Image models have gotten good enough that a generated photo can carry a slideshow slide. They have also flooded feeds with a recognizable look: glossy, over-lit, slightly too perfect, with hands that do not quite work. Viewers have learned to spot it and to swipe past it. Here is how to use generated images so they add to a deck instead of marking it as automated.
When a generated photo earns its place
- The opening scene. A hook like "Nail appointment tomorrow, no clue what to get" wants a picture of that moment: a hand holding a phone at a kitchen table in evening light. Stock rarely has the exact moment; a model can make it.
- Body slides that describe a physical situation. A saved design on a phone next to bare nails. Running shoes by a front door. If the headline names objects and a place, a photo of those objects in that place helps.
- Backdrops behind cards. A darkened photo behind a comparison or step list makes the slide feel like it happened somewhere instead of on a template.
When it hurts
- Anywhere a real screen belongs. Never generate an app interface. Viewers notice invented buttons, and it misrepresents the product. Use a real capture or typography.
- People's faces. A generated face on every post is the fastest way to look like a bot farm. Prefer hands, objects and places.
- Results and testimonials. A generated "before and after" of a user's outcome is a fabricated claim, whatever the caption says.
- Slides that are already strong as type. A one-line statement in big letters on a lit field is a design choice, not a gap to fill.
Briefing for a phone-gallery look
The default output of every image model is a beautiful professional photograph, which is exactly the wrong register. Push against it in the brief:
- Name the camera: "amateur smartphone photo, handheld, available light."
- Name the imperfection: "slightly off-center framing, lived-in room, ordinary objects."
- Name what to avoid: "studio lighting, retouching, cinematic grading, stock-photo polish, watermark, text."
- Keep subjects in the central 70 percent of the frame so a 9:16 crop does not lose them.
- Ask for no text, no logos, no interface. Your renderer adds the words; a model spelling them will misspell them.
Keep the deck consistent
One anchor image, then edits from that reference for the other scenes, keeps the same room, palette and objects across the deck. Independent generations per slide drift: different kitchens, different hands, different light. A deck that drifts reads as assembled from a bag of pictures, which is the stock-photo problem all over again.
The honesty rules
- Generated imagery may set a scene. It may not depict the product, a result, or a person's experience.
- Do not label a generated photo as a customer photo, and do not imply it in the caption.
- If the platform asks you to flag AI-generated content, do it. TikTok's AI label does not hurt reach the way people fear; a viewer catching a fake does.
- Never generate a real person, a competitor's product or a recognizable brand.
Cost and restraint
A generated photo costs money and seconds every time. A deck does not need eight of them. One strong opening scene plus two or three body photos, with the rest as typography, cards and real screens, looks more like a person made it than a wall-to-wall photo carousel does.
ViralSlides follows exactly this pattern: the planner marks the beats a picture would carry, one storyboard keeps the scenes consistent, generation is capped per deck, and real screens are never synthesized. If you want to see the difference on your own app, generate a slideshow with the custom opening visual switched on.