Article

How to Write Better GPT Image 2 Prompts

GPT Image 2 is OpenAI's latest image generation and editing model, available natively on HiArt. It excels at following detailed prompts, rendering readable text in layouts, and keeping visual consistency when you attach reference images — making it the default choice for brand assets, e-commerce hero shots, and polished social graphics.

Unlike faster budget models, GPT Image 2 rewards specificity. This guide covers the prompt patterns, reference workflows, and settings that consistently produce strong output on the HiArt canvas.

What you can do

Text-to-image — describe a scene from scratch; no uploads required
Reference-guided generation — attach up to 16 images to lock product shape, brand colors, or character look
In-canvas edits — select an existing image node and describe what to change; HiArt routes edits through GPT Image 2
High-resolution export — up to 4K for print-ready and large-format social assets

Specs at a glance

Setting	Options
Aspect ratios	1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 4:5, 5:4, 21:9, 1:4, 4:1, 1:8, 8:1
Resolution	1K, 2K, 3K, 4K
Quality	Auto (default), Low, Medium, High
Reference images	Up to 16 per generation
Output	Single image per generation

GPT Image 2 supports more reference images than any other image model on HiArt — use that headroom when you need SKU consistency across angles, packaging variants, or a full brand kit.

How to structure a prompt

Strong prompts usually follow this order. You do not need every section every time — but skipping subject and style is where most weak results come from.

[Subject] + [Action/Pose] + [Environment] + [Lighting] +
[Camera/Composition] + [Style/Medium] + [Text overlay (if any)]

Subject — be concrete

Name materials, colors, and scale. "A matte-black wireless earbuds case" beats "some earbuds."

Environment — anchor the scene

Studio sweep, marble countertop, rainy neon street, minimalist SaaS landing hero — one clear location prevents drift.

Lighting — cheap quality boost

Soft diffused studio lighting, subtle rim light — product and pack shots
Golden hour, warm backlight, long shadows — lifestyle and outdoor
Overcast daylight, even exposure — fashion and editorial
Neon practicals, wet pavement reflections — cyberpunk and nightlife

Style — lock the medium

End with a medium tag: Photorealistic commercial photography, Flat vector illustration, 3D render, Octane-style, Watercolor editorial, Anime key visual. Mixing two conflicting mediums in one prompt often causes muddy results.

Reference images — when and how

Without references, GPT Image 2 generates freely from text. With references attached, HiArt sends an edit request — the model uses your images as visual anchors while following the prompt.

Common reference roles

Goal	What to upload	Prompt tip
Product consistency	Existing SKU photos from 2–3 angles	Name what must stay: shape, logo placement, label text
Brand colors	Palette swatch or prior campaign key visual	Say "match the color palette of the reference" explicitly
Character look	Turnaround or portrait refs	Describe pose and scene as new; anchor face, hair, outfit to refs
Layout remix	Wireframe or competitor layout you like	Specify which elements to keep vs replace
Style transfer	Art reference with clear medium	"Same illustration style as reference, new subject: …"

Tell the model what each reference is for. Vague lines like "use the images" leave too much room for the model to guess.

Hero product shot of the earbuds case from the reference images.
Keep exact logo placement and matte-black finish. New scene: floating
above a reflective dark surface with soft studio lighting. 16:9 crop.

Typography and text in images

GPT Image 2 is one of the stronger models for readable in-image text — but you still need to spell out copy, placement, and hierarchy.

Put exact headline copy in quotes: Headline reads "SUMMER DROP" in bold sans-serif
Specify placement: top third of frame, bottom-left corner, centered over dark sky area
Separate headline, subhead, and legal: Tagline below in smaller grey text: "Free shipping over $50"
Name the type style: condensed grotesque, elegant serif, rounded friendly sans
For non-English copy, include the exact characters — do not rely on translation

Outdoor adventure poster. Massive orange camping tent in a pine forest
at night, campfire glow, cinematic atmosphere. Large rugged serif
typography "WILDCAMP" in bold orange dominates the upper sky. Tagline
at bottom: "Sleep under the stars." Small grey text top-right:
"Designed with GPT Image 2." Photorealistic commercial style.

Resolution and quality

Pick resolution to match delivery — not every draft needs 4K.

Setting	When to use
1K / 2K + Auto quality	Concept exploration, layout tests, Agent drafts
3K / 4K + High quality	Final social assets, print-adjacent exports, client deliverables
Auto quality	Default — good balance for most canvas work
Low / Medium	High-volume iteration when credit budget matters

HiArt shows estimated credits before you generate — draft at 2K, then bump resolution once the composition is locked.

Examples that work

E-commerce hero (text-to-image)

Studio product photo of a ceramic pour-over coffee set in matte sage green.
Steam rising from the carafe. Marble surface, soft diffused lighting,
shallow depth of field. Minimal props — single eucalyptus sprig.
Photorealistic, high-end DTC brand aesthetic. 4:5 vertical crop.

Social carousel slide

Flat illustration for a fintech app onboarding screen. Friendly character
holding a phone showing a savings chart. Pastel blue and coral palette,
rounded shapes, plenty of whitespace. Headline area left blank for overlay
in Figma. Clean vector style, no gradients.

Reference-guided product refresh

Using the attached product references, generate a summer campaign hero.
Same bottle shape, cap color, and label typography as refs. New scene:
condensation droplets, sliced citrus on ice, bright beach daylight.
16:9 landscape for web banner.

In-canvas edit

Change the background to a warm sunset gradient. Keep the product
position and lighting on the subject identical. Remove the extra props
on the right side.

A quick workflow on HiArt

Open the canvas and add an Image node — or select an existing one to edit.
Choose GPT Image 2 from the model picker; set ratio and resolution.
Attach reference images if you need consistency; write the prompt using the structure above.
Generate at 2K / Auto first; review composition and text placement.
Re-run at higher resolution or switch to High quality for the final export.
Branch nodes to A/B test backgrounds, copy, or crop without losing the original.

Mistakes to avoid

Vague subjects — "nice product photo" gives the model nothing to anchor.
Conflicting styles — "photorealistic watercolor anime 3D render" in one line.
Unspecified text — asking for "a headline" without the exact words or font mood.
Too many changes at once — in edits, stack 1–2 changes per generation.
Wrong ratio for platform — generating 1:1 then cropping for Stories loses composition.
Skipping references — when SKU accuracy matters, text alone is rarely enough.

Copy-paste templates

Product ad (21:9 banner)

Cinematic widescreen product hero. [PRODUCT] centered on reflective
black surface. Dramatic side lighting, subtle smoke or mist. Brand colors:
[COLORS]. Headline left third: "[HEADLINE]" in bold condensed sans.
Photorealistic commercial photography, ultra-detailed textures.

Instagram portrait post (4:5)

Lifestyle photo. [SUBJECT] using [PRODUCT] in [LOCATION]. Natural
daylight, candid moment, shallow depth of field. Warm color grade.
Leave lower 20% relatively clean for caption overlay. 4:5 vertical.

App store screenshot frame (9:16)

Mobile app mockup on iPhone frame, floating at slight angle. Screen
shows [UI DESCRIPTION]. Soft gradient background in [BRAND COLORS].
Headline above device: "[HEADLINE]" in modern geometric sans.
Clean SaaS marketing style, generous padding.

Want model specs and credit estimates side by side? See the GPT Image 2 model page. Ready to generate — open an image node on HiArt canvas, select GPT Image 2, and start with one clear subject, one lighting setup, and one style tag.