How to Reverse-Engineer an AI Image Prompt from Any Reference Photo

You saved a reference — a Pinterest board pin, a Midjourney render, a client mood board, a film still. You know what you want. You don't know how to say it in words the image model understands.

Reverse-engineering a prompt means: look at the result → describe what produced it → get text you can paste, edit, and reuse. You're not recovering the literal original prompt (that's usually impossible). You're building a visually equivalent description — often 80–95% of the look on the first pass.

This guide covers the manual breakdown framework, vision-LLM workflow, model-specific syntax, and when dedicated image-to-prompt tools like PromptMake /image beat doing it by hand.

What reverse-engineering can and cannot do

Can:

• Capture subject, composition, lighting, color palette, lens feel, medium/style
• Translate a look across models (MJ → FLUX → DALL·E dialect)
• Give you editable language for client revisions ("more like this reference")
• Speed up learning — you see which words map to which visual choices

Cannot:

• Recover the exact seed, hidden parameters, or private prompt from someone else's render
• Guarantee pixel-perfect recreation (generators re-roll noise every time)
• Replace style-reference images where the model supports --sref or img2img

Treat output as a strong starting prompt, not a forensic copy.

The 7-layer breakdown (manual method)

Before any tool, train your eye to scan in this order. Same layers vision models should extract:

1. Subject — who/what, pose, expression, clothing, key props
2. Environment — location, time of day, weather, background depth
3. Composition — framing (close-up, wide), angle, rule-of-thirds, negative space
4. Lighting — direction (side, back, rim), quality (soft/hard), color temperature
5. Color & grade — palette (teal-orange, muted pastels), contrast, film stock feel
6. Medium & style — photograph vs illustration, era (80s editorial, brutalist 3D, anime)
7. Technical cues — lens (85mm portrait), aperture bokeh, grain, aspect ratio intent

Write one sentence per layer, then merge into a single prompt. Cut redundancy.

Example merge (photo reference):

"Portrait of a woman in her 30s, three-quarter view, soft window light from camera left, shallow depth of field, warm neutral palette, shot on 85mm f/1.4, natural skin texture, minimalist beige interior background, editorial fashion photography."

For Midjourney add parameters separately: --ar 2:3 --style raw. For FLUX keep natural language. For SDXL split positive/negative.

Method 1: Vision LLM analysis (ChatGPT, Claude, Gemini)

Upload the image to a multimodal model. Use a structured instruction — not "describe this image."

Template:

Analyze this reference for text-to-image recreation. Output ONE prompt only. Include: subject, environment, composition, lighting, color palette, medium/style, camera/lens feel. Format for [Midjourney v7 / FLUX / DALL·E 3]. Do not include commentary.

Optional add-ons:

• "Prioritize lighting and lens over generic adjectives."
• "Max 120 words."
• "End with suggested --ar for Midjourney."

GPT-5.5 / ChatGPT Image — strong at natural-language photographic prompts; good for DALL·E and ChatGPT Image targets.

Claude Opus 4.8 — careful composition and mood wording; useful for editorial and product scenes.

Gemini 3.1 Pro — solid vision analysis; pair with Gemini image generation or export to FLUX.

Weakness: generic vision prompts produce generic captions ("a beautiful sunset"). Force specificity in your template.

Method 2: Midjourney /describe (MJ-native only)

In Discord: /describe + upload image. Midjourney returns four prompt suggestions trained on its own corpus.

Pros: syntax matches MJ behavior; fast for MJ-to-MJ iteration.

Cons: Discord-only; vague on lighting sometimes; not portable to FLUX/SD without reformatting.

Workflow: /describe → pick closest option → edit → add --style raw, --ar, --cref if character consistency matters.

For non-MJ targets, use describe as a draft, then run through model-specific rewriting (Method 4).

Method 3: CLIP interrogators (Stable Diffusion tradition)

Tools like CLIP Interrogator rank tags against a vocabulary — great for SDXL/ComfyUI pipelines, anime tags (Danbooru style), texture-heavy work.

Output looks like: (masterpiece:1.2), 1girl, rim lighting, film grain, ...

Pros: dense keyword coverage; community vocab for SD.

Cons: reads poorly in Midjourney v7 natural-language mode; needs cleanup for FLUX prose.

Use when: SD/ComfyUI is your target and you want weighted tags, not sentences.

Method 4: Dedicated image-to-prompt tools

Purpose-built tools run vision models with model-specific system prompts — same image, different output dialect per target.

Why they beat raw ChatGPT for this task:

• Midjourney v7 gets comma phrases + --ar, not essay prose
• FLUX gets natural language without MJ parameter noise
• DALL·E gets conversational scene description
• SDXL can include negative prompt suggestions

PromptMake /image workflow:

1. Upload JPG, PNG, or WEBP (up to 10 MB)
2. Pick target: Midjourney v7, FLUX.1, DALL·E 3, Stable Diffusion, Ideogram, etc.
3. Choose goal mode: Recreate Exactly, Change Style, Adjust Lighting, or Create Variation
4. Optional notes: "more cinematic," "remove text," "emphasize product"
5. Copy prompt → paste into your generator → iterate

Free tier: 3 image generations per day (guest and registered — separate pool from /text). No credit card.

Pro unlocks fine-tune panel: stylization, negative magic, output format options.

Difference from legacy image-to-prompt-guide on this blog: this article is the deep workflow; that post is the quick intro. Use both.

Goal modes: same photo, different intent

Recreate Exactly — maximal fidelity to reference. Use when you want the closest match.

Change Style — same composition/subject, new aesthetic (photo → illustration, modern → vintage).

Adjust Lighting — same scene, different light (golden hour → overcast, studio → neon).

Create Variation — inspired by reference with creative freedom; not a clone task.

Pick mode before generating. The vision instruction changes what the tool optimizes for.

Model-specific rewrite cheat sheet

Same visual idea, different syntax:

Midjourney v7 — concise phrases, subject first, parameters last: `--ar 16:9 --style raw --stylize 200`. Character work: `--cref [url]`. Style lock: `--sref [url]`.

FLUX.1 — flowing natural language; weight emphasis in parentheses if supported; avoid MJ flags.

DALL·E 3 / ChatGPT Image — conversational scene description; specify what to avoid inline ("no text overlay").

Stable Diffusion XL — positive prompt + negative prompt block; `(token:1.1)` weights; LoRA names if applicable.

Ideogram — explicit typography if text in scene; design-forward wording.

Never paste a Midjourney prompt into FLUX unchanged. Cross-model translation is half the job.

Step-by-step workflow (production)

A. Reference is a photograph (real world)

1. Crop to subject; remove watermarks
2. Run image-to-prompt with Recreate mode, target = your generator
3. Generate v1 in target model
4. Compare side-by-side: what's wrong? (lighting, lens, palette)
5. Edit prompt — add ONE layer at a time (don't rewrite everything)
6. Optional: use reference as --sref or img2img strength 0.3–0.5 while keeping text prompt

B. Reference is AI-generated (someone else's render)

1. Identify generator if possible (MJ texture vs FLUX clarity vs DALL·E simplicity)
2. Start with that target in image-to-prompt
3. Strip copied artist names / living artist triggers if policy-sensitive
4. Variation mode if you want inspired-by, not clone

C. Client says "make it like this"

1. Image-to-prompt the client reference
2. Extract 5–10 distinctive tokens (lighting, palette, framing)
3. Merge into your existing brand prompt — don't replace wholesale
4. Document approved prompt in DAM/Notion for reuse

Common failures and fixes

Output too generic — your instruction lacked layer specificity. Re-run with the 7-layer checklist explicitly.

Wrong medium — model thinks illustration is photo. State "35mm photograph" or "3D render" explicitly.

Lighting flat — add direction + quality: "hard side light, deep shadows, rim light on hair."

Composition drift — specify "centered subject, upper third horizon, 85mm compression."

MJ ignores details — shorten prompt; MJ favors front-loaded nouns; move vibe words after subject.

FLUX overwrites with prose fluff — trim adjectives; FLUX likes precise scene grammar.

Text garbled in image — Ideogram/DALL·E need explicit quoted text; other models: "no text" in negative.

Manual vs tool: when to use which

| Situation | Best approach |

| Quick one-off, learning | Vision LLM + 7-layer template |

| MJ-only workflow | /describe + hand edit |

| SD/ComfyUI tag soup | CLIP Interrogator |

| Multi-model export, client work | PromptMake /image or similar |

| Maximum style lock | --sref / img2img + short prompt |

| Privacy-sensitive refs | Local tool or self-hosted vision, not public Discord |

Ethics and policy

• Don't reverse-engineer to impersonate living artists by name — use style descriptors ("1970s editorial colour photography" not "in the style of [living artist]").
• Client references may be copyrighted — prompts you extract are for licensed work contexts.
• --cref on real people without consent is a policy violation on most platforms.
• Disclose AI-generated assets where your industry requires it.

Iterate like an engineer, not a gambler

Log: reference thumbnail → extracted prompt → generator settings → output v1/v2/v3.

Change one variable per iteration (lighting words OR lens OR palette — not all three).

When stuck, switch mode: Recreate → Variation, or swap target model to see which dialect captures the reference best.

Frequently asked questions

Can I get the exact original prompt from an AI image?

Almost never. You approximate the visual recipe.

Does reverse-engineering work on low-res images?

Poorly. Vision models miss detail; upscaling helps slightly. Use the highest clean source available.

Best format for upload?

PNG or JPG, well-lit, subject unobstructed. Crop clutter.

Midjourney /describe vs PromptMake?

Describe is MJ-native and Discord-bound. PromptMake supports multiple targets and goal modes in browser — 3 free/day on /image.

Can I reverse-engineer video frames?

Same workflow on a key frame; add motion language separately for Sora/Runway (see video guides on the blog).

How close is "Recreate Exactly"?

Close in mood, composition, and lighting — not identical pixels. Expect 2–5 edit rounds.

Related articles

Image to Prompt: Complete Guide (legacy) — quick intro to goal modes.

Midjourney --cref Guide — character consistency after you have a base prompt.

Prompt engineering for AI images: 7 elements (cluster 3 — coming next).

DALL·E 3 vs Midjourney vs Flux — which model needs better prompts.

Bottom line

Reverse-engineering a prompt is structured looking: subject, light, lens, palette, medium — then translation into the target model's dialect.

Do it manually to learn. Use vision LLMs for speed. Use image-to-prompt tools when you need model-calibrated output and goal modes without writing system prompts yourself.

Upload the reference on promptmake.net/image, pick your target and mode, copy the result, iterate once in your generator. That's the loop most creators actually run in 2026.