The difference between a headshot you can actually use on LinkedIn or a proposal and the ones that go straight to the trash is almost entirely in the prompt inputs, not the model.
People who run 40 generations and still don't have one usable photo are usually giving the model "professional headshot of a [vague description]" and hoping volume fixes the problem. Volume just multiplies the same set of average guesses.
The four inputs that change the hit rate
- A real reference photo (or detailed description of one) of the actual person, not "a 35 year old woman in business casual."
- The exact use case and output constraints (LinkedIn profile, 4:5 crop for proposals, must work in black and white, must keep glasses exactly as they are).
- Lighting reference from a real photo the user likes (not "good lighting").
- One negative rule that addresses the most common failure for that face or tool ("no extra fingers on glasses," "keep ear shape exact," "no smoothing of skin texture").
Without all four, the model is still guessing the most statistically common professional portrait.
Table: input quality vs. usable output rate
| Input given | Typical generations before one usable | What usually breaks |
|---|---|---|
| "Professional headshot, friendly" | 30-50 | Wrong age appearance, generic expression, changed glasses |
| One reference photo + "make it look professional" | 8-15 | Still smooths skin or moves hair |
| Reference + use case + lighting ref + one hard rule | 2-5 | Minor crop or color tweak only |
The teams and individuals who get one or two usable shots in under 10 generations are not using a secret model. They are using the AI Headshot and Image Prompt tools with the four inputs above.
A concrete before/after prompt
Weak: "Create a professional headshot for a female lawyer in her 40s, smiling, office background."
Strong (what the headshot tool produces when you fill the form):
"Real photo of [specific person description from reference: 42 year old woman, shoulder length dark brown hair with slight wave, wears thin rectangular tortoiseshell glasses, light olive skin with visible freckles across nose, small mole above left lip]. She is wearing a navy blazer over a cream silk shell, no necklace. Shot in a real law office conference room with large window on left side, soft overcast daylight from window mixed with warm 3200K practical lamp on the right. 85mm lens, shallow depth of field, eye level, three-quarter view turned 15 degrees toward camera. Keep glasses frames, ear shape, and skin texture exactly as reference. No beauty smoothing, no extra fingers, no jewelry changes, no hair length change. Output 4:5 crop for LinkedIn."
The second version tells the model what not to invent. The first version lets it average every lawyer headshot it was trained on.
Common headshot failure modes and the exact rule that fixes them
- Glasses turn into blobs or disappear: "Keep the exact frame shape, thickness, and reflection pattern from the reference photo. Do not simplify or remove glasses."
- Skin looks airbrushed: "Preserve real skin texture and visible pores. Do not apply beauty filter or frequency separation."
- Hair changes length or color: "Hair length, part, and color must match reference exactly. Only adjust flyaways for realism."
- Background looks fake: "Real interior photo background, not gradient or blurred stock. Match the reference room's wall color and window direction."
- Expression is the same generic smile on every generation: "Natural closed-mouth smile with slight eye crinkle, as if listening to a client. Not teeth-baring or overly wide."
Workflow that produces usable files fast
- Take or pick 2-3 real reference photos of the person in the actual clothes they will wear. One should have the lighting direction you want.
- Run through the AI Headshot tool. Paste the reference description or upload if the tool supports. Fill the use case (LinkedIn, website about page, proposal PDF, black and white print).
- Generate 6-8. Pick the closest 1-2.
- If still off on one detail (glasses reflection, ear), take that one image into an image-to-image or upscale tool with a very narrow fix prompt: "Only fix the left ear shape to match reference, keep everything else identical."
- Crop and color correct in 2 minutes in any editor. Export at 1200px on long side for web, 300dpi for print.
The 40-generation loop usually happens because step 4 never happens: people keep rolling the dice instead of editing the one variable that is wrong.
When to stop using AI and book a photographer
If after two different reference photos and three different tools you still cannot get the glasses or ear shape to survive, the model is telling you the reference is too low resolution or the angle is too extreme. Book a 20-minute headshot session with a real photographer who uses continuous light and will shoot 50 frames. It will cost $150-300 and you will have files you can use for five years without second-guessing.
Use the AI step for the 80% direction and speed. The last 20% is still a human decision on which file actually looks like the person when they walk into the room.
The AI Headshot tool plus a single hard rule per generation is how you get to a usable file in single digits instead of triple digits.