Can’t say it turned out like I was hoping at all.

KatyB

<ul><li>Scale only UNet blocks above depth 2 → stronger adherence to prompt.</li><li>Slightly downscale text encoder adapters → better facial consistency in img2img.</li><li>Update metadata with tags so it loads cleaner in UIs (“img2img tuned,” “balanced strength,” etc.). Recommended img2img settings (SD1.5)</li></ul><ul><li>LoRA weight: 0.7–0.9</li><li>Denoise strength: 0.20–0.40 (lower = tighter face match)</li><li>CFG: 3.0–5.0</li><li>Sampler/steps: DPM++ 2M Karras, 20–28 steps</li><li>Seed: reuse the source image’s seed when comparing variants</li><li>Extras (optional):<ul><li>IP-Adapter / FaceID: 0.6–0.8 for very high face lock</li><li>ControlNet (Canny/Lineart): 0.35–0.55 to preserve composition/edges</li></ul></li></ul>

Files with SHA256: 9c12ef7ed276b42a9930f777f599ce12902cbcc4db197a57c34da3447237f09c