NullReal is aimed towards a mix of realism and CG.TriggerWord: NullRealEthnicity Tags<ul><li>african</li><li>african_european</li><li>south_asian</li><li>south_asian_european</li><li>european</li><li>middle_eastern</li><li>latina</li><li>east_asian</li><li>north_african</li><li>indigenous_mexican</li></ul><h1 id="thoughts">Thoughts</h1>v5 Feels like it's in a pretty decent spot. The lora applying to a character is hit or miss. Typically lowering the strength of the character tag or removing it alongside adding the trigger word can help.<h1 id="training-info">Training Info</h1><h2 id="v5">V5</h2><ul><li>Resolution: 1024x1024</li><li>Epoch: 100</li><li>Steps: 24000</li><li>Image Amount: 240 images (Same dataset as v4)</li></ul><h2 id="v4">V4</h2><ul><li>Resolution: 1024x1024</li><li>Epoch: 40</li><li>Steps: 9600</li><li>Image Amount: 240 images (Same dataset as v3 with ethnicity tags added)</li></ul><h2 id="v3">V3</h2><ul><li>Resolution: 768x768</li><li>Epoch: 40</li><li>Steps: 9600</li><li>Image Amount: 240 images</li><li>New dataset was generated via H3 Minimax adding more ethnicities, backgrounds, and lighting conditions.</li></ul><h2 id="v2">V2</h2><ul><li>Resolution: 768x768</li><li>Epoch: 81</li><li>Steps: 9963</li><li>Image Amount: 123 images (Same dataset as v1)</li></ul><h2 id="v1">V1</h2><ul><li>Resolution: 768x768</li><li>Epoch: 25</li><li>Steps: 3075</li><li>Image Amount: 123 images</li><li>Dataset was generated via H3 Minimax</li></ul>

Null Real

NullReal-v4.safetensors

Files with SHA256: 2cbd8491a8905fbbf1fea2d607bdafccc88de904b153bf21c62e04216cef9bb3