01 Core Formula
GPT Image 2 responds best to plain, descriptive language. You don't need keyword stuffing — you need a clear sentence that names the subject, the style or medium, and the light. The rest is optional seasoning.
1.1 The Basic Prompt Formula
Start with three required parts — subject, style and medium — then add optional scene details. Write full sentences; the model rewrites your prompt internally, so vague one-word lists tend to drift.
[Subject] + [Style] + [Medium] + optional [Lighting] [Background] [Detail] [Mood]
Subject
The main person, object or scene. Be specific: 'a tabby cat' beats 'a cat'.
Style
The visual treatment: photorealistic, anime, 1970s film still, studio product shot.
Medium
How the image is made: oil painting, digital render, Polaroid, editorial photo.
Lighting
Golden hour, softbox, neon backlight, overcast window light — light is the fastest mood setter.
Background
Keep it simple unless the environment is the point.
Detail
One or two telling specifics: texture, color accent, worn edges, shallow depth of field.
Mood
A single emotional word: serene, tense, playful, lonely.
Botanical Still Life
Prompt
A botanical still life of eucalyptus, dried pampas grass and a single pink ranunculus arranged in a matte ceramic vase. Soft natural window light from the left, pale linen backdrop, gentle shadows. Editorial product photography, muted earth tones, high detail.
Subject + medium + light in one sentence; style and mood close it out. Preloaded in Easy mode at 4:5 / 1K — the cheapest way to test the composition.
1.2 Detail, Medium and Surface
GPT Image 2 is good at material and surface language. Tell it what something is made of and how the light hits it, and the texture usually follows. This is where the Advanced tier pays off: larger sizes and higher quality give fine detail room to survive.
Oil Painting Landscape
Prompt
An oil painting of a misty mountain lake at dawn, pine-covered slopes reflected in still water, a small wooden rowboat tied to a weathered dock. Visible brushstrokes in the clouds, impasto texture in the foreground reeds. Warm ochre and cool teal palette, cinematic composition.
Preloaded at 1024x1024 / high / webp. The high-quality setting preserves the brushstroke detail that low or medium would smooth away.
02 Image References and Edits
GPT Image 2 can take one or more input images as visual references. Upload a sketch to turn it into a render, upload several shots of a character to keep design consistent, or hand it a product photo and ask for a new finish.
2.1 Upload and Prompt
The simplest edit: upload one image and describe what you want changed or kept. The model processes the input at high fidelity automatically, so you don't need any extra fidelity switch.
Sketch to Render
Prompt
Turn this pencil sketch into a clean 3D product render of a minimal wireless speaker. Soft studio lighting, matte off-white plastic body, warm fabric grille, rounded corners. Floating at a slight angle on a neutral gray background, subtle reflection underneath.
Upload your sketch image in image-to-image mode. Preloaded in Easy mode at 1:1 / 2K for a quick, low-cost iteration.
2.2 Multi-Reference Consistency
For character or product consistency, upload multiple reference images and describe the exact pose or angle you want. Generating several variants at once makes it easier to pick the one that keeps the design intact.
Character Turnaround Pick
Prompt
A brave young explorer in a khaki safari jacket, leather satchel and wide-brimmed hat, standing in a jungle clearing. Use the uploaded character references to keep the face, outfit and color scheme consistent. Soft dappled sunlight through the canopy, adventure-storybook illustration style.
Upload 2–4 reference images of the same character. Preloaded in Advanced mode at medium quality, 4 images at once, so you can choose the best match.
2.3 High-Fidelity Edits
When the input is already a high-resolution photo and you want to preserve that fidelity, use the Advanced tier at a larger size and high quality. Good for product reskins, background swaps on portraits, and material changes.
Product Reskin
Prompt
Take the uploaded product photo and change the sneakers to a deep navy suede with off-white sole stitching, keeping the same studio lighting, shadow and camera angle. Premium e-commerce product photography, crisp detail, no extra props.
Upload the original product shot. Preloaded at 2048x2048 / high / png — the settings that best preserve source detail.
03 Quality, Size and Output Format
GPT Image 2 pricing depends on quality and size. Use low quality and small sizes for fast drafts; reserve high quality and larger sizes for final assets. Output format matters too: PNG for archival, JPEG/WebP for smaller files.
3.1 Draft Quality: Iterate Cheap
Set quality to low and size to 1024x1024 when you're exploring composition, style or color. The cost is a fraction of high quality, and the output is already good enough to judge direction.
App Icon Exploration
Prompt
A modern app icon for a habit tracker: a friendly paper airplane leaving a dotted loop trail, rounded square icon, soft gradient background in mint and cream. Flat vector style, clean shapes, no text.
Preloaded at low quality, 1024x1024, 4 images at once — the cheapest way to compare variations before committing to a final render.
3.2 Final Quality: Large and Sharp
When the concept is locked, rerun the winning prompt at high quality and the largest size you need. Resolutions above 2560x1440 are still experimental, but they're available for high-stakes deliverables.
Hero Cinematic Frame
Prompt
A cinematic wide shot of a lone astronaut standing on a rust-red Martian ridge at sunset, a small habitat module glowing warm in the valley below. Dust particles catch the golden light, vast alien sky with two small moons. Epic sci-fi concept art, highly detailed, film grain.
Preloaded at 3840x2160 / high / png. This is the most expensive preset — use it only once the prompt is already tested at a smaller size.
3.3 Format and Compression
PNG is the default and the safest for further editing. Use JPEG or WebP when file size matters; WebP usually gives the best size-to-quality ratio.
Social Media Photo
Prompt
A candid lifestyle photo of two friends laughing over coffee at a sunlit café, warm wooden table, soft bokeh background. Natural window light, authentic moment, slightly desaturated film look.
Preloaded at 1536x1024 / medium / jpeg — a good balance of quality and file size for web and social use.
04 Moderation and Limitations
GPT Image 2 has the same content policy as other OpenAI models, plus a few practical limits. Knowing them upfront saves you from surprise blocks and disappointment.
4.1 Content Moderation
All prompts and outputs are filtered. The default `auto` level is standard. If a request is blocked, the error code is `moderation_blocked` — retrying the same prompt won't help.
Fantasy Creature
Prompt
A friendly baby dragon with iridescent scales curled around a stack of old spellbooks in a cozy wizard's attic. Warm candlelight, dust motes, storybook illustration style.
Fantasy and stylized content is usually fine. Avoid requests for realistic violence, gore or sexual content — those will hit the moderation filter regardless of artistic framing.
4.2 Text, Layout and Consistency
Text rendering is still unreliable: spellings can drift and placement can wander. Layout-sensitive compositions — precise grids, aligned UI mocks, exact packaging dimensions — are also hard to control. Treat the output as a starting image and plan to refine text in a design tool.
Packaging Flatlay
Prompt
A flatlay product photo of a premium chocolate box, a single open sleeve revealing dark chocolate squares, gold foil accents, scattered cocoa nibs. Soft overhead light, cream marble background, luxury packaging photography.
The layout and product form come out well; any text on the box will likely need to be added or corrected afterward in a design tool.