Prompt Adapter

Convert prompts between AI models while preserving intent and maximizing output quality.

Adaptation Workflow

Step 1: Identify Source and Target

Determine:

1. Source model: What model was this prompt written for?
2. Target model: What model should it run on?
3. Priority: Preserve style fidelity or optimize for target strengths?

Step 2: Analyze Source Prompt

Break down the prompt into components:

• Core subject/action
• Style modifiers
• Technical parameters (model-specific)
• Negative prompts (if any)
• Aspect ratio / dimensions

Step 3: Apply Model Translation Rules

Load {PROMPT_ENGINE_DIR}/references/model-guide.md for detailed rules. Key translations:

Midjourney -> Flux:

• Remove --ar, --v, --style, --s, --chaos parameters
• Expand shorthand into natural language descriptions
• Flux prefers longer, more descriptive prompts
• Remove :: weight syntax, integrate naturally

Midjourney -> DALL-E:

• Remove all -- parameters
• Rewrite as clear, direct descriptions
• DALL-E prefers straightforward language over artistic jargon
• Remove negative prompts (DALL-E doesn't support them well)

Flux -> Midjourney:

• Add --ar for aspect ratio
• Add --v 6.1 or appropriate version
• Condense long descriptions into key phrases
• Add style parameters (--style raw for photorealistic)

Any -> Sora (Video):

• Add camera movement descriptions (pan, zoom, tracking, etc.)
• Add temporal flow ("the scene transitions from... to...")
• Specify duration if possible
• Focus on motion and action over static details

Any -> Leonardo AI:

• Reference specific Leonardo models (Phoenix, Alchemy, etc.)
• Use Leonardo-specific quality tokens
• Adapt negative prompts to Leonardo format

Step 4: Search for Target Model Examples

Find reference prompts in the target model:

python3 {PROMPT_ENGINE_DIR}/scripts/search_prompts.py "SUBJECT" --model TARGET_MODEL --limit 3

Use these as style references for the adaptation.

Step 5: Present Adaptation

Output format:

1. Original prompt (source model labeled)
2. Adapted prompt (target model labeled)
3. Translation notes (what changed and why)
4. Parameter mapping (source params -> target params)
5. Confidence level (High/Medium/Low -- based on model compatibility)

Common Pitfalls

• Midjourney weight syntax (::2) has no direct equivalent in most models
• DALL-E ignores most style parameters -- weave them into descriptions
• Sora needs temporal language that image models don't use
• Aspect ratios must be specified differently per platform
• Some styles only work well on specific models (e.g., --niji is Midjourney-only)