How to use and evaluate “A table for two”
This portrait prompt is designed as an editable starting point, not a guaranteed one-click result. The notes below translate its JSON into a readable visual plan, identify the controls that matter, and show how to diagnose a weak generation without simply adding more text.
Prompt-specific blueprint
Scene and atmosphere
- Location
- Ordinary small Kadıköy neighborhood restaurant, two simple dinner plates, water glasses and warm wall light, no luxury decor
Visual treatment
- Medium
- Unretouched everyday smartphone photo. Natural pores, slight blemishes, ordinary fabric creases, honest low-light grain. No portrait mode, cinematic grade or beauty filter.
Composition and wardrobe
- People Count
- couple
- People Count Label
- two distinct adults in a romantic couple
- Framing
- waiter takes a knowingly requested rear 1x phone couple photo from 1.2m at seated eye height; table partially in foreground, no portrait mode or fake bokeh
- Aspect ratio
- 4:5
- Wardrobe · Description
- the main reference person wears a navy crew-neck sweater; the adult girlfriend a cream fine-knit sweater, same shoulder-length straight dark brown hair and small silver studs as her reference
Recommended workflow
- 01
Prepare the input
Choose reference images you have permission to use. For A table for two, prioritize a clear subject and avoid unrelated people or private details in the frame.
- 02
Lock the visual plan
Confirm subject count, framing, scene, and lighting before adding finishing language. Resolve conflicts such as close-up versus full-body composition.
- 03
Adapt for one model
Start with ChatGPT Image. Keep a baseline generation and document the exact model version, references, aspect ratio, and visible settings.
- 04
Revise from evidence
Name the visible failure, change one instruction block, and compare it against the baseline. Keep only revisions that improve the intended criterion.
Model adaptation notes
ChatGPT Image
Keep the JSON structure, attach references in the same conversation, and use follow-up edits for one local correction at a time.
A compatibility label means the visual intent can be adapted; it does not promise identical output across providers, versions, or settings.
Prompt-specific quality checklist
- The output preserves the intended portrait subject and scene.
- Lighting direction, camera framing, and environmental details agree.
- The result can be compared against the prompt’s stated purpose.
Failures to check before publishing
- pasted face
- oversized head
- twisted neck
- plastic skin
- extra fingers
- extra limbs
- merged identities
Why the editable controls matter
Gender
Changes Prompt.Identity.Gender without requiring a full prompt rewrite.
People count
Changes Prompt.Composition.People Count Label without requiring a full prompt rewrite.
Photo format
Changes Prompt.Composition.Aspect Ratio without requiring a full prompt rewrite.
Identity or subject drift
Reduce competing style language, improve the reference, and separate stable traits from pose, gaze, and expression.
Composition breaks
Check whether crop, shot distance, aspect ratio, subject count, and required objects can all be satisfied in one frame.
Artificial lighting or texture
Name a motivated light source and believable material behavior before adding grain, grading, sharpness, or resolution terms.
@cankirkgoz
Comments
Share the result you got, attach an image, and reply to other people's results.
You need to sign in to comment.