I'm pleased to present my latest motion concept LoRA, a kind of successor to my Riding Dildo Video LoRA. I've done a lot of experimenting since then and you should find this LoRA far more flexible and higher quality.
V2.0
New version is trained for 8400 steps vs 3600 steps of v1.0, an extra 25 videos in the dataset and the faces have been censored so it should work well now with character LoRAs. You should also be able to grab the workflow from the new videos I've uploaded, which is a slightly modified version of Flexbility's workflow.
Prompting
The key phrase is She is using a sex machine with a dildo attached to a metal rod. The dildo thrusts back and forth into her {asshole. anal object insertion|vagina. vaginal object insertion}. Some of the dataset had squirting or ejaculate which you can try prompting for.
This is the prompt I used for the sample videos (using impact wildcard processor):
A video of a nude woman with __haircolors__ hair in the style of __hairstyles__, {lying on her back|squatting|standing|on all fours facing away from the camera looking behind} on a {black|white} {couch|wood table|metal table|kitchen counter}. She has {medium|large} breasts. She is using a sex machine with a {huge|large|medium|small} {orange|pink|flesh colored|black} dildo attached to a metal rod. The dildo thrusts back and forth into her {asshole. anal object insertion|vagina. vaginal object insertion}. Her legs are spread wide. The background is a {dimly lit|brightly lit}. Camera is slightly to the side.Nothing fancy and lots of room for experimentation. The videos were generated using Flexability's excellent upscale workflow found here.
Videos were generated in 10 steps with FastVideo lora, FP8, guidance and shift both 10 and a single upscale v2v pass. Around 3 minutes per gen on my 3090Ti.
Training Details
This LoRA was trained using difusion-pipe on a dataset of around 50 videos at 300px res and 60 frames on a 3090Ti. Learning rate 2e-5, repeats 4. I'll upload the configs soon for more details if anyone is curious.
Future Plans
There is headroom for further training and I have some ideas for improvements, depending on how this is received so keep an eye out of for new versions.
Description
FAQ
Comments (23)
This Lora Looks awesome. Definitely going to try this one + it looks clean. You said "This LoRA was trained using musubi-tuner on a dataset of around 50 videos at 300px res and 60 frames on a 3090Ti. Learning rate 2e-5, repeats 4. I'll upload the configs soon for more details if anyone is curious."
I'd love to see the configuration. How long did it take on your setup?
Thanks, hope you enjoy! I get 120 steps per hour on my setup, so about 30hrs for this version of 3600 steps. I will probably train for another day or 2 as the loss is still reducing.
@Adaptalab0r I've uploaded my configs. Apologies I used diffusion-pipe not musubi, it was late when I posted this Lora! I experimented with musubi quite a lot but found diffusion-pipe to work better for concepts, but I may just not be using musubi right.
@definitelynotadog
Thank you. I didn't have any repeats in my training. Might be a key factor. If I made another Lora, I'll give you feedback.
@Adaptalab0r no problem. I think repeats can be used just to make sure the dataset is evenly trained. Because I use gradient accumulation steps of 4 (which is the default I think) there's a chance some videos would be under-trained or even skipped without repeats.
This LoRA's character identity seems too strong. It cannot be used with other character LoRA. The more prompts add, the more distinct the appearance of this LoRA's character identity becomes.
Thanks for the feedback. I can fix this for the next version by censoring faces, which I have seen other loras do.
@dowefa5467894 you might be happy to hear I found a script to censor all the faces in my dataset and a v2 lora is underway. I also added an extra 20 videos and improved some of the captions. Stay tuned!
Would this be in part because the Lora is using normal words that might ignore/overlap/overwrite tokens? The loras that use obfuscated words like ohx or 4n4l etc tend to avoid this from what I've read. If the model has no recognition of the trigger word prior to training, the lora should in theory be safe from competing to take over everything in the scene. If a hugely-connected word like 'woman' is used in the trigger, but meant for a different part of prompt, it would in theory be making a mess of the output.
Also. Other things might cause the issue. But... I hope this makes sense.
@nerfme Well, the video captions are about a paragraph long and use natural language so it would be impossible to only use made up words. As it uses a text encoder for training I don't think those trigger words have the same effect like back in the SD1.5 days.
It should work though as the missionary sex lora which is very popular here uses similar natural language and censored faces in the dataset and seems to work well. I'm happy with how the lora is following prompts otherwise so I think it's just learning the faces too much because there's a lot of them in my dataset and the learning rate is quite low.
@definitelynotadog
Oh hey sorry my reply in this thread was kind-of off-track anyhow, since @dowefa5467894 was mentioning face and prompt.
Yeah the captions are using natural language I get it now, that's cool. Chat-Gpt has made some of the most amazing nsfw prompts that worked straight off the bat and I've been hell impressed. I'm currently into about chapter 8 of the most excruciating mastb scene I've ever heard a bot come up with. I'm just past the point where I'm getting excellent natural language prompts for explicit actions. I'm keen to test more of them, as the weaker ones were working 100% on Kling. Getting Chat-GPT to give out the goods on these prompts is also a win hah.
My prompting method when I have a single made-up word for a person etc, I just exchange it for woman or man, or add it as an adjective to that word. (ohx or "ohx woman" etc). If it's an action, I fit it in as a verb in the prompt. It works very well. ("they are doing xyz" etc.). It cuts down the word length a LOT. And you can mix it in with natural language. :)
I love the Lora btw, running it right now :)
@definitelynotadog I really appreciate your hard work. I hope your hard work will bear good results. Thank you for your attention.
@definitelynotadog You are awesome dude.
@nerfme That's what I've found too. For character LoRAs, I don't include a class and just use ohwx. If clothing or location is repetitive, I'll caption like this, ohwx, ohwx wearing such and such in an office standing in front of a desk. In general, for characters, the less captioning, the better
For concept LoRAs like this one, instead of using woman or even person, I'll use a rare token that I'll avoid using at inference. Something like, sex_machine, a ohwx with a sex_machine thrusting in and out of a ohwx's vagina.
@definitelynotadog If you're talking about the face_blur_app mentioned in that long TIPS post on Civitai someone made, I created one that has some improvements to it. If you have a better one than mine would you mind sharing? https://github.com/garyfnbusey/face_blur_app
why not leave the meta in? i am seeing this a lot with other doing the same with hunyuan?
Ah I forgot to sorry, but it's just the flexible workflow I linked with ImpactWildcardProcessor swapped in. Flow/Guidance at 10 and the 1st upscaler denoise at 0.6
thanks for the info, but please provide the actual workflow. just post one up in the examples. Easy enough.
@StarW I'm using all my GPU training the next version, I'll post some when that's done and it should be an improvement anyway
Thank you so much for your hard work! Keep them coming!
Very solid Lora
The versatility and variance in positions is exceeding expectations.
I’ve even had it spit out double penetration pretty accurately albeit accidentally and attempts at reproducing with intention produced fewer solid results.
Combined well with deepthroat Lora as well, but the combination occasionally gives the woman a ballsack which I didn’t sign up for 🤣
Previews look 10 times better that the monstrosities that come out of my gens. Please post some metadata.
I tried with the workflow you've linked but it's even worse somehow.
Comes out great! Works well with all my LoRAs but does do a bit more dp than I would like. It's still great!