AI Image Generation for Marketing Sites: What Works, What Trips the Slop Detector

AI image generation has gotten remarkable in 2026. Midjourney v8, Stable Diffusion 4, DALL-E 4 — the output quality on the right prompt is genuinely indistinguishable from stock photography. The wrong prompt produces the same AI-slop tells (melted hands, plastic skin, six-fingered grips) that have become a SERP-wide cliche. Here’s the curation framework we use at FH to decide whether an AI image earns its place on a client site.

#Where AI imagery works for marketing sites

Abstract or conceptual visuals (gradient backgrounds, geometric compositions, atmospheric scenes) where photorealism isn’t required.
Product mock-ups for products that exist but you don’t have professional photography of yet.
Backgrounds and texture overlays — non-foreground assets where the AI tells are invisible.
Icon and illustration generation in a defined style (consistent line weight, color palette, etc.).
Reference imagery for mood boards before a real photo shoot.

#Where it doesn’t

Photos of your team. Don’t. The uncanny valley is too obvious. Hire a photographer.
Customer testimonial photos. Same reason, with added authenticity risk.
Photos of completed client work that the client could verify. Misrepresentation risk.
Anything with prominent human hands or text. Both are AI’s persistent weakness.
Anything where authenticity is the marketing message (local roots, real craftsmanship, real people).

#The FH curation rubric

Every image headed for FH client storage runs through a curation flow that scores it on five axes:

Brand fit (0-10): does it match the brand’s color palette, mood, visual language?
Slop risk (0-10): how many AI-tells visible? Hands, faces, text, plastic skin?
Stock-cliche risk (0-10): generic handshake? Office laptop? Diverse team posing?
Channel appropriateness (0-10): right aspect ratio? Right resolution? Right format for the intended channel?
License clarity (0-10): is the source documented? Is the license commercially usable?

Total score below 30 = blocked. 30-40 = revise. 40+ = ship.

#AI-slop tells to check for

Hands with wrong finger counts, melted fingers, or impossible joints.
Text in images that’s gibberish or partially-formed letters.
Faces with asymmetric eyes, mismatched ears, or skin that looks like plastic.
Background details that don’t hold up under scrutiny — buildings with too many windows, vehicles with logical impossibilities.
Lighting that doesn’t obey physics — multiple light sources from inconsistent directions, shadows that don’t match.
Compositions that are eerily symmetric or perfectly centered.
Excessive bokeh or depth-of-field blur masking detail issues.

#Stock-cliche patterns we block

Even when an AI image looks technically perfect, it can still scream ‘stock photo.’ The patterns we block on FH client work:

Generic handshake imagery (every B2B site in the world).
Office team gathered around a laptop pointing at the screen.
Whiteboard-with-marketing-jargon-written-on-it shots.
Diverse-team-posing-by-the-window shots.
Lone-developer-hoodie-at-rooftop-laptop shots.
Floating-keyboard-with-graphs-overlay shots.

#The metadata sidecar

Every image in FH storage has a sibling `meta.json` documenting source, prompt (if generated), license, intended channel, human subject release status, and score. This is the audit trail. Missing or incomplete metadata = blocked at the storage layer.

{
  "source": "midjourney",
  "license": "owned",
  "prompt_or_camera": "abstract gradient backdrop, FH brand palette purple to blue, soft grain, --ar 16:9 --v 8",
  "intended_channel": ["site-hero", "og"],
  "human_subject_release": "n/a",
  "attribution_required": null,
  "scored_at": "2026-05-10T14:32:00Z",
  "score": {
    "overall": 8,
    "brand_fit": 9,
    "slop_risk": 2,
    "license_clear": true
  }
}

#The automated curation gate

When a file lands in an FH inbox prefix, an automated flow runs: image evaluator (visual quality rubric), brand-fit checker (against brand tokens), slop detector (checks for AI tells), license auditor (verifies metadata). Anything that fails any check gets quarantined and surfaced for human review. Curation is enforced at write time, not after the fact.

#Prompts that work

Specific over generic. ‘Modern workshop with natural light through tall windows, woodworking tools on a clean bench, no people, shot on a Sony A7R with a 35mm lens, color graded warm’ beats ‘carpentry workshop.’ Reference photos in the prompt help. Style descriptors (‘shot on’, ‘color graded’, ‘depth of field’) pull the model toward photographic output. Specify aspect ratio and resolution explicitly.

#Licensing posture

Midjourney’s commercial-use license is clear if you’re on a paid plan. Stable Diffusion outputs are CC0 in most cases but model-dependent — verify the specific model’s license. DALL-E’s commercial use is allowed with their TOS. For client work we always use the paid commercial tier and document the source in meta.json. Never use ‘free trial’ outputs commercially without paying — the license usually requires it.

#When to commission real photography

Whenever the image will appear above the fold on the homepage. Whenever the image features your team. Whenever authenticity is the brand’s positioning. Real photography is more expensive up front and the only way to actually win on those use cases. FH client work uses commissioned photography for hero images on every site.

#How this lands across FH client work

Every image headed for FH storage runs through the curation rubric. AI-generated imagery is allowed for backgrounds, mood, abstract concepts, and product mocks. Real photography for everything else. The rubric blocks ~30% of submitted AI imagery — usually slop risk or stock-cliche pattern. If you’re using AI imagery on your site and not sure whether it crosses the slop line, book a consultation — we’ll run your asset library through the rubric and surface the candidates that should be replaced.