v Krea 2
Krea version 2 works without any bypass LoRAs or similar censorship modifications. It uses the same dataset as the latest Qwen-Image version, meaning various camera angles and positions are possible (e.g., from the front, side, behind, above, below, etc.). Krea 2 offers a wide range of possibilities.
You can also experiment with different strength levels—settings between 1.0 and 2.0 produce interesting results. In my experience, the best results are achieved with a strength from 1.0 to 1.25.
For the highest quality, generate at 4K resolution.
v0.3
This version introduces greater variety in the poses – not only from the front and the side, but also from behind, close up, and so on. It also includes the anus, visible from the front and from behind. The quality of the genitalia should be better – the images aren’t fried now. I know the logo appears in the images, but I was too lazy to remove it from the original photos. :-P
v0.2
The second version of LORA focuses on small penises, but in an erect state. It no longer requires support in the form of another LORA focused on penises – it works independently. It works best in the range of 0.5 to 0.8 – personally, I think the best level is 0.6. The best photos are from the front, side, bottom, standing, sitting, lying down, squatting, etc.
I encourage you to test it, because sometimes LORA surprises me with how effective it can be. There are also anatomical errors, but rarely, with complicated position descriptions, several people, etc. Tested on Qwen-Image 2512 and works great.
v0.1
LORA created based on photos of femboys and shemales with small genitals. Training in photos from the front and at a slight angle from above and below. Early version, may cause blurring of the photo, as the samples that meet the requirements were not always in the highest resolution. Trained on the body parts below the navel, hips, and mid-thighs so as not to affect the overall body structure.
I recommend using it with other NSFW lora at a level of (0.4-0.6) to improve quality.
I used lora Qwen4play_v2 (0.4-0.6) in the sample photos. The photo was generated on a Mac Air M1 with 8GB of RAM, so it should be tested on better hardware with a higher number of steps and higher resolution.
Comments and feedback are welcome.
Description
Krea version 2 works without any bypass LoRAs or similar censorship modifications. It uses the same dataset as the latest Qwen-Image version, meaning various camera angles and positions are possible (e.g., from the front, side, behind, above, below, etc.). Krea 2 offers a wide range of possibilities.
You can also experiment with different strength levels—settings between 1.0 and 2.0 produce interesting results. In my experience, the best results are achieved with a strength from 1.0 to 1.25.
For the highest quality, generate at 4K resolution.
FAQ
Comments (15)
Do your training images have watermarks?
Then you should consider masked training to prevent them from bleeding in.
Yeah I got the same problem...
I was too lazy dto manually edit all data images to remove watermarks ;)
@jakko sad, as that's destroying the quality of the image generation.
And there are automated tools that can help you (e.g. YOLO model for watermark & logo detection) and make that very easy. With a good captioning tool (e.g. my fork of TagUI https://github.com/StableLlamaAI/taggui_flow) you can quickly assess the placement and adjust when necessary.
@StableLlama thanks for the advice! Maybe next version will be without watermarks.
yeah I always manually remove
I use Krita because I can't use Photoshop anymore on my OS
Krita AI diffusion let's you generatively do it
But yeah if it's a huge dataset, I could imagine it'd be annoying
20-30 images, a total walk in the park
@Intosica with TagGUI Flow I'd say up to 1000 images are easily doable in high quality (i.e. with manual control). And any higher number is ok as well. Perhaps when it comes to 10.000+ then writing your own pipeline (still using the same YOLO models) might become beneficial.
I've had reasonable success with https://github.com/jferments/watermark_remover - I always check the results but I'd say it's 90-95% accurate with no manual intervention at all, just point it at a directory full of images
Works well with character Loras. Shame about the logo, its otherwise a great Lora.
Wow, amazing work!
Has a watermark
Its the best penis lora ever, but the watermarks kills it. :( :(
How you going to release a lora to the public and not even take the few minutes to crop out the watermarks? What a shame
I am father of two, I have a wife and I am on director position in a international company. I dont have time for this shit. :-D
works okay, but yeah the water mark... kinda lazy bro