Edit Oct 13:
After several users approached me and mentioned that 2GB is too much for a breasts LoRA (very rightfully so) I have released an additional version that's 1GB (Rank 64), the results are not as good as the full 2GB version but definitely works.
I have done multiple training runs on 2xH200 on RunPod with ranks 16/32/64/128 and ended up uploading the 128 variant as it seems to be generating the best results.
Dataset was trained with the following key words:
tiny/small/medium/large sized areoles
tiny/small/medium/large sized breasts
pale/brown/dark/ghost areoles
hard/erect nipples
These definitely help with generating variations in results.
Most of my testings were done with my Qwen character LoRAs (none of them had naked images in the dataset)
I've also did some testing without my character LoRAs and the results were similar.
This is not perfect but much better than base Qwen.
Join my Discord server for updates on new LoRAs, tips & tricks, Workflows, RunPod templates and the holy right to be close to an amazing charismatic person like myself:
https://discord.gg/fyha5Pzm
Enjoy!
Description
This is the same dataset and captioning just at a lower rank.
Performance is a bit worse than V1.0 (Rank 128) but it's half the size.
FAQ
Comments (29)
Hey there @HearmemanAI, been using your runpod templates since I started Comfy long time now. This past week for some reason, your hearmeman/comfyui-qwen-template:v5 has caused both of the diffusion models to create images of pure noise. I've deleted them and redownloaded straight from the source and that seems to fix it. But upon restarting the pod I run into the same issue again. Have you dealt with this before?
I have the same problem. The Lora seems to apply noise to the whole image and the model doesn't target specific zones
I am not sure how this relates to this post?
You can send me a message or consult in my Discord server.
Anyways, I've deployed the template multiple times in the past few days and did not have any issues.
I've used an RTX Pro 6000 and an H200 SXM.
Have you tried deploying different GPUs and observed the same result?
Do you have a network volume? if yes, try to deploy without it.
test a bit, work, but also affect the race/face/skin some how, need to lower the weight. The difference between tiny and large is not significance.
Thanks.
I really tried to find the sweet spot between consistent results and not influencing the generation too much.
I found it to work very good with character LoRAs, at least the ones I trained.
@HearmemanAI I'm using it in my qwen workflow. At 1.0 strength, it does modify the face everytime. Drop it to 0.50, and you will retain the face plus get the effect intended by this lora.
@fedupscribe687 That's why the strength I recommend is 0.8 and not 1.
Looks very good from your previews but rank 128 is way overkill for a lora that is literally just for breasts. I guess it would be acceptable if it handled the full scope of the female anatomy. But for breasts only - a lora that is bigger than SD 1.5 at FP16 - is just too much. I'll give it a try when I can anyway :D Thank you for doing this
Totally agree, I probably wouldn't have download it myself lol.
Having said that, I do see distinctly different results from rank 64 to 128.
I might release the 64 rank variant as well, it's working roughly the same for close shots but lacks at a distance.
@HearmemanAI Qwen struggles with nipples a lot for some reason. I'm trying to train a LoRA for the Edit model myself - different concept though. Its struggling a lot. Granted, its only rank 16. Seeing yours makes me believe I'll have a much easier time if I increase it to 64. Over that I really think its just too much. I'd probably not even consider it until I saw yours so once again, thank you.
EDIT: please release the 64 one if you are willing - half the size is significantly more reasonable even if its not as good
@huj0ps1t6 I have released the Rank 64 version as well.
Does this work with qwen edit 2509?
Haven't tested
For future readers - it works fine
@Kiefstorm How do you use it with 2509? I tried and it did nothing at all, though sometimes it made my results extremely artifacted at random. Are you using it on only specific high or low, or like high none, high with, low with? Certain value like 0.5 strength? CFG 3.5 or 1 with/without lightning? Thanks.
@kudon44 i found that strength over 0.80 reduces the quality of the generated image greatly. But under 80 it works fine. The way I use it is in a normal workflow with a subject I say something like "remove the shirt. She has medium breasts and small sized areoles". It seems to work fine to add the breasts, but the "areoles" lol doesn't pickup as well. I'm sure it works fine for QWEN image. It also works well for WAN. But for just generating realistic breasts in 2509 it works every time
Udder lora pls!!
yeah saggy breasts would be a big lora. It's available for a lot of other models as lora's on here. You can roll the dice enough on this lora and eventually get some natural/sag, but you'll probably go through a lot of perky/fake type boobs first.
You trained with areoles? Not areolas.... okay
lmao huge catch.
Looks amazing works but I can't use it directly in my AI test because it's too much heavy files. Please, do you think it's possible to do a "LITE" version below 900mo ??? Please Thank you sir
This is actually such a great realism lora if you just leave out any of the trigger words
Yep, I use it a lot just for realism
Noise the entire image.
next tain only on closeup breast and a view upper full body...
The Rank 128 version seems to work pretty well in QwenImageEdit2511_bf16 with fp8_e4m3fn.
im always getting noise in my images
looks nice, would be much needed on Qwen Image 2512
can this add nice fertile (not elderly / sickly) looking veins?