Loading page
Model guide · Image · Stability AI
LoRA plus IP-Adapter, run locally
The full toolkit: LoRA for the identity, IP-Adapter for a per-shot reference, ControlNet for pose. It is the only stack where you control every variable, and the only one where the persona never leaves your machine.
What it is good at
Where it stops
The most setup of anything here, and base-model quality is behind the closed models unless you invest in the fine-tune.
Routes it can take you on
The syntax
This is Maya Rao from the cast, rendered into Stable Diffusion's own syntax. Square brackets are the things only you have.
<lora:[your lora trigger word]:0.8> Photograph of a woman, late twenties, South Asian features, an oval face, dark brown almond eyes, thick straight brows, a straight nose, full lips, a single dark mole high on the right cheekbone, mid-back length wavy jet black hair with a centre part, golden skin, natural texture and visible pores, average height with an average build. Wearing plain unbranded basics, a warm neutral palette of sand, bone and camel, in a small modern city apartment, soft daylight from a large window, camera at eye level. Medium portrait, three-quarter angle, looking into the lens. Neutral expression, hands relaxed and out of frame Negative: beauty filter, airbrushed skin, plastic skin, waxy skin, over-smoothed, perfectly symmetrical face, extra fingers, malformed hands, warped jewellery, text artifacts, watermark, logo on clothing, duplicate person, changing facial structure, retouched cheek, poreless skin Seed: 1753468036, CFG: 6, Steps: 30, Sampler: DPM++ 2M Karras
Build your own in the studio and this same block comes back filled with your persona instead — same structure, your locked line, your seed.