Image to Prompt Goal Modes: Recreate, Restyle, Relight, Vary

Goal modes decide what your image to prompt tool optimizes for before it writes a single word. Recreate Exactly locks subject, composition, style, and lighting for closest-match prompts. Change Style keeps the bones of the shot and swaps medium or aesthetic. Adjust Lighting holds subject and style while rewriting light direction, quality, and color temperature. Create Variation treats the upload as mood and layout inspiration with room to invent.

This article shows when to pick each mode, how Midjourney v7, FLUX, DALL·E / GPT Image, SDXL, and Leonardo read the same photo under different goals, and how optional notes steer the result on tools like PromptMake /image.

What goal modes change in an image to prompt run

An image to prompt tool runs a vision model over your upload, then formats what it sees for a target generator. The photo stays the same. The instruction does not. Goal mode is that instruction: fidelity, restyle, relight, or free take.

Without a mode, you get a generic caption dressed up as a prompt. Captions list objects. Generation prompts choose what to protect and what to rewrite. Recreate protects almost everything. Change Style protects subject and framing while swapping paint, film, or render look. Adjust Lighting protects subject and medium while swapping key, fill, and Kelvin. Create Variation keeps mood and layout cues and drops the clone mandate.

Pick the mode first. Then pick the target model. Then add short notes if you need a specific twist. Order matters because notes layer on top of the mode. "Make it watercolor" under Recreate fights the fidelity goal. The same note under Change Style matches the job.

PromptMake /image exposes four labeled modes: Recreate Exactly, Change Style, Adjust Lighting, and Create Variation. Upload JPG, PNG, or WEBP up to 10 MB, choose Midjourney, FLUX, DALL·E, Stable Diffusion, or Leonardo as the target, and copy the prompt into your generator.

Recreate Exactly: when fidelity is the job

Recreate Exactly tells the vision stack to reproduce the source as close as text can get. Composition stays. Style stays. Lighting stays. New objects stay out unless the scene needs them for coherence. You use this when the reference is the brief itself.

Closest match means approximate. Seeds, private parameters, and noise rerolls mean no tool recovers a pixel clone from words alone. Treat the output as a strong first draft: subject identity, framing, palette, lens feel, and light direction packed into the dialect your generator wants. Midjourney v7 gets front-loaded nouns and trailing --ar / --style raw. FLUX gets flowing scene grammar. DALL·E / GPT Image gets a short paragraph. SDXL and Leonardo get denser tag-friendly phrasing.

Client work loves this mode. A brand sends one approved hero frame. You extract a reusable prompt, paste it into Midjourney v7 or FLUX, and iterate on small edits instead of rebuilding the look from memory. Product teams do the same with packaging shots: keep the bottle angle, label readable-ish intent, softbox direction, and white sweep language intact.

When to pick Recreate Exactly

• You need the closest text match to a photo or AI render you like
• A client said "make more like this" and pointed at one approved frame
• You are building a reusable identity block from a portrait or product hero
• You plan to pair the text with Midjourney --sref / --cref or img2img at low strength later

Skip Recreate when you want a different medium, a night version of a day scene, or a looser remix. Those jobs belong to the other three modes. Burning a Recreate run then hand-editing every style token wastes quota.

What Recreate output looks like across models

Midjourney v7: short stacked phrases, subject first, lighting and lens mid-prompt, parameters last. Expect something like portrait language plus --ar 2:3 --style raw --v 7. Recreate pushes literal photographic cues over vague "beautiful" fillers.

FLUX: one or two dense sentences. Textures, materials, and light quality get extra weight. No Midjourney flags. Parentheses emphasis appears when the formatter supports it.

DALL·E / GPT Image: conversational scene description. Soft negatives sit inline ("no watermark, no extra limbs") instead of a separate negative block.

SDXL / Leonardo: comma-heavy positives, quality boosters where useful, and room for a negative prompt if your pipeline uses one. Recreate leans on concrete tokens (85mm, softbox, marble surface) over vibe adjectives.

Change Style: same bones, new medium

Change Style keeps subject and composition while swapping artistic medium, era, or aesthetic. The person sits three-quarter left. The product stays on the same table plane. Paint, ink, 3D, anime, or 1970s print replaces the source medium.

This mode answers "what if this photo were an oil study?" or "what if this UI mock were a vaporwave poster?" You protect structure so the restyle reads as the same shot in a new register. You do not ask the model to invent a new pose, crop, or prop set. Notes carry the destination style: watercolor, isometric 3D, Risograph print, editorial fashion illustration, claymation still.

Creators use Change Style to expand one photoshoot into a multi-channel set. Photographers turn a clean studio headshot into ink, charcoal, and flat vector versions for campaign variants. Game artists restyle a real location photo into concept-art language without losing camera height and leading lines. Marketers keep product silhouette and shelf angle while moving from e-commerce white to lifestyle magazine grade.

When to pick Change Style

• The composition is right and the medium is wrong
• You want photo → illustration, modern → vintage, clean → textured, without a new crop
• You are testing brand looks against one locked framing for fair comparison
• Optional notes name a clear destination aesthetic in a few words

Avoid Change Style when lighting is the only problem. A flat office selfie that needs golden hour belongs under Adjust Lighting. Mixing "cyberpunk neon" into Change Style can drag color temperature and style together; that is fine if you want both, noisy if you only wanted a light shift.

Notes that steer restyles

Write destination style, keep constraints short. Good notes: "oil painting on linen, visible brushwork," "flat vector poster, limited palette," "anime key visual, soft cel shading." Weak notes: "make it cooler and more artistic." Cooler means nothing to a vision formatter. Artistic invites fluff.

Name what to keep if the restyle risks deleting it: "keep the red jacket," "preserve label text intent," "same three-quarter crop." Change Style biases toward composition lock; a one-line keep-list stops hard medium swaps from eating brand-critical props.

On PromptMake /image, put those notes in the optional field before you generate. The mode sets the goal. Notes set the destination look. Target Midjourney v7 when you want aesthetic punch and --sref later. Target FLUX when photoreal materials matter under a stylized wrap. Target Ideogram-adjacent workflows when typography in frame is the point (PromptMake formats for Leonardo and SD-family targets too).

Adjust Lighting: same scene, new atmosphere

Adjust Lighting holds subject and style while rewriting light type, contrast, and color temperature. The room stays the room. The product stays the product. Hard noon sun can become overcast soft fill. Flat window light can become rim-lit night with a practical lamp in frame language.

Photographers know lighting is half the grade. Image models know it too, if you feed direction and quality instead of "moody." Adjust Lighting is the mode that forces the prompt to talk like a DP: key side, hard or soft edge, Kelvin, rim separation. You keep medium tokens (photograph, oil, 3D render) so the relight does not turn a photo into a painting by accident.

Use cases stack fast. E-commerce needs the same SKU under softbox white, warm lifestyle window, and dramatic low-key for ads. Architecture visualization needs dusk vs midday from one daytime plate. Portrait series need Rembrandt, butterfly, and backlight versions without reshooting. Upload once, run Adjust Lighting with a note like "golden hour backlight, 3200K warm rim," paste into Midjourney v7 with --style raw, compare side by side.

Pair notes with concrete light language. "Add rim light from behind" beats "more drama." "Overcast soft top light, low contrast" beats "softer vibe." If you also need a style swap, run Change Style as a second generation rather than stuffing both jobs into one Adjust Lighting pass. One goal per run keeps the vision instruction clean.

Create Variation: mood and layout with room to invent

Create Variation uses the upload as mood and layout reference while allowing freer invention. Subject can shift. Props can change. The tool reads palette, energy, and rough framing, then writes a prompt that feels related without promising a twin.

Pick Variation when the reference is a spark rather than a fixed brief. Mood boards, competitor ads you admire for energy, film frames you cannot clone for policy reasons, and AI renders you want to riff on all fit here. Recreate on those assets fights you: the prompt over-constrains, then Midjourney or FLUX collapses into near-copies that miss details and burn edits.

Variation also helps exploration days. Upload one strong composition, run Create Variation three times with different notes ("colder palette," "wider environment," "younger cast"), and collect divergent prompts instead of one fidelity draft. Save the winners in your library. Promote a winner to Recreate only when you decide that frame is now the master.

Ethics stay simple. Variation does not license you to impersonate living artists by name. Prefer medium and era descriptors. Skip --cref on real people without consent on platforms that forbid it. Treat competitor visuals as directional references inside your own brand rules.

How mode and model change the same photo

Mode sets intent. Target model sets dialect. Run the same café portrait four ways and you will see the split in the text before you spend GPU time. Recreate lists window direction, cup material, bokeh, and jacket color. Change Style with "watercolor" drops lens jargon and adds pigment and paper. Adjust Lighting with "neon night interior" rewrites Kelvin and practical sources while keeping the sitter. Variation loosens the cup brand cues and opens seat count or weather.

Cross-model paste fails a lot. A Midjourney v7 Recreate string full of --ar and --stylize confuses FLUX. An SDXL tag soup with (masterpiece:1.2) reads stiff in DALL·E / GPT Image. Generate per target, or rewrite after you copy. PromptMake /image does the per-target formatting when you select the model up front.

Vision backends in 2026 sit on the same class of multimodal models you use for chat analysis: GPT-5.6 Sol / Terra / Luna class stacks, Claude Fable 5 / Opus 5 vision paths, Gemini 3.5 Flash for speed and Gemini 3.1 Pro when the frame is dense. The goal mode matters more than which logo analyzed the pixels, because the system instruction changes what gets emphasized in the final prompt string.

Midjourney v7 and FLUX

Midjourney v7 + Recreate: front-load subject, keep phrases short, end with --ar, --style raw when you want literal light, moderate --stylize so beauty defaults do not melt hard shadows. Add --sref on the original URL after you lock text if style lock beats prompt alone.

Midjourney v7 + Change Style: same structure, medium tokens early ("gouache illustration," "chrome 3D render"). Style mode loves --stylize a bit higher; name the medium so v7 does not invent a random aesthetic.

Midjourney v7 + Adjust Lighting: keep subject nouns stable, rewrite light clauses, prefer --style raw. Pair with notes that name key direction and Kelvin.

Midjourney v7 + Variation: shorter prompts win more runs. Let chaos or higher stylize explore; avoid packing every prop from the reference.

FLUX + any mode: natural language, material and light precision, no MJ flags. Recreate sentences run longer. Variation sentences drop brand-specific clutter. FLUX rewards "brushed steel bead-blasted, soft rectangular softbox from camera left" over "premium product shot."

DALL·E / GPT Image, SDXL, and Leonardo

DALL·E / GPT Image: prose paragraphs. Recreate reads like a careful photo brief. Change Style needs an explicit medium sentence. Adjust Lighting should spell motivated sources ("warm tungsten pendant over the table"). Variation can add "inspired by this scene" language without listing every object. Inline avoid-lists work better than SD-style negatives.

SDXL: positives plus negatives. Recreate fills both: keep detail in positive, push watermarks and bad hands into negative. Change Style swaps medium tags and artist-movement tokens with care (prefer dead movements and medium words). Adjust Lighting edits light tags and contrast words. Leonardo workflows sit close to SDXL: clean tag stacks, game and stylized presets, quality boosters where your fine-tune expects them.

Cross-check habit: generate Recreate for your primary model first. If the look is right but the medium is wrong, rerun Change Style. If only the grade is wrong, rerun Adjust Lighting. Jump to Variation when you are stuck in local minima and need a new prompt skeleton.

Mistakes that waste image to prompt generations

Wrong mode for the ask. Want watercolor? Change Style. Want dusk? Adjust Lighting. Want twin? Recreate. Want riff? Variation. Fix the mode before you rewrite fifty adjectives by hand.

Notes that fight the mode. "Build a new scene from scratch" under Recreate cancels fidelity. "Match every pore and prop" under Variation cancels freedom. Align the note with the goal label.

Vague notes. "Better," "epic," "more cinematic" add noise. Replace with medium, light direction, or one keep-constraint.

Dirty uploads. Heavy watermarks, tiny subjects, and crushed shadows make vision guesses worse in every mode. Crop to the subject. Prefer clear light. Use at least ~512×512 when you can.

Pasteless cross-model reuse. Translate or regenerate per Midjourney v7, FLUX, DALL·E / GPT Image, SDXL, or Leonardo. Dialect is part of the product.

One-and-done expectations. Plan two to five generator iterations after the prompt lands. Change one layer per try: light words, or lens, or palette.

Run goal modes on PromptMake /image

Open https://promptmake.net/image. Upload the reference. Choose the target model. Choose the goal mode that matches the job above. Add a short note if you have a concrete twist. Generate, copy, paste into Midjourney v7, FLUX, DALL·E / GPT Image, SDXL, or Leonardo, then compare against the upload.

Free tier: about 3 image generations per day as a guest, about 5 per day registered, separate from the text enhancer quota. No credit card to start. Soft path: try Recreate on one clean photo today, then rerun Change Style or Adjust Lighting on the same file tomorrow so you feel the instruction shift in the text itself.

When the prompt is close but wordy, trim by hand or send a tightened brief through PromptMake /text for model-ready cleanup. Keep /image for vision. Keep /text for language polish. That split matches how the product is built.

FAQ

What is an image to prompt goal mode?

A goal mode is the creative intent you set before vision analysis writes a generation prompt. Recreate Exactly aims for closest visual match, Change Style keeps composition and swaps medium, Adjust Lighting rewrites light while holding subject and style, and Create Variation treats the upload as loose inspiration. The same upload produces different prompt emphasis under each mode. Pick the mode before you pick adjectives.

Which image to prompt mode should I use first?

Start with Recreate Exactly when the reference is the look you want to keep. Switch to Change Style when framing is right and medium is wrong, Adjust Lighting when only the grade and light need to move, and Create Variation when you want related energy without a clone brief. One clear job per generation beats stacking every wish into a single note. Rerun the same file under a second mode if the first intent was wrong.

How is Change Style different from Create Variation?

Change Style locks subject and composition while swapping artistic medium or aesthetic on purpose. Create Variation loosens that lock and treats the upload as mood and layout fuel. Restyle when you need the same bones in new paint; vary when you want divergent ideas that feel related. If you need both a new medium and a new pose, run Variation with a medium note, or run two separate generations.

Do goal modes work the same on Midjourney v7 and FLUX?

Intent carries across models, but syntax does not. Midjourney v7 wants concise phrases and parameters such as --ar and --style raw, while FLUX wants natural-language scenes with precise materials and light. Pick the target in your image to prompt tool so recreate, restyle, relight, or vary instructions land in the right dialect. Pasting a Midjourney string into FLUX unchanged drops performance.

Can Adjust Lighting turn day into night in the prompt?

Yes, when your notes name the destination light with direction, quality, and color temperature. Example note: "clear night, cool moonlight key from camera right, warm practical window fill." The mode keeps subject and style tokens while rewriting atmosphere clauses. Midjourney v7 with --style raw and FLUX both respond well to that DP language.

How accurate is Recreate Exactly compared to the photo?

Vision analysis on clear, well-lit subjects lands in a strong approximate range for mood, composition, and lighting, in the mid-80s to mid-90s percent for readable frames. Complex mashups, heavy filters, and tiny subjects drift more. No mode recovers seeds or guarantees pixel identity. Plan a few edit rounds in Midjourney v7, FLUX, DALL·E / GPT Image, SDXL, or Leonardo after you paste.

How do I start on PromptMake for free?

Go to https://promptmake.net/image, upload a JPG, PNG, or WEBP under 10 MB, pick a target model, pick a goal mode, and generate. Guests get about 3 image-to-prompt runs per day; a free account raises that to about 5 per day, separate from /text. Copy the prompt into your generator, compare to the upload, then rerun a different mode on the same file if the first intent was wrong.