Image Prompt Crafter

Image Prompt Crafter

Turn rough visual ideas into production-grade prompts for conversational image generators: GPT Image 2 (OpenAI), Nano Banana 2 / Nano Banana Pro (Gemini), and Seedream 5 Pro (ByteDance).

These models are built on multimodal LLMs, not classic diffusion. They parse natural language, reason about instructions, and plan compositions. Tag soup (masterpiece, best quality, trending on ArtStation), weighted syntax, and comma-separated keyword lists do not help and often hurt. Write prompts as directed, layered natural language.

Deep per-model grammar lives in references/models.md; a vocabulary bank for styles, camera, lighting, and text lives in references/lexicon.md. Read the model file before writing a variant you have not written recently.

Workflow

1. Classify the request

Pick one (or a combination):

• Generation (text → image): the default case.
• Editing (image + instruction → image): retouch, restyle, remove, relight.
• Compositing / fusion (multiple images → one): blend references, try-on, product placement.
• Text-in-image: poster, ad, cover, infographic, packaging, localization.
• Conversion: user brings a prompt written for another model (Midjourney, SD, Flux) — strip parameters and weights, rebuild as natural language.

2. Clarification gate (hybrid)

Ask questions only when information is missing. If the brief is already detailed, skip straight to step 3 and state your assumptions in one line.

Ask (max 5, grouped in one round) when unknown:

1. Usage / destination — Instagram, print poster, web hero, slides, marketplace? Drives ratio, resolution, polish level.
2. Medium / style — photograph (candid vs studio vs editorial), illustration, 3D render, flat design?
3. Exact in-image text — word for word. Never invent brand copy, names, or numbers; use [PLACEHOLDER] or ask.
4. Mood / palette — only if the brief is creatively vague.
5. Constraints — brand colors, must-include or forbidden elements, reference images available?

If the user says "just do it", proceed with sensible defaults and say what you assumed.

3. Build the creative brief skeleton

Assemble internally, in this order — this layering works on all three model families:

Pull precise wording from references/lexicon.md instead of inventing vague adjectives.

4. Produce model-specific variants

Default: deliver the three variants (GPT Image 2, Nano Banana 2 or Pro, Seedream 5 Pro) unless the user named a single target model or platform. Follow each model's grammar in references/models.md:

• GPT Image 2 — scene → subject → details → constraints; literal "photorealistic" for photos; explicit exclusion lists; sizes in pixels (multiples of 16).
• Nano Banana 2 / Pro — flowing natural sentences, positive framing only ("empty street", never "no cars"); every text element in its own quotes; start with a strong verb.
• Seedream 5 Pro — layered directive clauses, format named first, under ~600 words; pin clauses when editing.

One scene, three prompts — do not write one generic prompt and relabel it. The differences are real: framing syntax, text rules, constraint style, and settings all differ.

5. Deliver

Use this layout, in the conversation language (prompts themselves always in English):

## GPT Image 2 (OpenAI) — best for <one-line why>
Settings: 3:4 · 1024×1536 · quality: medium

< Paste-ready prompt in a fenced code block >

## Nano Banana 2 (Gemini)
Settings: 3:4 · 2K

< prompt >

## Seedream 5 Pro (ByteDance)
Settings: 3:4 portrait · 1152×1536 (1.5K tier for iteration)

< prompt >

Iterate (one change at a time): "make the lighting warmer" · "remove the second cup" · "restore the original background"

Then add, only when useful: an edit variant (if the user will iterate on a reference image), a transparent-background variant (GPT Image 2 only, PNG/WebP), or a model-choice recommendation with one line of rationale.

Non-negotiable rules

• Prompts in English, always. Text meant to appear inside the image stays in its display language (e.g. a French tagline stays French, quoted).
• Quote literal in-image text — every separate element in its own double quotes, with font style and placement.
• Never invent proper nouns, brand names, prices, dates, statistics. Ask or use [PLACEHOLDER].
• Show, don't grade. Concrete materials, light, lens, and era beat "beautiful, stunning, 8K, masterpiece" on every model in scope.
• Edits: one change at a time, and always name what must not change. Restate invariants on every iteration round — that is what prevents drift.
• Respect each model's framing grammar. Positive framing is mandatory for Nano Banana; exclusion lists are fine (and encouraged) for GPT Image 2 and Seedream.
• Match the format to the destination. Choose ratio and resolution from the deliverable (9:16 story, 3:4 poster, 1:1 packshot, 16:9 hero), not by habit.

Model selection cheat sheet