I've captured the tokenized identity <a target="_blank" rel="ugc" href="https://twitter.com/naturevrm">NatureVRM</a> as a full-body LoRA. The training process is described in this blogpost <a target="_blank" rel="ugc" href="https://hackmd.io/@reneil1337/avatar-lora">https://hackmd.io/@reneil1337/avatar-lora</a> and can be applied to your own VRM avatars. Publishing the SD LoRAs under cc0 licence. Enjoy the model and please share your creations.<h3 id="prompt-via-&quot;..-naturevrm-a-person..-&quot;-kjrtpjkp4">Prompt via ".. NatureVRM a person.. "</h3>Image training dataset - SD15 v1.0 was trained with 100 images in 512x512 resolution - SDXL v1.0 was trained with 100 images in 1024x1024 resolution - Flux.1 D Beta was trained with 100 images in 1024x1024 resolution

NatureVRM

Increased training data size to 100 images. Improved descriptions. Integrated regularization images by aitrepreneur (person prompt) into the training. Used training data in 512px after realizing problems with higher resolution images.

Files with SHA256: 0f4abb1a5fcac437d881b8a97e9f89d69f5456b65b08f7b1537a67955e266b84