WARNING: Consider this a beta, it still needs a lot of work.
Triggers / Tips
For style : 2d anime style (always use this).
For specific words, it was consistently trained on these terms:
penis, vagina, vaginal, testicles, penetration, insert, breasts, bouncing, jiggle, anus, anal, buttocks, plopping sound (for both sucking and penetrating), cowgirl, reverse cowgirl, doggy style, missionary, deep throat, thrust (for thrusting hips), blow job, suck, cum, cumming, deep breathing, heavy breathing, moan, grunt, insert, pull out, tentacle,
It is trained on sound: moan, grunt, use "plopping" sounds fucking / sucking sounds.
Adding other NSFW or motion loras (which are trained on 3d/IRL) will make the generation look less 2d. Just a warning.
V0.3
edit: added 8100 steps checkpoint, use that one, seems better.
Added back in general NSFW data, so vagina, cumming, anal etc. are back in the dataset. I also added in an extra 50 or so images, making the total now: 130 images and 412 videos. It can now do pretty much all the stuff the v0.1 could do but actually stable and working now.
I didn't have much time to check the checkpoints but the 5k steps one seems stable enough. Some of the samples seem weird with the vagina if you are prompting insertion. That I think can be fixed with better prompting. Sometimes bj the lips dont attach but rerolling seed or adjusting prompt fixes. I'll see if more training steps will help or not.
I'll work on another version to improve the vagina shape and try to get the anus to appear, but I think might take a little longer, I'm getting a little burnt out working on this so fast. And I finally feel the lora is in a good place now. Please post your examples and constructive feedback.
Did limited tests with lightx 8 step lora, and it does work with this lora (1str light 0.75-0.8 my lora) but you need to play around with the strength of both loras otherwise it goes off the rails. I recommend to not use them as they effect style quite a bit, let me know good settings and I can test. All samples are generated with 20 steps fully no speed ups. I also recommend the euler beta57 scheduler but simple seems more stable now.
V0.2
Note: Do not use euler simple otherwise it will be distorted, try euler beta57 (my favorite so far) or some other scheduler/sampler. Strength 1 recommended, maybe 0.75 if you feel distortion too strong but I keep it at 1. Not tested with lightx.
I had to go back to the drawing board and start over. I pruned the dataset down to 105 high quality clips involving only hand jobs or blow jobs and truncated them down to 75max frames. I also removed all furry content to remove any confusion for how dicks should be shaped. Then I added in a dataset of uncensored 2d anime style penis images. I also converted my captions from LTX format to H3 format with the help of gemma4 12b.
I trained the videos on 672x384 resolution and the images on 1024x1024 resolution. Batch 3 for images, 1 for videos.
I adjusted h3_guidance_distillation_scale down from 4 to 3.5.
By reducing frames, resolution, adding in images, I went from 20s/it to around 4s/it! And I reached a good enough check point at around 6k steps to conclude this approach is a success! So I will release this now, but understand that the model has drifted away from the 1300 clips base knowledge. Penetration is not working anymore including doggy style. But blow job, hand jobs, and penis shape has improved ten-fold.
So the next step is to add back in more data and train again so all the concepts are back.
There is a 5.7 and 6.1k steps version, 5.7k has less defined penis shape but seems more like normal size, 6.1k seems to make the penis bigger but better shape. I don't wanna train further because I want to add more data and concepts and start over.
V0.1
Trained on 1300 clips, so there is a lot of variety beyond just these. Some furry stuff in there but not a ton.
Write the prompts in h3 syntax, start with "2d anime style"
Works for t2v and i2v, i2v is better you can play with the strength, but I like 1 strength on both. Not tested well with lightx loras. In I2V if the genitals are already in frame it will keep their shape better.
I think it does blow jobs especially well for 2d. The base model cannot do blowjobs right for 2d. Not tested with lightx, these are generated with 20 steps on euler simple with comfykitchen optimization only. Everything else is native/kijai's nodes.
Training
This was trained on musubi fork on 1300 uncensored high quality hentai clips at 832x480 resolution 124 max frames. And it only reached around 2 epochs on this version. I will train more tonight and see if there is improvement. I think its under trained, but at the same time I had to revert to a earlier checkpoint to get the best result. In this state it's usable for certain things, especially with i2v so I wanted to put it out. While I maybe just rework everything.
The dataset and captions were originally meant for an even bigger LTX lora, but H3 was released so I decided to pivot this to H3. This is more of a learning experiment for me. I used a few LLM's to do a first pass on the captions, and then by hand brushed up around 1600 captions, I took out 300 dataset pieces due to odd aspect ratios that I didn't wanna deal with.
The size of the dataset is my biggest yet, and this is still not finished, consider it a beta/experimental. I'll do another round of training and if nothing improves, I'm going to start over from square one with a smaller dataset and fit everything around H3 instead of LTX.
I trained 4 versions before landing on this. H3 is very sensitive. I did 5k steps on V1, then loaded from safetensors and did 2400 steps in V4. Genitals in t2v are still not good enough, but it's picked up the sex positions and motions that go along with them.
Plans for v2.0:
Reduce dataset down to like 300-500 clips max. Rewrite captions in H3 format. Brush up on terminology ("plopping" is a bit awkward word, and I need to caption sucking and fucking sounds separately).
Big Thanks
Thanks to JonXL who helped me test and tinker with pretty much every version, not possible without his help. Also to everyone in discord Sulphur and Banodoco discord with their feedback and tips.
Description
Added back in general NSFW data, so vagina, cumming, anal etc. are back in the dataset. I also added in an extra 50 or so images, making the total now: 130 images and 412 videos. It can now do pretty much all the stuff the v0.1 could do but actually stable and working now.
I didn't have much time to check the checkpoints but the 5k steps one seems stable enough. Some of the samples seem weird with the vagina if you are prompting insertion. That I think can be fixed with better prompting. Sometimes bj the lips dont attach but rerolling seed or adjusting prompt fixes. I'll see if more training steps will help or not.
I'll work on another version to improve the vagina shape and try to get the anus to appear, but I think might take a little longer, I'm getting a little burnt out working on this so fast. And I finally feel the lora is in a good place now. Please post your examples and constructive feedback.
Did limited tests with lightx 8 step lora, and it does work with this lora but you need to play around with the strength of both loras otherwise it goes off the rails. I recommend to not use them as they effect style quite a bit, let me know good settings and I can test. All samples are generated with 20 steps fully no speed ups. I also recommend the euler beta57 scheduler but simple seems more stable now.
FAQ
Comments (23)
Bro, can I help you with data collection and description?
Also, I'm curious, did you train this LoRa based on the basic checkpoint or use a fine-tune model?
its a rank 32 lora, I think though I should've done it at 16. I'm afraid to change anything at this point until I hit a wall again. Though I did load weights from the 1300 clips dataset version and trained the 100 clips v0.2 on top that. I'm now training v0.3 which has around 400 clips and 130 images and I'm gonna start from 0 on that one.
if I need help with captioning the data, or getting more will let you know. I think now I just have too much data. I collected this for LTX, which didnt even know what anime is, and now h3 seems to know quite a lot so it doesn't really need that much data.
ok v0.3 is out, I may slow down a bit with updates. Since this seems like it can do all the general nsfw stuff now. I'll try to get more samples up later.
isnt training from hentai anime a bad idea to begin with ? since they have bad animation, quality, etc
There are quite a few good hentai animations out there. Not all of them are bad.
Can you please do similar lora based on Blender/3d animations?
Sorry I don't have interest in 3d animations, though I think this works with 3d animations already if they're anime style. Give it a try. Maybe I will when I port over my dispatch lora though.
@tazmannner379 Alright.
Ive heard that Fizgig is really good for training Minimax (and Krea2) and you can even train on 12-16gb of vram... and you can get very fast good quality with way fewer steps.
Musubi fork is very good though. This trained with 20 block swap and used around 23gb vram, later no blockswap it went around 30gb vram usage 5s/it on a 5090. I think with more blockswap you could maybe a lot more even. Fizgig is also only images atm?
Could I please inquire. Are there any fundamental difficulties that prevent Lora from training to produce a properly shaped pussy and anus?
yes there is, have you seen all the other loras here with similar issues?
@tazmannner379 I think almost everyone has this problem. I've only seen one Lora for H3 that can generate normal, realistic vaginas.
@yuduz367 This lora does vagina's fine but has issues with certain interactions with them. I think it can be trained around with the right data.
The model is super new my dude. It took forever for LTX2.3 to render anything at all (and it's still very bad at genitalia). And I've had better success with H3 than with anything so soon into the model's life. LTX2.5 is just a better audio generator for WAN2.2. H3 is just remarkable in all respect, especially with maintaining character likeness and predicting their likeness no matter the angle.
I'm having to learn that I can be more dynamic with my prompts because H3 will follow the commands. LTX2.3/2.5 absolutely falls apart after 3-4 seconds and rarely understands what you want. WAN2.2 looks and responds better than LTX2.3/LTX2.5 but has no native audio support.
If WAN and LTX had a baby, and sent that baby to a good school and it graduated college, then you get H3.
You can get realistic vaginas with H3 right now but you need to adjust strengths and the model and turbo lora (and its strength) matter a lot.
The full sized model is almsot always rendering thing better than the smaller ones for example.
IT takes a bit of experimenting but once you find a combo that works, SAVE THAT WORKFLOW. Duplicate to create a new flow for further experimentation but don't touch the original. That way you can compare changes if you make any within the duplicate.
@DaddyWolfgang nothing statement, helpful to no one.,,well done
@ntrlover he was pretty clear, adjust your lora strengths until you find a combination that works for your workflow. It's going to be different per workflow depending on what loras you use. It's been found going over a summation of 2.0 with lora strengths tends to start deepfrying the video, in my experience and others.
You're way too uppity for someone who provided nothing in return. Typical of an ntr consumer.
@smd123 you're right I skimmed through it, @DaddyWolfgang I apologize.
I think playing with strength and seed helps. I also had luck looking at the preview and sometimes at 17 steps or so the result was good and 20 steps bad so locking seed and adjusting shift to like 11 worked well. Or just roll a new seed.
I added a 9450 step checkpoint, I feel maybe the shape of things is better but its less stable. Can someone help test the few that are uploaded on this v3 version and let me know which is working better?
yeah seems like 8100 is better than 9450, I'll probably remove it for now.
Splendid work. May I ask if you use any ai model to assist the prompt writing? H3 seems to understand complex prompt and if only using manual input it is going to be very tedious....
Yes, claude is very good at it but it wont do NSFW. So for that grok will work but grok IMO needs like 20% of the prompt rewriten, even with custom instructions not to use certain terms. Feed the model these files first: https://github.com/MiniMax-AI/MiniMax-H3/tree/main/skills/h3-prompt-writing/references
Though just learning the syntax and writing by hand will almost always give better results. I tend to ask claude how to do certain shots or camera angles etc since it knows what h3 does to get those shots.
I haven't tried local llm for prompt writing but I did use uncensored gemma4 12b to convert LTX captions to H3 format captions and it did that well enough.