A double handjob blowbang.
Include the keyword bl0wb4ng in your prompt and wan will handle the rest.
v2
I removed the most messy cuts from the dataset, then re-captioned and re-trained it.
🟢Much more stable
🟢More controllable (see below)
🟢Better compatibility with other LoRAs
🟢No more men grabbing each other's winner in the background lol
Prompting
Typical prompt
bl0wb4ng, one woman and {two|three|four|five|six} men, {blowjob|handjob|switching}, {static|dynamic} shot, ...Example
Unused words can be added in the Negative prompt to increase adherence.
Positive: bl0wb4ng, one woman and two men, handjob, ...Negative: three men, four men, five men, six men, blowjob, switching, ...Usage Tips
If you don't know where to start, the first showcased video contains an easily reproducible workflow, with as few custom nodes as possible. Just download the video from its dedicated page, then drag-and-drop it into the ComfyUI canvas to display the workflow.
The dataset is distributed around four men, with two and six being the least represented number of people. Wan will usually draw them with a precision of +/- 1 person, it's not 100% reliable.
It can draw a progressive transition instead of a cut, by setting the strength of the high-noise lora to ~0.5-0.65 and adapting the prompt accordingly (setting the number of men to two seems to work best).
This is what the action words imply:
handjob: double handjob
blowjob: double handjob + blowjob
switching: double handjob + blowjob + changes
Training
v1.0: 67.5 epochs, 72 5-seconds clips (4860 steps).
v1.1: 80.0 epochs, 72 5-seconds clips (5760 steps).
v2.0: 75.0 epochs, 56 5-seconds clips (4200 steps).
Description
v2 for Low-noise
FAQ
Comments (14)
Great job! Much more stable indeed!
Omg, I couldn't figure out how to get the other guys to stop jerk in each other off, some times they would pass their winners around too. It's so hard to tell a lot of times if it's the loras in general or you didn't apply something right. Thanks for the update
Yes, it was a frustrating side effect. I used to add words like "people in background", "gay", in the Negative prompt to avoid it, it helped a bit, but normally it is not needed anymore.
@qdr1en lol, I had a few in this new version were the guys started kissing each other and gave a couple tugs on another guy but is much less than before. I think it's not so much the lora as it is the model just try's to use what's the closest and most convenient. I read that the first 20 tokens of a generation are the most vital that's where all the decision making is done so maybe there's something in that could help with changing the way it acts but it's like how do you know how long 20 tokens lasts?
@burn1one1one812 for me it's a sign you prompt too hard. Probably you won't get anything like this if you remove all the NSFW stuff from your prompt, and replace them by more suggestive words. Some wan's censoring mechanism is likely triggered here.
I don't know for the first 20 tokens, but I read that adding the same prompt translated to chinese below the english one helps at maintaining prompt adherence.
"No more men grabbing each other's winner in the background lol"
God, I love this website 😂
Having an issue where she keeps turning around and facing her back to the camera how do you stop that? I took switching out of the prompt and tried going up and down the strength nothing seems to work, its like every other generation.
It's strange, I never got that. Maybe start again with a clean simple prompt like the example I gave, and empty the negative prompt as well. Make sure you use the base wan models (not a custom merge). Also, avoiding NSFW words in the positive prompt helps smoothing things up, as well as adding "porn" in the negative.
If it works that way, then you can re-add elements to your prompt progressively, until she turns around again. At this point you will know which word causes that
@qdr1en So does that still stand if I'm using the NSFW text encoder? I've seen a few people say to never use them but it always seemed I ended up switching back to the nsfw because the regular would get too vanilla. Yes, I used both the base fp16 and fp8 models. I'm thinking it may have possibly been something to do with the color of her shirt and the color of her hair were both black so the model maybe couldn't tell if she was facing forward or back. I noticed the guys would sometimes all be wearing the same outfit the woman is wearing idk why it does that either lol. I'm learning so much from your information I really appreciate your help and your work. Thanks
@burn1one1one812 I deleted that encoder from the disk. From my tests, it did not add anything good or i did not see the benefits.
For the clothing, maybe it's because of the model limited attention: the guys' clothing will be different if you prompt for it, but if you don't, it focuses on the main action instead.
Very nice Lora, sometimes a bit hard to control but in general great results, especially 2 handjobs, 1 blowjob seems to come out good.