Hello,
Here is the BulkedUp AI LoRA that was made with Kohya's GUI. The purpose of this LoRA was to create bigger buff dudes on different Stable Diffusion models. I personally use the LoRA between 0.3 to 0.5 strength, with 0.3 strength adding a bit of muscle and 0.5 going even bigger. Although this LoRA can work on different models, this LoRA works really well with a normal model merged with a furry one. One model merge that I'd like to recommend is a model merge between Gay621 Solo Male Focus and Homoerotic v2.0, with a weighted sum multiplier of either 0.3 or 0.5.
Using default E621 tags with spaces, like huge muscles, works really well with the LoRA
One thing I would like to mention about this LoRA is that if it output dudes that are too huge or have bizarre anatomy, sending them to img2img or inpaint with high denoising strength (between 0.4-0.7) could really help fix them.
= Training =
- This LoRA was trained on photorealistic AI images, with each image containing only one muscular male.
- The LoRA was trained on a model merge between Yiffy-e18, Analog-Diffusion-1.0, and HassanBlend 1.5.
- The training settings for the LoRA is from Aitrepreneur's LoRA training video, which is found here: https://www.youtube.com/watch?v=70H03cv57-o&t=960s
= NOTE: The LoRA training used the LoraBasicSettings.json file from the linked video, with "mixed_precision": "bf16" and "save_precision": "bf16", changed to "mixed_precision": "fp16" and "save_precision": "fp16" due to bf16 giving an error when used.
Description
VERSION 1.5: BULKEDUPAIR1.5
= CHANGELOG =
-Added 15 more images to the training dataset, featuring bigger AI generated photorealistic muscular male anthros.
Image Grid Model Comparisons from seeds 1-10:
-Gay621 Solo
-Gay621 Solo/Homoerotic v2.0 (0.5 weighted sum merge)
-Homoerotic v2.0
FAQ
Comments (22)
Where can I find models G621a-H2.0(0.5)?
Oh god, my apologies for the late reply.
The G621a-H2.0(0.5) model is a weighted sum merge of Gay621 and Homoerotic 2.0 with a 0.5 multiplier. Gay621 and Homoerotic can be found here:
- Gay621
https://civitai.com/models/10319/gay621-solo-male-focus
- Homoerotic 2.0
@Mr_Yakuza Thanks
Great work on this, I use it a lot! :-)
Thank you! I'm glad that my model is helping you out in getting some stacked studs 🤘
@Mr_Yakuza Cross-posting hasn't been working for me for ages. Now maybe you'll see a few more of mine show up, as it seems to be working again. Maybe. Hopefully.
这是个很好用的lora，可以帮助调节体型，而且我目前还没遇到与其它lora相互污染的情况，在-0.3~0.3内使用较为符合人类审美
is there a way the prompt words can be changed from balls to testicles?
Every now and then when I put the prompt "big balls" or "huge balls", it literally creates a giant ball as in the toy or sport equipment.
I have the same issue. I'm looking into it and will comment back if I find a solution.
The creator of this model is right. Mix in a furry model and it will alleviate this issue.
Hello,
The LoRA can make some huge muscle dudes on other models, but it struggles when generating genitals on base models that weren't trained on e621 tags, genning actual balls. The problem seems to get resolved when merging a furry model with the base model, getting rid of the weirdness.
So far, all the loras I've tried that use trigger words composed of two words, have weird behaviors as the individual words of the trigger words can have different meaning (as in your example) and can "bleed" to other aspect of the image, just like "blue eyes" in a prompt will pretty much always bleed the blue color on clothes and other object in the image.
On the other hand, all the "good" loras I've tried that actually gives the expected results have unique keyword(s) that are nowhere to be found in a dictionary, thus can only have the meaning they have been trained for in the lora. In the case that interest you "bigballs" is not an intelligible word, while "big balls" actually are, and worse: they can have multiple meanings, toghether as an expression or each words of the keywords seperately. Not to mention that the scientists who developped SD most probably didn't trained the base 1.5 model with dirty/pervert expressions and slangs.
Considering your focus seems to be big dudes with exagerated penis, you should probably change the name of your lora to something that better reflects what it is trained for. I like fitness women, but I hate to get random pictures of a woman with an exagerated big penis.
Plus, think about all the missing opportunities to get your work known to dudes loving big penis who don't see your lora because the search engine didn't prioritized your page.
Thank you for the recommendation! I'll adjust the tags to better reflect the LoRAs.
great, but it keeps making multiple nipples on models.
Thank you for your comment. I've noticed that too with the gens on different models. I can't tell if it's just SD 1.5 jank, or bad LoRA training settings. Hopefully I can find a better model/LoRA training combination to alleviate that.
I'm having a problem where it seems to want to put a black cloth over the darker areas of a model's chest that increases as the strength goes to 0.4 and above. Is this a known issue?
Thank you for your comment. Try the newer version, version 2, of the LoRA to see if it alleviates that issue. I'll also keep watch if that happens with the gens I make and see what I can do to tackle it. 👌
COuld you do a hyper balls only version?
My apologies for responding so late. I wanted to do this LoRA to be flexible with generating hyper muscles/balls/penis. I don't have any plans in making a LoRA for each specific part unfortunately.
I suddenly came across it again. I really like your lora, but it seems like it hasn't been updated for a long time. Are you interested in making an il or nai version?
Thank you for your comment. And you bet. After learning to train a LoRA with NoobAI and Chroma with my 7900XTX, I'll def be improving this LoRA in the future. 🙏