Realistic Mannequin Character LoRA Training
I enjoy challenges, so I wanted to work with generative AIs in a way that avoids any issues with image rights or copyrights for my character. That’s when I decided to create a unique, unrecognizable character and thought of mannequins.I began training a LoRA model, but it wasn’t easy because the images I found were very limited, with almost no clothes and very few poses. The initial results were somewhat poor since training for this kind of model is more complex than a typical LoRA. I needed to account for the face, body, poses, and clothing. Generating images was frustrating because I often ended up with multiple figures in the image, hallucinations, and other poor results. I owe a lot to the Tensor Discord chat rooms for helping improve my generations.Here are a few tips for getting better results: generate images around 512x512 pixels because larger sizes tend to result in multiple people in the image. You can then use HiRes to upscale and ControlNet for the poses.https://tensor.art/models/705013938401653144Once I had a good number of new mannequin images, I moved on to version 2.0. With a better and more varied dataset, the results improved significantly. The generated images had fewer hallucinations, better compositions, and it was easier to achieve a good outcome.https://tensor.art/models/727667919498229943I initially created the first LoRA for SD1.5 and the second for SDXL. I’m still figuring out which checkpoint is better, as I continue to discover the strengths and weaknesses of each.#AIArt #DigitalArt #LoRA #LoRATraining #AIModeling #CreativeAI #SDXL #SD15 #AIChallenges #TechArt #DigitalCreativity #AIExperiments #TensorCommunity #TensorArt #ControlNet #ImageGeneration #ArtisticProcess #AITraining
