Why order still beats adjectives
Modern image models parse language in sequence. Early tokens tend to anchor the subject, wardrobe, and key materials; later tokens refine environment, lighting, and camera. That is why scattershot lists of buzzwords underperform compared with a short, ordered stack of visual facts (concrete nouns, measurable light, real lens cues) rather than praise like “epic” or “masterpiece.” Community guides such as fal.ai’s GPT Image 2 prompting notes emphasize swapping hype for inspectable detail and ending with constraints.
A six-line template you can reuse
- Subject: who or what, scale, pose, wardrobe or materials.
- Spatial relation: foreground vs background, distance, interaction.
- Setting: interior or exterior, time of day, weather if it matters.
- Light: source direction, quality (hard, soft, bounced), color temperature.
- Camera: focal length, height, depth of field, grain or cleanliness.
- Medium + guardrails: editorial photo, gouache, UI mock; then “no watermark,” “no extra logos,” etc.
Example: product still life
Single glass bottle of cold brew on slate, condensation beads, label facing camera. Bottle centered lower third; negative space above for headline crop. Overcast daylight through north window, soft shadows, neutral white balance. 35mm product lens, f/5.6, crisp label text, shallow background falloff. Clean studio tabletop, minimal props, editorial beverage photography. No watermark, no brand logos except the bottle label, no people.
Style without style-tag soup
Instead of stacking genre tags, describe what a viewer would actually see: paper texture, ink bleed, corner radius on cards, rim light on metal. If you need a film look, name stock or halation behavior (“subtle halation on highlights”) rather than “cinematic” alone.
Related guides on this site
For quoted lettering and composition control, read GPT Image 2: Text, Typography & Negative Prompts. For paste-ready blocks across genres, see GPT Image 2 Ready Prompt Examples.