Fine tuned Lora on different mostly up close images of alita. If you use the lora the weight 0.3 to 0.8 gave me good results. Sometimes reducing the strength is better. Both Lora versions can be used at the same time with low stregth to get better results in some cases. Also if you mention the name alita in the prompt the alita style is stronger. You can also try to describe her emotions. I used words futuristic suit, tense look, serious look, smiling, sad, each of the words can be combined with slightly (slightly tense look, slightly smiling...). The Lora by itself generates safe for work images and was trained on the model sunshinemix_sunlightmixPruned.safetensors It’s possible to generate some hot images in combination with other loras. Here is the colab I used. It should be reproducible using the dataset here <a target="_blank" rel="ugc" href="https://colab.research.google.com/drive/1f2q75FOdMkpV_4o3RynVopjjM_bG41I6?usp=sharing">https://colab.research.google.com/drive/1f2q75FOdMkpV_4o3RynVopjjM_bG41I6?usp=sharing</a>

Alita

This Version is more general, therefore not as close to the original alita. Finetuned this on more variable poses/images of alita. Training data added here. Works great with the sunshine mix model from civitai (sunshinemix_sunlightmixPruned.safetensors). The lora is not trained on NSFW images, I just used it with some NSFW Loras to generate those images. Using both versions of this lora at the same time with low strength gave also good results. Worked well in combination with some different other loras. But the lora strengths have to be adapted to get better results depending on the image.

Files with SHA256: c42aef4a8fc7cdaccd04db6d3551df7503177ee165677213f0022211e5faaf38