I run the creative side of our agency and after building AI avatar tooling for a few years, I've gotten the process for a realistic result down to five steps. Not five prompting tricks — five structural decisions that happen in sequence and that each constrain the problem space for the step after them. When all five are done in order, the result is consistently in the "realistic" range. When any one is skipped, the generation has to compensate with luck. I'd rather have a framework than rely on luck.

None of these steps require you to be a professional photographer or a prompt engineer. They require you to make deliberate decisions at the right moments instead of defaulting to generalities. The framework is the same whether you're generating a personal brand headshot, an executive profile image, or a content creator avatar — the questions are the same, even if the answers differ.

Step 1 — Define the Anchor Image

Before any prompting begins, identify one photograph that represents the visual target as closely as possible. Not a celebrity reference — a real photograph of the actual person, in conditions as close as possible to the intended result. This anchor image does two things: it gives you specific details to translate into prompt language rather than making up specifications from scratch, and it gives you a comparison baseline to evaluate results against. If you don't have a photograph that's directionally close, find a lighting or composition reference that captures the quality you're aiming for. The anchor is the true north of the process.

Step 2 — Specify Light Before Anything Else

Light is the single largest determinant of whether an AI-generated image reads as realistic or generated. Before specifying appearance, expression, or background, lock the light: direction (which side, what angle), quality (hard/sharp vs. soft/diffused), and the shadow behavior you want. Catch light in the eyes should be specified explicitly — it's a detail that separates alive-looking from flat. Once the lighting specification is in place, it orients every other specification that follows.

Step 3 — Describe, Don't Evaluate

Write every prompt element as a description, not an evaluation. "Professional" is an evaluation — it tells the model a judgment, not a specification. "Dark navy blazer over white shirt, one button open, no visible tie, clean analog watch" is a description — it tells the model exactly what to produce. This distinction applies to expressions ("slight natural smile, soft focus, direct eye contact" not "friendly and approachable"), to backgrounds ("light warm gray seamless" not "clean studio background"), and to skin treatment ("visible natural skin texture, slight variation across the face, no obvious retouching" not "realistic skin"). Descriptions constrain; evaluations invite interpretation.

Step 4 — Generate in Batches, Score Against the Anchor

Single-shot generation is a lottery. Batch generation is a selection process. Generate 4-8 results from the same prompt and score each against the anchor image on four dimensions: physical accuracy, light match, expression quality, and background coherence. Pick the one that scores highest across all four — not the most beautiful one, not the most flattering, but the one that scores highest against the specification. Then identify what the winner did better than the others: that's the variable to reinforce in your next batch. Two to three rounds of batch scoring gets you to a result that isn't achievable in a single generation.

Step 5 — Lock the Configuration in Kyndrify

Once you've found a configuration that reliably produces results in the realistic range, the work of this framework is only half done if you don't preserve it. Manual prompting drifts over time — you simplify, you forget details, a new model behaves differently and the same text produces different results. Kyndrify addresses this at the structural level: the platform's button-based framework encodes your working configuration so that future generations start from the same specification. Step 5 is not optional — it's what turns a one-time success into a repeatable process. Without it, you're back to the lottery after every model update.

Five steps, each one constraining the problem space for the next. Light before appearance, description before evaluation, batch before selection, preservation before repetition. Follow the sequence and realism becomes a likely outcome rather than a lucky one.

Sources