Skip to main content

Prompt Optimization

Purpose: provide rules and practical examples for creating effective prompts for image generative models in the ceramics/marble sector.

General principles​

  • Describe the desired outcome concretely (material, colors, veining, finish, texture).
  • Order information by importance: first critical elements, then secondary details.
  • Avoid misleading or ambiguous references (e.g., do not use "photographic" if not necessary).

Initial formulation (how to prepare the description)​

  • Indicate the exact material: e.g., "natural Amazzonite marble slab".
  • Specify base color and possible variations: e.g., "turquoise-green with slight variations of light green and beige".
  • Describe the veins: color, thickness, distribution.
  • Define texture and finish: e.g., "realistic texture, slightly honed surface".
  • Make explicit what to avoid: artificial symmetries, repeating patterns.

Example short description: A natural Amazzonite marble slab, elegant and uninterrupted. Turquoise-green base with soft variations of light green and beige. Extremely white cream veins, thin and evenly distributed. Realistic texture, slightly honed surface. No artificial symmetry or repetition.

Interaction with ChatGPT / Prompt to send​

Request template:

  • Request in Italian: "Formula un prompt da fornire a un modello generativo di immagini per il settore della ceramica sulla base del seguente testo. La risposta deve essere intorno alle 100 parole e poi tradotta in inglese. Sulla base di questo prompt, genera anche un negative prompt."
  • Insert the prepared description.

Example of expected result:

Prompt (IT) Natural and elegant Amazzonite marble slab, seamless surface. Turquoise-green base with soft variations of light green and beige. Extremely white, thin, evenly distributed cream veins. Realistic texture, slightly honed finish, no symmetry or repeating patterns.

Prompt (EN) Natural and elegant Amazzonite marble slab, seamless surface. Turquoise-green base with soft variations of light green and beige. Extremely white, thin, evenly distributed cream veins. Realistic texture, slightly honed finish, no symmetry or repeating patterns.

Negative prompt artificial symmetry, tiling, repeating patterns, blurry texture, noise, harsh contrasts, unrealistic veins, plastic look, reflections, metallic shine

Review and correction of the negative prompt​

  • Do not insert direct negations in the negative prompt (e.g., avoid "no tiling"); instead list the elements to exclude as single words/phrases: "tiling", "repeating patterns", "artificial symmetry".
  • Remove ambiguous terms or ones contrary to the context (e.g., "a ceramic slab" → "a marble slab").

Note on the model​

  • When inserting the prompt into the generation pipeline remember to add the model reference with @model.

Optional: use of reference images​

  • You can provide an image to ChatGPT and ask: "Formulate a descriptive prompt of the material based on the following image...". Then proceed with the same textual review described above.

Final checks before generation​

  • Verify the order of importance of elements.
  • Check linguistic accuracy of EN/IT translations and avoid misleading terms or references within the prompt (for example: ‘A ceramic slab’ → ‘A marble slab’).
  • Ensure the negative prompt contains only terms to exclude, without negations.
  • Avoid words that introduce undesired photographic style if the result should be neutral.