The trained checkpoint gives you the best quality possible - it's a clone of the trained model. LoRAs are fantastic for lots of things, but for human likenesses they don't match up to Dreambooth trained checkpoints.
"I need a LoRA for my workflow"
It's easy to extract a LoRA from an SD checkpoint on your own if you prefer a LoRA for your workflow. For instance you can use the Auto1111 extension "Supermerger" to extract a LoRA. That said, LoRA will not match the quality of the original Dreambooth model (loss incurred). The original Stable Diffusion v1.5 ema only (~4gb) checkpoint was the base this model was trained on (For anyone who opts to make their own LoRA extraction). If there's any requests for a LoRA version I'll upload one as an additional version to this post.
Trigger with "Charliekirk Person"
(yes, put person also). Use "Hi-Res fix" with the settings you'd normally use for best results. Use prompt weighting and/or Controlnet to get maximum flexibility out of this likeness model. Trained locally using JP's repo (24gb VRAM req / arguably the best for likenesses).
Note on the Samples:
Some same images were generated using Control Net (particularly the 16x9 film style images). I don't think Civitai notes if Controlnet was used but you may not be able to reproduce some of the images due to that.
High Quality Dreambooth (JP's 24gb repo) trained checkpoint of Charlie Kirk created using high resolution images of the influencer. Trained with Dreambooth.
Why a full checkpoint/2GB model for 1 subject?
The trained checkpoint gives you the best quality possible - it's a clone of the trained model. LoRAs are fantastic for lots of things, but for human likenesses they don't match up to Dreambooth trained checkpoints.
"I need a LoRA for my workflow"
It's easy to extract a LoRA from an SD checkpoint on your own if you prefer a LoRA for your workflow. For instance you can use the Auto1111 extension "Supermerger" to extract a LoRA. That said, LoRA will not match the quality of the original Dreambooth model (loss incurred). The original Stable Diffusion v1.5 ema only (~4gb) checkpoint was the base this model was trained on (For anyone who opts to make their own LoRA extraction). If there's any requests for a LoRA version I'll upload one as an additional version to this post.
Trigger with "Charliekirk Person"
(yes, put person also). Use "Hi-Res fix" with the settings you'd normally use for best results. Use prompt weighting and/or Controlnet to get maximum flexibility out of this likeness model. Trained locally using JP's repo (24gb VRAM req / arguably the best for likenesses).
Note on the Samples:
Some same images were generated using Control Net (particularly the 16x9 film style images). I don't think Civitai notes if Controlnet was used but you may not be able to reproduce some of the images due to that.