NEW V2 Update as of 11/13/25
This Lora is trained on (mostly) hands free tit POUNDING AND titfuck-blowjobs, if your looking for something involving the woman using her hands i recommend the other titfuck lora I have linked below. Trained on 256x256 videos at 16fps. 10k steps
Titpounding0 is the trigger word
PLEASE READ: Recently I have found that the BEST results come from square resolutions, 256x256, 512x512, etc... please also use the recommended "M4CROM4STI4" lora for better effect. The best sampler and scheduler is Bong tangent with res 2m, res multistep seems to be good aswell. cfg 3.5. Im still experimenting with Lightning Lora's but they seem to mess with the movement speed but enhance the quality so its up to you. I generally avoid them.
My personal settings are 720x720. 1 High, 0.8-1 Low for MY LORA, and for the Macromastia Lora i'd suggest 0.7 High and 0.5 Low.. Generally you want other loras to be around 0.7ish
General TAGs:
Positive Prompting
A [Add Detail Here] woman wearing [Add Detail Here], with [Add Detail Here] hair getting her huge breasts fucked by a [Add Detail Here] Male,Negative Prompting
Low quality, bad quality, blurred, blurry, pixelatedOther TAGs:
Male thrusting - To have the male doing the work
Dominant female - To have the woman doing the work
Camera angles
pov, male pov, side view, three quarter view, Hand placement
the woman is [squeezing/holding] her breasts together with her [hands/wrists/elbows]Dong visibility
the man's penis is completely enveloped and buried in her deep cleavage, the man's penis [is/is not] visible**NEW**The V2 model has been trained on TFBJ's as well (Note: This is experimental, so might be mixed results).
titfuckbj, fellatio, a video of a chubby woman sucking a man's penis between her huge breastsHappy Humping!
Description
This is Version 2 of the Titpounding Lora, hopefully the quality is better. All of the dataset has been recropped and resized so the quality is crisp. Still trained on 256x256 clips, 5 seconds. I added a lot more data to the set this time so it should be better with motion. The model uses "breasts" instead of "tits" now just to make it easier to prompt. Trained on 110 Videos of intense titpounding action i'v scrubbed the interwebs to find. I'm still new to this but i generally use weight 1-1.25 for the HIGH and 0.7-1 for the low, seems to work but try it out and let me know.
Improvements over v1.0
-More videos for the dataset in general.
-Better video quality for the dataset
-Side view should be easier to prompt now
-POV shots should be easier to prompt now
-Retagged dataset
TAGS
"titpounding0" is the trigger word and its in everything in the dataset so....
Huge breasts= Big boobs
Massive breasts= BIIIIIIIIIIIG boobs
Dominant female = To have the woman doing the work
Male thrusting = To have the male doing the work
You can specify if you want the head of the penis visible or not or not. Something along the lines of the mans penis is not visible or is visible.
FAQ
Comments (14)
I think, if someone speaks about "higher quality", then it would be helpfull to provide a couple of examples where it can be seen.
More samples coming soon I'm away
As others have state, the effect is great, but 256x256 sources means the final results lower the quality of the output to give it an "old video" look. A lot of the detail Wan is capable of is lost. If you can recreate your sources at 512 or higher, that would help tremendously.
Thanks ill try to upgrade the resolution for the next version. Im going to update the negative prompts because I actually did tag to avoid low quality like the last version. Negative: Low Quality, blurry, pixelated, bad quality should help
Another thing you can do, is use 256x256 videos, but with higher resolution images in the same dataset.
@Kwijiboe "low quality" probably doesn't cut it alone. 256x256 is very small, so I'd guess the LoRA is "seeing" that. I'm only academically aware of training quirks at the moment though, as I haven't gotten there yet.
@boobkake22 Thanks, would you reccomend maybe using higher resolutions like 512x512 or higher?
@Kwijiboe I'd suggest doing some reading around what people suggest. I would guess it's mostly an issue with low noise? Since it's the detail phase.
I cannot say if this is a good reference point or not, but I can say the LoRA works well, and they noted doing low noise at 400 with some 720 images to support. Tho I've also read using images encourages less motion? Again, I'm poor on experience here.
I checked your embedded workflow. You are using 30 steps and 4 CFG, right? That is too long to create a video. Does this work with lighting loras?
Nah, generally i dont use lightning loras it messes with the speed of the movement. Usually between 24-28 steps and 3,5 cfg
Ok, I tried 8 steps (4high & 4low) with Kijai lighting lora and it works perfect. And it works with 768 x 1152 resolution aswell. Perfect lora
@Kwijiboe Can you share your dateset and captions you used to train this lora aswell on this page? This would make a good use
@zirtapod ill share the dataset in the next version and captions should be in the description, i use prompt enhancers like dolphin ai alot but not always