A Schema for Turning Visual Evidence into Reusable Image Prompts

Reverse-engineering an image into a useful prompt is not the same as writing a caption. A caption tells a reader what the picture is about. A reconstruction prompt must preserve the visual decisions that made the picture look the way it does.

That means the process needs a schema. Without one, observers usually over-describe the subject and under-describe composition, light, optics, materials, and constraints.

The evidence-first rule

Begin with what is visible, not with a story about how the image was made.

For every observation, assign one of three confidence levels:

• Observed: directly visible in the pixels.
• Inferred: likely, but not certain.
• Unknown: cannot be established from the image.

For example, “warm light from camera left” is an observation. “Shot with a specific brand of softbox” is an inference. Good prompt reconstruction keeps those categories separate.

A seven-layer visual schema

1. Frame and composition

Record the structural choices first:

• aspect ratio,
• subject position,
• crop,
• horizon or vanishing point,
• foreground, midground, and background layers,
• empty space,
• and dominant visual direction.

Composition has more influence than many stylistic adjectives. “Centered portrait” and “subject on the right third with negative space on the left” can produce completely different images even when every other phrase is identical.

2. Subject identity

Describe only visual identity that matters to generation:

• object or person type,
• approximate age range when relevant,
• silhouette,
• pose,
• clothing or surface details,
• expression,
• and distinctive features.

Avoid assigning a real identity unless the image clearly requires it. The goal is visual reproducibility, not speculation.

3. Environment and scene state

A prompt should distinguish the location from its current state.

“City street” is a location. “Wet city street at blue hour with sparse traffic, reflected signage, and light fog” is scene state.

Useful environment fields include:

Field	Example
Location	narrow urban alley
Time	after sunset
Weather	recent rain
Background density	sparse
Atmospheric depth	light haze
Surface state	wet pavement

4. Lighting

Lighting should be described as relationships:

• key-light direction,
• hardness or softness,
• fill level,
• rim or backlight,
• contrast,
• color temperature,
• and practical light sources visible in frame.

Instead of “dramatic lighting,” write “hard cyan rim light from behind, soft magenta key from camera left, deep shadow fill.” The second version gives a model something it can act on.

5. Camera and optics

You may not know the exact lens, but you can describe the visual effect:

• wide, normal, or telephoto feeling,
• camera height,
• viewing angle,
• depth of field,
• focus plane,
• perspective compression,
• and motion blur.

Mark an exact focal length as inferred unless metadata is available. “Telephoto compression with shallow depth of field” is often safer than pretending to know a specific millimeter value.

6. Material and texture

List surfaces and how they respond to light:

• brushed metal,
• translucent plastic,
• matte fabric,
• glossy ceramic,
• skin texture,
• film grain,
• dust,
• condensation,
• or micro-scratches.

Material words are especially important in product and interior images, where the same geometry can look cheap or premium depending on surface behavior.

7. Exclusions and failure controls

A reconstruction prompt should say what must not drift.

Examples:

• no additional people,
• no text or watermark,
• preserve the asymmetric crop,
• keep the product label readable,
• avoid over-smoothing skin,
• no extra props,
• and do not change the dominant light direction.

These controls are not a generic negative-prompt dump. They should respond to risks in this specific image.

Convert the schema into a prompt stack

Assemble the final prompt in a stable order:

1. format and composition,
2. main subject,
3. action or pose,
4. environment,
5. lighting,
6. camera and focus,
7. materials and texture,
8. mood or finish,
9. exclusions.

This order makes revisions easier. If composition is wrong, change the first block instead of rewriting the whole prompt.

A tool such as ImageToPrompt can generate a first-pass prompt from an uploaded JPG, PNG, or WebP and tune the result for Flux, Midjourney, Stable Diffusion, or Nano Banana. Treat that first pass as structured draft material, then compare it against the seven layers above.

Score coverage before testing

Use a simple coverage matrix:

Layer	Present?	Specific enough?	Needs correction?
Composition	Yes	Partial	Add negative space direction
Subject	Yes	Yes	No
Environment	Yes	Partial	Add wet surface state
Lighting	Yes	No	Add key direction and fill
Camera	No	—	Add eye-level telephoto feeling
Material	Yes	Partial	Add brushed-metal response
Exclusions	No	—	Add no extra props

A prompt that covers every layer is not automatically good, but missing layers create predictable drift.

Test one variable at a time

Run the first generation as a diagnostic. Compare it with the source image in this order:

1. overall composition,
2. subject silhouette and pose,
3. lighting direction,
4. color relationships,
5. camera perspective,
6. material response,
7. fine detail.

Fix the first major mismatch before polishing small details. If the subject is framed incorrectly, adding more texture adjectives wastes iterations.

Preserve an audit trail

Save four things with every experiment:

• source-image identifier,
• prompt version,
• changed field,
• and observed effect.

A useful note might be: “Version 3 changed only camera from normal perspective to telephoto compression; background scale now matches, but subject crop is too tight.”

That record turns prompt iteration into a repeatable method rather than a sequence of guesses.

The practical standard

A reconstructed prompt is successful when another person can read it, identify which visual decisions are fixed, and change one layer without breaking the rest. The best result is not the longest description. It is the smallest structured description that preserves the image’s important evidence.