AI Image Prompt Structure: Subject, Setting, Light, Lens

By the end of this post you'll be able to take a vague idea ("a coffee ad") and turn it into a four-part prompt that tells an image model what to draw, where, under what light and through what lens. You'll also know how to fix a weak result by changing one part at a time instead of rewriting everything.

The ai image prompt structure below isn't a magic formula. It's a checklist that matches how the model makers themselves describe good prompts.

What the model makers actually recommend

Three official guides are worth reading, and they agree more than you'd expect.

Google Cloud's Ultimate Nano Banana prompting guide gives a core formula: subject, action, location or context, composition, style. It also tells you to design the lighting explicitly ("Tell the model exactly how the scene is illuminated") and to use photographic and cinematic terms for the camera.

The fal.ai FLUX.2 prompting guide uses a similar order: subject, action, style, context. It says the model weighs earlier information more heavily, so the things you can't live without should come first.

Black Forest Labs' own prompting summary describes the job as covering "the subject, setting, style, lighting, and composition."

Different wording, same skeleton. Here's how I'd compress it into four slots.

The four slots

1. Subject: who or what, in concrete terms

This is the slot most people under-write. "A woman with a coffee" leaves the model to invent everything. Compare:

A woman in her thirties with short dark hair, wearing a mustard knit sweater, holding a white ceramic mug with both hands

Name the age range, clothing, colours, materials and what the hands are doing. If there's a product, say what it is and what it's made of. If you need the same person across many images, a text description won't hold exactly. That's what saved characters are for (more on that below).

Put the subject first. Per the FLUX.2 guide, earlier words carry more weight.

2. Setting: where, when and what's around it

Setting covers location, time of day, background and props.

...sitting at a small window table in a quiet corner café, rain on the glass behind her, a half-eaten croissant on a plate

Keep it to a few objects that matter. A long list of background items tends to crowd the subject, and some will get dropped or merged anyway.

3. Light: the slot that changes mood the most

Light does more for the feel of an image than any adjective like "beautiful" or "stunning". Be specific about direction, softness and colour. Google's guide offers examples such as a "three-point softbox setup" for evenly lit products, "chiaroscuro lighting with harsh, high contrast", and "golden hour backlighting creating long shadows". The fal.ai guide adds "soft diffused studio lighting from above" and "hard directional sunlight."

Practical pairings:

• Product on white: soft diffused light from above, gentle shadow under the object.
• Portrait with mood: window light from the left, deep shadow on the other side of the face.
• Outdoor lifestyle: golden hour, low sun behind the subject, warm rim light on hair.
• Dramatic hero shot: hard directional light, high contrast, dark background.

4. Lens: how the picture is framed

Lens and camera terms control perspective, background blur and crop. The guides suggest:

• Wide-angle (roughly 14 to 24mm): big scenes, dramatic perspective. The fal.ai guide lists 14-24mm as wide and dramatic.
• Standard (35 to 50mm): natural-looking, documentary feel.
• Portrait (70 to 85mm): flattering faces, compressed background.
• Aperture: f/1.4 to f/2.8 for a blurry background, f/8 to f/16 for everything sharp.
• Macro lens: small details like texture and stitching, per Google's guide.

The FLUX.2 guide makes the point that a prompt ending "shot on 85mm lens at f/2.0" gives a markedly different result than "24mm lens at f/11." Try both on the same scene and see for yourself.

Putting it together

Here's the full prompt, written as one flowing description rather than a pile of keywords:

A woman in her thirties with short dark hair, wearing a mustard knit sweater, holding a white ceramic mug with both hands. She sits at a small window table in a quiet corner café, rain on the glass behind her, a croissant on a plate. Soft window light from the left, warm tones, gentle shadow on the right side of her face. Shot on an 85mm lens at f/2, shallow depth of field, eye-level, medium close-up.

Notice what's missing: no "masterpiece, 8k, ultra-detailed" string. Google's guide specifically favours describing a scene over stacking keywords, and good photographic terms do more work than quality buzzwords.

Add a style line only if you need one

Style is the optional fifth part. "Fashion magazine editorial, shot on medium-format analog film, pronounced grain" (from Google's example) pushes the whole look in a direction. If you don't care, leave it out and let light and lens carry the image. Stacking several conflicting styles usually produces mush.

How to iterate without starting over

Most weak results come from one weak slot. Change one thing per run so you can see what did what.

1. Run the four-slot prompt as written.
2. Look at what's wrong: wrong person, wrong place, flat mood, or awkward framing.
3. Edit only that slot. Flat mood? Rewrite the light. Cluttered background? Shorten the setting or open the aperture.
4. Run again and compare.
5. Save the version you like.

Results vary from run to run even with the same prompt, so generate a few before you decide a change worked or failed.

Common mistakes

• Negatives instead of positives. Google's guide advises describing what you want ("empty street") rather than what you don't ("no cars"). Some models handle negation badly. BFL's summary also mentions working without negative prompts, which suggests you shouldn't rely on them.
• Vague light. "Good lighting" tells the model nothing. Name a direction and a quality.
• Contradictions. "Harsh midday sun" plus "soft moody dusk glow" makes the model pick one at random.
• Buried subject. If the key requirement is at the end of a long prompt, it may get less attention.
• Expecting lens terms to work like a real camera. They're a strong hint about the look, not a physics simulation. Exact framing and blur vary by model.
• Long text in the image. Models differ a lot in how well they render words. If text matters, test a short phrase first and pick a model known for typography.

Using this in Mediafy Studio

In the Studio, pick Image, choose a model and paste the prompt. If you're starting from a short idea, the Enhance button rewrites it into a fuller prompt, which is a handy way to see how the four slots get filled in. Then edit the light and lens lines yourself, since those are the details you'll care about most.

The credit cost shows on the Generate button before you run it, so you can draft on a cheaper model such as Seedream 5 Pro (2 credits) to check composition, then switch to a pricier model for the final. Different models read the same prompt differently, so it's worth running your best prompt on two of them. Save keepers in My Prompts. If you want the same face across a series, create a character and type @ in the Studio to bring it in. For a walkthrough of the basics, see Your First Image in 5 Minutes.

If you want structured practice beyond one post, the AI Creator Lab course covers it in more depth.

A reusable template

Copy this and fill in the brackets:

[Subject: age, clothing, colours, materials, pose]. [Setting: place, time, 2-3 props]. [Light: direction, quality, colour]. [Lens: focal length, aperture, angle, crop]. [Optional style line].

Try it on your next idea at mediafystudio.com/signin.

Sources

• Google Cloud Blog, "Ultimate prompting guide for Nano Banana": https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-nano-banana
• fal.ai Learn, "FLUX.2 Prompt Guide": https://fal.ai/learn/devs/flux-2-prompt-guide
• Black Forest Labs Docs, "Image Prompting Summary" (FLUX Prompt Framework): https://docs.eu.bfl.ai/guides/prompting_summary