRa on other versions of CogVideo
Now I'm thinking about 5B-1.5-I2V or xfun
I completely forgot about 2B that someone made it.) I'll need to study the training of this model since it should be more accessible than 5B
hi im new to the ai thing but it doesnt work for me, i put both the whole pasitve thing and the tregger word but it only give me normal pic not a video
If you use LoRa correctly for CogVideoX5B-I2V, it is better to first try to write positive and negative propmt from the description, because mostly the model was trained on such cues (you can remove the last 2 lines about live2d and emotions).
And yes it is better to use 720x480 image with not very intricate pose (for example pov cowgirl).
@Cixiao thx for answering, wait you mean I should bring a pic then put all the positive and negative propmt on it?
Can't I generate a new pic with this lora immediately?
Like let's say I put a loran for yor forger then put this lora, is it gonna work?
@Toya_00 You need to generate an image using another neural network or take a photo from the Internet
Then generate an animation using the CogVideoX-5b-I2V neural network and my LoRa
You don't have to describe anything in particular if you want and use the standard hint that I wrote in the description.
By the way, I hope that you use ComfyUI and https://github.com/kijai/ComfyUI-CogVideoXWrapper to run this LoRa?
Well, you also need at least 8~10 GB of VRAM on the video card to run it
I've been trying to gen at 20 steps but I'm not having very good results, it's very hit on miss. I had a lot more consistency with the paizuri and BJ ones.
Maybe the 5000 step model is too overtrained for you... Try 2500
I also noticed that it works best on images generated with pov and cowgirl tags
Couple questions! Does this work with the new 1.5 5B CogVideoX version, or only 1.0?
Also to use this in ComfyUI, do I add a CogVideo Lora Select node and plug it into the (Down)load CogVideo model node?
I don't know if it will work with 1.5.
You can check it yourself.
@Cixiao So the answer is it works, but only sort of. It tends to cause some really drastic and broken motion on 1.5.
@Diodamyr I saw that someone is working on setting up the LoRa training process for 1.5
Maybe I'll try it soon
@Cixiao Lemme know if you get that working. I've got some video concepts that I'd like to have a crack at training as well, lol :)
i musst say 1.5 is much better on my tests. Hope we will see some awesome finetunes/loras :)
So, nobody can directly answer the question? Is it "add a CogVideo Lora Select node and plug it into the (Down)load CogVideo model node"?
@nojojos2234 Yes/ but I had to manually update bitsandbytes to the latest version to get it to work
Great work! But I have a question. I find it is difficult to ask man to move his thing up and down. What kind of prompt I should add in? I use Qwen 2.5 3B to make prompt itself instead of other LLMs because of national limit policy. I use your prompt example in Qwen but result still doesn't have motion of male's thing or little motion has. Anything suggestion? I also try 2500 cycles file but result is same.
Hello, one question, what would be the ideal configuration for a 12GB 3060?
Is the trigger word Penetratiom or penetration? Would you be able to share more of the prompts used to train this? having a difficult time trying to figure out how to prompt this off of the single example.