Reference Consistency with GPT Image 2.5: Characters & Products

Series work fails when every reroll invents a new face or label. GPT Image 2.5 stays steadier when references lock identity and the prompt is only this shot’s brief—plus a hard rule: change one variable per turn.

What you are actually locking

Type	Lock first	Safe to change per shot
Person / character	Face, hair, age feel, signature props	Outfit, expression, scene, camera
Product / SKU	Shape, materials, colorway, logo placement	Background, light, staging angle
Brand series	Color temperature, lens language, finish	Props and copy scenarios

Locking the wrong layer (style background, not the person) yields “same vibe, different human.”

Five-step checklist

1. Write one identity sentence: age feel, hair, materials, logo placement.
2. Attach 1–3 clear references on Create (front or 3/4 for people; clean 3/4 for products—no heavy filters).
3. Use the four-layer prompt stack, but let refs own identity; the text states this shot’s deltas.
4. Change one variable per turn: outfit or background or camera.
5. Explore in Flare; finish and refine in Sunburst (Flare vs Sunburst).

Choosing references that work

Good refs usually have:

• Unobstructed key features
• Normal lighting (not beauty filters that reshape bones)
• The same identity across multiple images
• Real materials and logo regions for products (even if finals omit text)

Weak refs: heavy compression, collage of unrelated styles, “pretty but not the same person,” mixed SKUs.

Prompt vs reference division of labor

Treat the prompt as a shot brief, not a character redesign:

same woman as reference, linen blazer, evening city street, soft neon rim light,
eye-level medium shot, natural skin texture

Product:

same matte black earbuds as reference, on oak desk, morning window light from the left,
45-degree overhead, shallow depth of field, no readable text

same … as reference is a reminder; the uploaded images do the real locking. Do not describe a conflicting face or bottle shape in text.

Why one variable per turn works

Changing hair, outfit, scene, and lens together forces the model to re-guess the hero. Split the loop:

1. Freeze refs + subject clause
2. Swap background only → confirm identity holds
3. Swap lighting only
4. Swap camera only

When drift appears, you know which step caused it. Works for KV sets, PDP galleries, and character sheets.

Extra tips by type

People

• Lock a hero portrait first, then derive scenes
• Expression can move; bone/age language should not be rewritten every turn
• For big poses, keep hair and age clauses stable

Products

• Shoot a clean “passport” frame, then derive lifestyle scenes
• If logos wobble: use no readable text and composite type later, or try Sunburst with fewer adjectives
• For SKU lines, change only color/accessory words; keep materials and lens language

60-second QA before you deliver a series

Cheaper than rebuilding the set later:

1. Recognize — can you still ID the same person/SKU from thumbnails?
2. Materials — do metal, fabric, and glass share one lighting logic?
3. Logo zone — is placement roughly stable? If not, composite type later or reduce prompt noise.
4. Single variable — relative to the last frame, did you change only what you meant to?
5. Model stage — still exploring on Sunburst, or shipping finals from Flare?

Log the winning ref set, identity line, and which variable changed—your next series will start faster.

Common mistakes

• Rewriting the whole prompt when only lighting is wrong
• Mixing refs that fight (photo + heavy paint + cel-shade)
• Skipping simple negatives for blur, extra limbs, or stray text
• Attaching too many refs during early exploration
• Switching model, prompt, and refs at once
• Rerolling for luck instead of locking references—credits burn while identity still drifts

Wire it to the rest of the site

1. Unstable structure → prompt structure
2. Model split → Flare vs Sunburst
3. First loop → Getting started
4. Practice → Create; templates → prompt library

Next step

Pick one real character or SKU: attach 1–2 refs, change only the background twice in Flare, then polish one Sunburst final. If you can still recognize them across frames, the consistency loop is working. Need a composable product cutout next? See GPT Image 2.5 transparent cutouts: real alpha and edge QA on a dark backdrop.