One of the most powerful workflows in AI image generation is recreating an existing image — a photo you love, an artwork that inspires you, or an AI image you want to remix. But you can't just feed an image into most generators and say "make this."

You need a prompt. And extracting an accurate prompt from a reference image is a skill that unlocks enormous creative possibilities.

This guide covers the complete workflow for recreating any image with AI, from extracting the prompt to generating and refining the result.

Why recreating images matters

There are several legitimate creative reasons to recreate an image:

Style transfer — you love the lighting and mood of a photograph and want to apply it to a different subject. Extracting the prompt captures the style elements separately from the content.

Learning — understanding what prompt produced a given image teaches you how the generator thinks. It's the fastest way to build intuition for prompt writing.

Iteration — you have an AI-generated image you like but want to vary the subject, change the setting, or try it in a different generator. The extracted prompt is your starting point.

Reference matching — a client shows you a mood board image and wants you to create something similar. Extracting the prompt gives you a technical foundation to work from.

Step 1: Analyse the reference image

Before extracting a prompt, understand what makes the image work. Look for:

Lighting — where is the light coming from? What quality is it (soft, harsh, directional, diffused)? What color temperature (warm golden, cool blue, neutral)?

Composition — how is the subject framed? What's the perspective (eye level, low angle, bird's eye)? Is there depth or is it flat?

Style — does it look photographic, illustrated, painterly, cinematic? Is there a specific aesthetic era or genre?

Color palette — what are the dominant colors? Is there a specific grade or tone (muted, saturated, faded, contrasty)?

Subject — who or what is the main focus, and how is it described or posed?

Understanding these elements helps you verify and refine the extracted prompt.

Step 2: Extract the prompt

The fastest and most accurate way to extract a prompt is to use a dedicated tool. Upload your reference image to PixelPrompt and select the output mode that matches your target generator.

  • Midjourney mode — outputs prompt with correct --ar, --style, and --v parameters
  • Flux mode — outputs natural language description optimized for Flux
  • Stable Diffusion mode — outputs keyword-based prompt with quality tags
  • General mode — works for any generator
  • Structured mode — breaks the analysis into labeled components (subject, lighting, style, etc.) — useful for understanding and manual editing

The Structured mode is especially valuable for learning — it shows you exactly what the AI identified in each category, so you can see which elements it prioritized.

Step 3: Verify and refine the prompt

The extracted prompt is a starting point, not a final answer. Compare it against your visual analysis from Step 1:

  • Does the lighting description match what you see?
  • Is the style accurately captured?
  • Are important composition elements mentioned?
  • For Midjourney, are the parameters correct (aspect ratio, version)?

Make adjustments based on what you notice. Add missing elements. Remove anything that doesn't match the reference. For Stable Diffusion, consider what should go in the negative prompt.

Step 4: Generate and compare

Generate your first image from the extracted prompt. Place it side-by-side with your reference and compare:

What's accurate? — note what the generator got right. This tells you which prompt elements are working.

What's different? — identify specific differences. Is the lighting wrong? Is the style too different? Is the composition off?

What's impossible to replicate exactly? — AI generators are probabilistic. The same prompt produces different images every run. Some elements — a specific person's face, a very specific composition — can't be exactly reproduced through prompting alone.

Step 5: Iterate systematically

Change one variable at a time when iterating. If you change the lighting description, the style keyword, and the camera angle all at once, you won't know which change produced which effect.

Iteration order:

  1. Get the overall style right first (photographic vs painterly vs cinematic)
  2. Then lighting and atmosphere
  3. Then subject and composition
  4. Finally, technical details (film stock, grain, color grading)

For Midjourney, use the "Vary (Subtle)" option to generate variations that stay closer to a successful generation while exploring small changes.

For Stable Diffusion, adjust the CFG scale (how strictly the model follows the prompt) if results feel too literal or too free.

Practical example: Recreating a portrait

Let's walk through a concrete example. Say you want to recreate this type of image: a dramatic black and white portrait with strong Rembrandt lighting, deep shadows, and a sense of gravitas.

Extracted prompt (General mode):

A dramatic black and white portrait of an older man with deeply lined face,
strong Rembrandt lighting creating a triangular highlight on the cheek,
deep shadows on the opposite side, direct gaze into camera,
neutral dark background, fine art photography style,
sharp focus on eyes, medium format aesthetic

Midjourney version:

dramatic black and white portrait, older man with weathered face,
Rembrandt lighting, triangular cheek highlight, deep shadows,
direct gaze, fine art photography, medium format --ar 4:5 --v 6.1 --q 2

Stable Diffusion version:

(masterpiece:1.2), (best quality:1.1), black and white portrait photography,
elderly man, deeply lined face, Rembrandt lighting, triangle highlight,
dramatic shadows, direct eye contact, fine art, medium format, sharp focus

Negative: (worst quality:1.4), color, cartoon, smooth skin

What you can and can't recreate

You can recreate:

  • Lighting setups and moods
  • Art styles and aesthetics
  • Color grading and film stocks
  • Compositional approaches
  • Subject types and general descriptions
  • Atmospheric and environmental qualities

You can't exactly recreate:

  • Specific real people's faces
  • Exact compositions down to pixel level
  • A specific AI image's unique details (every generation is different)
  • Proprietary styles that are legally protected

The goal isn't pixel-perfect recreation. It's capturing the essence — the lighting, the mood, the style, the feeling — and using it as a foundation for your own work. The extracted prompt is a creative starting point, not a copy machine.

Speeding up the workflow

The full workflow — analyse, extract, verify, generate, iterate — takes time but produces much better results than guessing at prompts from scratch. The extraction step is where most time is saved.

Using a tool like PixelPrompt for extraction means the analysis and translation work happens in seconds. The time you save on writing prompts goes into the creative iteration that actually matters — getting the image right.

For ongoing projects where you're building a consistent style, save your best extracted prompts and use them as templates. Swap out specific details (subject, setting, time of day) while keeping the lighting and style elements that are working.