V7 WE DID IT!
The results are very consistent.
Pro tips: Use detailed prompts like:
<lora:Twinks_Legs_Up_V7:1> artistic intent to portray the subject in a natural and unobstructed manner, with light skin and straight brown hair, wearing only white ankle socks and a thin silver chain necklace. The man's nude body is lean and muscular, adding a tactile element to the scene. The setting appears to be indoors, with slightly parted lips and direct eye contact with the viewer. His body is slender with a toned physique, lying on a bed with his legs spread wide, and a small plant. The walls are painted a neutral color, with his hands resting on his lower legs.
The model still has a penis deficiency use img2img with androflux or twinkx to fix that if needed
Square images work the best!
Good images at least in my tests: 8/10
If you like the work of Human Diffusion consider tipping Buzz or Posting Pictures with our LoRAs
Description
FAQ
Comments (2)
So what did you learn training a concept that Flux doesn't know?
What should others do better in the same situation?
For anatomy like this i would say big Datasets is the most important. The significant jump in accuracy resulted when i went from 25+ Images to 45-50.
Also in complex structures like legs etc lot's of steps and smaller batch size is much needed. For example 2 or even 1 Batch Size so the Network can chew the images over more steps will results in way better results without pushing epochs too high. Which will result in the Network learning unnecessary things. I had a case with a different LoRA where it actually instead of learning the subject it learned the decor of a bathroom which was present in 3-4 pictures.
Network Dimension also plays a role especially if it needs to learn a lot of details.
In the case of the Pup Mask for example where the only thing it had to learn was a Pup Hood i got away with a small LoRA of size 2 but this where it had to learn body type, legs, ass, hole, anatomy etc. Higher Dimensions certainly made it way more capable.