Image to Prompt: How to Generate an AI Prompt from Any Photo
You save a reference photo to your moodboard. The lighting is perfect — soft, directional, slightly warm. The camera angle is exactly what you want. The skin texture reads as genuinely candid, not retouched.
Then you open Nano Banana and stare at a blank text box.
You type "similar vibe." You get a completely different image. You try again with more words. Still wrong. You spend 40 minutes guessing keywords that might describe what you are looking at.
This bottleneck kills more AI workflows than any technical limitation. The problem is not Nano Banana — it is the translation layer between what your eye sees and what a language model can process. This guide solves that translation problem completely.
Quick answer: To convert a photo into an AI prompt, decompose it into 8 layers — subject, action, outfit, environment, camera/lens, lighting, texture, and negative constraints. Each layer needs specific technical vocabulary, not general adjectives. This takes 15–25 minutes manually, or under 10 seconds with CookedBanana.
Why Manual Description Always Falls Short
Human perception is holistic. When you look at a photograph, you instantly absorb hundreds of micro-signals simultaneously — the direction the light is coming from, the lens compression ratio, the color temperature, the skin texture, the depth of field fall-off, the grain structure. You experience all of it at once.
When you try to convert that experience into text, you lose most of it. You write "nice lighting" instead of "soft diffused window light from upper-left at approximately 45 degrees, warm 4800K color temperature, no fill light, natural shadow across right side of face." You write "realistic" instead of the eight specific texture terms that actually produce realistic output.
The gap between what you see and what you can describe is where AI generations go wrong.
What Reverse-Engineering an Image Actually Extracts
A proper image-to-prompt reverse engineering process does not just describe the photo in general terms. It decomposes the image into the exact parameters Nano Banana needs to replicate it:
1. Subject & Identity Layer Who or what is in the image. Physical descriptors, age markers, asymmetries, distinguishing features. Not "a woman" but "late 20s woman, slight dark circles, natural asymmetrical face, no heavy makeup."
2. Action & Narrative What is happening and why. The implied motion, the emotional state, the social context. "Glancing down at a phone, mid-step, slightly distracted — candid street capture."
3. Outfit & Styling Exact garment descriptions, fabric types, fit, layering, accessories. Every undefined item gets invented by the model.
4. Environment & Background Setting details that anchor the scene. Architecture, surface materials, ambient objects, spatial depth.
5. Camera & Lens Simulation The focal length, aperture, and body that would have taken this shot. This single parameter determines how Nano Banana renders depth, compression, and micro-detail. See the complete camera reference in our 8-part prompt guide.
6. Lighting Architecture Source direction, quality (hard vs. soft), color temperature, shadow behavior, secondary fills.
7. Texture & Realism Modifiers The micro-details that separate a photograph from a render: skin pores, fabric grain, environmental clutter, film grain, chromatic aberration.
8. Negative Constraints What the model is forbidden to invent. The visual attributes that must not appear in the output.
This is exactly the 8-macro-reference structure that Nano Banana responds to best — and the one CookedBanana is built to output automatically.
The Manual Process: Step-by-Step
If you want to reverse-engineer an image manually, follow this process in order:
- Cover the image and write what you remember. Your first description reveals what your brain prioritized. These are the most salient elements.
- Uncover and audit layer by layer. Go through each of the 8 layers above and note what the image actually contains. Most people miss layers 5, 6, and 7 entirely.
- Translate photography instinct into vocabulary. "The light feels morning" becomes "soft diffused daylight, overcast sky, slight cool cast, even shadows." "The lens feels close" becomes "35mm, moderate depth of field, slight barrel distortion."
- Build the negative constraints. Look at what is not in the image. If the skin is textured and imperfect, add
no smooth skin filter. If the composition is candid, addno studio pose. - Assemble and test. Write the full prompt, run it, compare outputs, adjust the layers that diverged most from the reference.
A complete manual reverse-engineering session takes 15–25 minutes per image. For a single personal project, that is manageable. For an agency managing weekly content production across multiple clients, it is a bottleneck.
A Real Reverse-Engineered Prompt
Reference image: a candid street portrait, late afternoon, subject mid-walk, slightly underexposed, film look.
What most people write:
Candid street portrait, afternoon, film style, realistic.
What a proper reverse-engineering produces:
Late 20s woman, natural asymmetrical face, slight squint, dark hair loose — mid-stride on a city sidewalk, slightly looking right, weight on left foot — oversized dark wool coat, slightly worn collar, black jeans — urban street, blurred pedestrians in background, warm low light bouncing off building facades — 35mm f/2, slight vignette, modest barrel distortion — late afternoon sun from right side, warm 5200K, strong directional shadow left, no fill — visible skin pores, slight sheen on nose bridge, film grain ISO 1600, slight chromatic aberration at edges — no smooth skin, no posed stance, no HDR, no symmetrical face, no digital glow.
The gap in output quality between these two prompts is categorical. The second one leaves Nano Banana almost no interpretive room — which is exactly what consistent, replicable generation requires. If the skin texture still looks off after reverse-engineering, see our guide on fixing plastic skin in AI portraits.
How CookedBanana Compresses This to 10 Seconds
Generic image-to-text tools — CLIP interrogators, basic vision describers — return a raw paragraph of what is visible in the image. That paragraph is not a Nano Banana prompt. It requires significant manual restructuring before it becomes usable.
CookedBanana is purpose-built for this specific output format. Upload your reference image and the engine does not just describe it — it outputs a fully structured 8-layer prompt, pre-formatted for Nano Banana's architecture, with:
- Every layer correctly separated and labelled
- Negative constraints automatically generated from what is absent
- Camera simulation extracted from the optical characteristics of the image
- Texture modifiers inferred from the skin and surface rendering
If you want to keep a specific jacket from the reference but change everything else, activate Outfit Refs before generating. CookedBanana isolates the garment into its own reference layer (image_ref.2) so Nano Banana treats it as a locked asset while treating the rest of the scene as variable.
If you want to preserve the subject's identity — face, bone structure, distinguishing features — activate Lock Ref. The subject becomes image_ref.1: a frozen anchor the generation cannot modify.
Start your free trial — 3 generations, no credit card required.
Frequently Asked Questions
Can Nano Banana directly replicate a reference photo?
Nano Banana can come very close to replicating the style, mood, and technical characteristics of a reference photo — but it cannot reproduce an exact copy. What it excels at is maintaining the aesthetic DNA of the reference while applying it to new subjects, environments, or scenarios. The closer your prompt matches the actual parameters of the reference (lighting, lens, texture), the closer the output will be.
What is the difference between a basic image-to-text tool and CookedBanana?
Basic image-to-text tools (CLIP interrogators, generic vision models) return a raw description of what they see in the image. This description is not structured for AI image generation — it is a general paragraph that requires significant manual work to become a usable prompt. CookedBanana outputs a structured 8-layer prompt specifically formatted for Nano Banana, with negative constraints, camera simulation, and texture modifiers already included.
How many reference images can I upload to Nano Banana at once?
Nano Banana Pro supports up to 14 reference images in a single prompt (6 with high fidelity). This means you can reference a subject from one image, an outfit from another, a background from a third, and a lighting setup from a fourth — all in one generation. CookedBanana's Outfit Refs and Lock Ref features are designed to work within this multi-reference architecture, assigning each locked element its correct image_ref.N index automatically.
What types of reference photos work best for reverse-engineering?
High-contrast, well-lit images with a clear single subject produce the most accurate reverse-engineering results. Blurry, heavily post-processed, or low-resolution references make it difficult to extract accurate camera and texture parameters. For lifestyle photography references, candid shots with visible environmental context (background details, natural lighting direction) give the most useful source material.
Can I use this workflow for video content, not just static images?
The reverse-engineering workflow applies directly to still frames extracted from video. For a consistent visual style across a video production, extract the most representative frame from a reference clip, run it through CookedBanana, and use the output as your master prompt template. Every generated image in the series will share the same aesthetic fingerprint as the original frame.
How do I describe lighting direction in an AI image prompt?
Lighting direction is described by its source position relative to the subject. Common formats: soft window light from the left, overhead daylight with soft fill, 45-degree raking sidelight from lower-right. Always specify the quality (hard vs. soft), color temperature (warm 5200K / cool 5500K+), and whether there is a secondary fill light. Missing these parameters causes the model to default to flat, uniform lighting.
What prompt structure captures a photo's color grade?
Color grading is captured through color temperature descriptors (warm 5200K, cool overcast light), film stock references (shot on Kodak Portra 400), and processing style tokens (unedited RAW export, no color grade, slight desaturation in highlights). For a specific grade, describe the dominant shadow color and highlight color separately. For the full framework, see the 8-part prompt formula.
Topics