AI Portrait Prompts: Describe, Don't Judge

Fashion portrait of a woman, head tilted back, chin raised, eyes fully obscured by a large bioluminescent snake coiled around her head, snake body wrapped three times around neck and shoulders like sculptural jewelry, individual scale definition with diamond lattice pattern, translucent scales emitting warm 2200K amber-orange light from within, subsurface scattering spill across adjacent jaw and neck skin, snake head with black elliptical eye and pale cream snout emerging from crown, deep burgundy matte lipstick with visible lip texture, pore-level skin detail on neck and shoulders, fine vellus hair catching the key light, subtle clavicle definition, pure white seamless background, single hard key light at 45 degrees upper left, Rembrandt lighting with defined shadow falloff, tack-sharp scale edges with background falling to pure white, photorealistic skin rendering with visible sebaceous sheen, shot on Hasselblad H6D-100c with 120mm macro lens, f/8 --ar 2:3 --style raw --s 250

The text encoder does not read a prompt as a sentence. It tokenizes the string, embeds each fragment separately, and cross-attention redistributes weight across those embeddings at every denoising step. Early token groups fix the coarse layout — subject, pose, framing — while later tokens refine material and light. That ordering is why the first noun phrase in a prompt is not a stylistic preference but a structural commitment the remaining tokens must negotiate with.

Where Evaluative Words Land in the Latent

Evaluative adjectives are the weakest tokens in any portrait prompt. Words like "beautiful," "professional" and "hyperrealistic" appear in millions of training captions attached to wildly different images, so their embedding centroid is a blur. Cross-attention still assigns them real weight, but they return almost no spatial, material or spectral instruction. The sampler spends denoising steps on tokens that cannot resolve into pixels, and the budget those steps consume is not refunded elsewhere.

Descriptive tokens behave in the opposite way. "Diamond lattice scale pattern" maps to a narrow region of the latent: repeating geometry, regular spacing, hard edge definition. "2200K amber-orange" pins a spectral band. "Subsurface scattering" specifies that light enters a surface, scatters and exits elsewhere. Each is resolvable into pixels. Engines tokenize differently, so the same evaluative word behaves inconsistently across platforms such as Adobe Firefly and Midjourney — description-first phrasing is portable precisely because it does not depend on one encoder's quirks.

Why "Hyperrealistic" Produces Plastic Skin

"Hyperrealistic" is a genre label, not a material instruction. In caption data it sits beside over-sharpened HDR retouching and genuine macro photography alike, so its centroid drifts toward the retouched end of that cluster. The symptom is recognizable: pores rendered as uniform noise, skin that reads as vinyl, specular highlights that stop abruptly instead of rolling off. Naming a quality is not the same as producing it, and the sampler has no mechanism for checking whether the output actually matches the word.

The corrective is to describe the artifacts of realism instead of realism itself. "Pore-level skin detail on neck and shoulders," "subtle clavicle definition" and "visible lip texture" are all resolvable constraints with visible outcomes. The same discipline drives the ultra-luxe golden skeleton portrait, where anatomical terms replaced mood language and the render tightened as a direct result.

Lighting Stated as Physics

"Rembrandt lighting" works as a token because it names a specific pattern with a known shadow shape. It only holds, however, when the geometry around it is stated. "Single hard key light at 45 degrees upper left" supplies an angle, a direction and a quality in one phrase. Remove the angle and the named pattern activates a diffuse cluster of studio setups; the shadow falloff then lands wherever the model's prior prefers, which is rarely on the cheekbone where it belongs.

Emissive light is harder to control. "Warm glow" drifts toward a generic orange because warmth is relative and has no anchor. "Translucent scales emitting warm 2200K amber-orange light from within" fixes both the spectrum and the origin point — inside the geometry, not painted onto it. The companion clause, "subsurface scattering glow on adjacent skin," is what makes the light behave like light: it must spill onto whatever it touches, and the jaw and neck are what it touches here.

The Stylize and Style Raw Trade-Off

--s 250 is a moderate stylize value, and it pushes the sampler toward the model's own aesthetic prior. That prior was trained on attractive imagery, so it also pushes toward retouched skin — the exact failure mode described above. --style raw counteracts this by reducing the model's default aesthetic weighting. Midjourney documents both flags in its official parameter reference, and the interaction between them matters more than either value read in isolation.

The trade-off is genuine. Raw yields tighter adherence to descriptive tokens but flatter, less composed frames. At --s 250 with raw, the prompt's own lighting and framing language has to carry the composition. Drop to --s 100 and the result turns clinical; raise to --s 600 without raw and the snake's lattice softens into a smooth tube, because aesthetic smoothing outbids texture tokens at high stylize values.

This is why the camera block is not decoration. "f/8," "120mm macro" and "Hasselblad H6D-100c" are grounding tokens for depth of field and perspective compression. They tell the sampler how much of the frame stays sharp and how the head flattens against the lens. Without them, "sharp focus plane on scale texture" is an instruction the model cannot parse — it has no concept of a focal plane, only of relative sharpness across depth.

The Next Variation to Run

Shift the snake's emission to 3200K and drop the key light to a 30-degree angle. Expect the amber spill on the jawline to cool toward neutral, the shadow side of the neck to open up, and the scale lattice to read as texture rather than silhouette. The frame becomes less theatrical and more clinical — useful when a portrait has to sit beside product photography in the same layout.

Then strip every evaluative adjective from the prompt and re-run it. If the output does not lose quality, those words were never carrying weight. That audit takes two minutes and applies to