First draft of a concept LORA that helps to generate women riding bicycles. The sweet spot for strength, I have found to be 0.4. It depends on what other LORAs or prompt words are competing with it. Higher strengths from 0.7-1.0 tend to cause distortion in the coherence between bike and rider. I was surprised at how well this turned out.
Retrained with same dataset of images but added a lot more detail to captions describing the style of bike (road, comfort, hybrid, cruiser), view angle (side, front, three-quarters, front, behind, above, below etc.), lens length (wide, perspective, foreshortening, zoom etc.), and clothing style (plaid, corduroy, silk, tight, loose, etc.). I tried to reduce the total number of keywords in the training set, so that similar subjects get grouped together and features get strengthened. I believe that this will help to make prompted images be more coherent
This is a bit squirrelly as the geometry of a bicycle is complex, and I trained on many different angles, orientations, and focal lengths. Many realistic checkpoints generate pretty good girls on bikes, but they are better at women posing with a bike rather than women riding bikes, This LORA does a good job of that. Also, many prompts on realistic checkpoints generate women riding bikes for recreation. This LORA generates good images of women riding bikes in a city setting, commuting to work, meeting friends etc. They aren't wearing helmets and lycra, and they aren't riding road bikes.
Self explanatory. This LORA uses a different dataset of training images than v1. Quite likely NSFW.
The bike geometry, feet and legs need more development in the model. I am going to try to increase the training dataset by an order of magnitude and train for longer to see if I can get more consistent, coherent results.