Every major image model in 2026 — Flux 2, Midjourney, GPT Image 2, Stable Diffusion — is a diffusion system at its core, even where the vendor has layered a transformer, a language-model planner, or a “reasoning” step on top. The underlying training objective is the same one it’s been since 2022: start from noise, learn to predict and remove it, and land on something coherent.

That objective has a side effect nobody markets: a pore, a strand of vellus hair, or a faint capillary blush is mathematically almost the same signal as noise. It’s high-frequency, low-amplitude, and irregular — exactly what the model has been rewarded, over billions of training steps, for erasing. Practitioners in image-generation communities call this the denoising trap: the model doesn’t fail to render texture because it lacks capacity, it fails because texture and noise sit in the same statistical neighborhood, and removing noise is the entire point of the architecture.

This is compounded by the VAE (variational autoencoder) that most diffusion pipelines use to translate between pixel space and the compressed “latent space” where the actual generation happens. At typical 8x compression, a single pore can occupy the equivalent of one to three latent pixels — right at the boundary the decoder has to guess whether it’s signal or artifact. Left with no other instruction, it usually guesses artifact.

A prompt that reliably produces believable skin has three components, and it needs all three — dropping any one collapses the effect back toward the model’s smoothed default.

Texture descriptors + specific lighting terms + exclusion negatives. All three, or the result reverts to default.

Layer one (texture descriptors) forces the model to preserve what it would otherwise treat as noise. Layer two (lighting) is what actually reveals that texture in the render — flat, frontal light hides fine detail even when the model has generated it, which is why so many “photorealistic” outputs still look synthetic under diffuse studio light. Layer three (negatives) explicitly disables the smoothing behaviors baked in from training on retouched stock photography and beauty content.

Subsurface scattering (SSS) describes what happens when light penetrates the outer layer of skin, scatters internally off blood and tissue, and exits at a different point than it entered. It’s the reason a backlit ear glows amber-red, and the reason skin looks like it’s lit from within rather than simply bouncing light off a surface — a distinction the eye registers instantly, even without being able to name it.

Without SSS cues, a portrait reads as a lit mannequin. With them, it reads as tissue under real light.

Golden hour is the most reliable SSS trigger precisely because every major model was trained on enormous volumes of photography shot in that exact light. Ears, the tip of the nose, and finger edges are where SSS is most visible — and where its absence is most quickly noticed by a trained eye.

Image models are trained on millions of photographs whose captions and metadata frequently include studio and cinematography terminology. That vocabulary isn’t decoration — it maps to measurably different, reproducible visual results, which is more than can be said for vague phrases like “nice lighting” or “professional photo.”

The practical rule: describe every light source with three parameters — position, quality, and color temperature. “Soft Rembrandt light from a large window camera-left, slightly warm” produces a consistent, repeatable result. “Good lighting” produces almost nothing you can rely on across generations.

A catchlight is the small reflection of a light source in the cornea. Its absence is the fastest way to make an otherwise well-lit face read as dead — the iris starts looking like painted plastic rather than a wet, reflective surface. Working photographers will spend minutes repositioning a reflector purely to place that one point of light; in prompting, it takes one precise phrase.

Counterintuitively, catchlights matter more, not less, in low-light portraits. A single bright pixel in an otherwise dark pupil does more to humanize a face than almost any other single prompt term — which is exactly why photographers carry a reflector even outdoors in full daylight, not for the key light, but for that one point in the eye.

A perfectly symmetrical human face doesn’t occur in nature — identical cellular growth on both sides of a developing face, inside a non-identical physical environment (different muscle use, different sleeping position, different sun exposure), would contradict ordinary biology. Despite that, diffusion models default toward near-ideal symmetry, because symmetrical faces are statistically overrepresented in polished training data such as studio headshots and stock photography. The human eye clocks the mismatch instantly, even without being able to explain why.

Imperfections aren’t flaws in an AI portrait — they’re the biological signature of authenticity.

Even a well-built prompt can’t guarantee the same character looks identical across multiple generations — that’s the structural ceiling of prompting alone. Three different mechanisms address this in 2026, and they aren’t interchangeable.

LoRA itself comes from a 2021 research paper by Hu et al. describing low-rank adaptation as an efficient way to fine-tune large models without retraining every parameter — the technique was built for language models first and adopted by the image-generation community shortly after.

Creators who consistently produce convincing portraits rarely ship the model’s raw first output. They run a two-stage pipeline: generation for composition and light, then enhancement for the biological micro-detail that denoising erased.

Most negative prompts are a generic dump — “bad quality, blurry, ugly” — which wastes the tool’s real function. A well-built negative prompt targets the specific default behaviors you want switched off, not vague quality complaints.

What you exclude in the negative prompt is as important as what you request in the positive one.

Weighted terms in parentheses (supported in Stable Diffusion / Flux workflows through A1111 or ComfyUI) intensify the exclusion. airbrushed:1.3 is the single highest-value term in that list — it directly counters the cosmetic smoothing every model applies by default on faces.

The lineup changed meaningfully since early 2026, and using the wrong tool for a given job still costs hours. Here’s the state of the field as of late July 2026, checked against each vendor’s own release documentation.

For a portrait built to withstand close scrutiny: Flux 2 Pro for the base generation, Stable Diffusion (via ComfyUI) for texture enhancement and controlled upscaling. That two-tool combination, not any single model, is what professional AI-art pipelines lean on in 2026.

For strongly cinematic, mood-driven portraiture, Midjourney V8.2 with --style raw reduces automatic stylization and preserves more photographic realism.

For commercial work where clean IP provenance matters, Adobe Firefly remains distinct from every model above: Adobe states its models train only on licensed Adobe Stock content, openly licensed material, and public-domain images, and it extends IP indemnification to paying Creative Cloud, Firefly Premium, and Enterprise plans (with narrower terms on free tiers and beta features). No other major image generator in this list currently offers a comparable contractual guarantee.

Before spending a generation budget on upscaling and inpainting, it helps to score the base output honestly. To be clear about what this is: the Photoreal Fidelity Index (PFI) is a simple five-axis checklist, not a validated psychometric instrument — there’s no inter-rater reliability data behind it, just a consistent way to triage a base generation before you commit enhancement time to it. Score each axis 0–10 by eye, average them, and treat anything under 6 as not worth enhancing further; fix the prompt instead.

In this example, an average of 7.5 says the base image is worth enhancing — but the eye/catchlight score of 6.0 flags exactly where the two-pass pipeline (Section 8) should focus first, rather than running a blanket upscale and hoping the eyes improve along with everything else.

• Using only positive prompt terms and skipping negatives entirely
• Generating at 512px and upscaling afterward instead of generating natively at high resolution
• Applying skin enhancement after upscaling instead of before
• Stacking more than five imperfection descriptors, which drifts the result toward “aged” rather than “authentic”
• Describing lighting as “good” or “professional” instead of naming a specific pattern (Rembrandt, split, butterfly)
• Ignoring catchlights entirely, especially in low-light or moody portraits
• Expecting prompt-only consistency across a multi-image set instead of using LoRA, –cref, or multi-reference conditioning
• Choosing a model based on hype rather than the specific strength the job needs (see Section 10)

• 3–5 imperfection terms, never more, never zero
• Name a specific lighting pattern with position + quality + color, never “good lighting”
• Include one catchlight term matched to the light source
• Generate natively at high resolution; don’t upscale a small render
• Enhance skin texture before upscaling, not after
• Use ControlNet Tile for any upscale pass
• Reserve inpainting for eyes, teeth, hands at 0.4–0.6 denoising strength
• For repeat characters: LoRA (SD/Flux) > Flux 2 multi-reference > Midjourney –cref, in order of precision
• Score the base output on the PFI (Section 11) before committing to enhancement

Every technique in this guide exists because current models optimize for global image coherence, not the biological truth of fine detail. That’s an architectural choice, not a permanent ceiling — and it’s already shifting. Flux 2’s move toward multi-reference conditioning and GPT Image 2’s “thinking before generating” step are both early signs that vendors are building planning and detail-preservation directly into the model rather than leaving it entirely to the prompt.

What won’t become obsolete is the underlying literacy: the physics of light, the biology of skin, and the specific vocabulary that translates both into something a model can act on. That’s the part no future model release replaces on its own.