Prompting Runware models

Different model families read prompts differently. Match the phrasing to the model. When in doubt, confirm the model's strengths via runware-models.

Cross-family principles

• Layer the scene for complex images: subject → environment → camera/framing → lighting → mood. Name objects and their positions, not just objects.
• Short and interpretive vs long and structured. Modern models reward concise intent; older/diffusion models reward dense descriptive stacks. Pick per family below.
• In-image text is quoted, exact, and placed. To render words in an image, quote the exact string and state placement and style. Models render what you quote, not what you describe.
• Negation: some models take inline negation in the positive prompt (write "Negative prompt: X, Y" as a clause and the model obeys); others have a real negativePrompt field. Check the schema.

Image families

• LLM-based image models (gpt-image-2, Nano Banana 2): parse full natural language. Use structured briefs, inline negation, even pseudocode; say "photorealistic" explicitly; use camera language for composition. Nano Banana 2 has a thinking setting for hard prompts and renders legible in-image text well; it also grounds on real facts via providerSettings.google.webSearch.
• Seedream / Recraft: short interpretive prompts work; layer (subject → lighting → composition) for control. Recraft's Utility variants give flat, predictable output for mockups; constrain color with settings.colors.
• Ideogram: operates on a structured JSON (reserved keys: description, style, background, elements), not a sentence. Text is a first-class text element. Use Magic Prompt to expand natural language, or hand-craft the JSON for repeatable layouts.
• Grok / typography models: quoting the exact text is non-negotiable; placement and script/style are refinements.

Video families