<h2 id="flux-1e8942u5m">Flux</h2>Theo James in 2014.. All sample images are with no face detailer or hi-res fix, they are 1024x1024 resolution. They are generated on Euler at 20 steps. This lora is trained on 1024x1024 resolution as well. I used flux dev in the sample images. I've done testing and there are 2 ways my loras usually work best:<ol><li>Strength at 0.9-1 with no trigger word</li><li>Strength at 0.85-.0.95 with trigger word: th30</li></ol>i use comfy ui and theres a strength_clip option in the lora node and i set it to 0.85. Idk whats best but thats what i do<h2 id="sdxl-nhq9670qs">SDXL</h2>Theo James 2014. Strength 0.9-1.put this in negative prompt: ugly, deformed, airbrushed, photoshop, rendered, (multiple people), blurry, child, mouth open<h2 id="sd1.5-bk3sfmao9">SD1.5</h2>Theo James 2014. Trained on 768 resolution. This works with mostly realistic checkpoints. Strength 0.8-1. Here is the checkpoint I recommend using: <a target="_blank" rel="ugc" href="https://civarchive.com/models/462596/photonic-fusion">https://civarchive.com/models/462596/photonic-fusion</a>

Files with SHA256: 59e94d46b7e2ecfab9cc385c643a25c6468a2c461da6b5bfbcee19ff7663fbee