The complete masterclass on OpenAI's gpt-image-2 released April 21, 2026. The 7-Part Prompt Formula. 100 paste-ready prompts. 8 production templates. 5 reusable skill packages. Thinking Mode, mask-based editing, multi-image reference, API engineering. Beginner to expert, no gatekeeping.
ChatGPT Images 2.0 (model ID gpt-image-2, snapshot gpt-image-2-2026-04-21) is OpenAI's flagship image generation system released April 21, 2026. It is autoregressive, natively integrated into the GPT architecture, and replaces both DALL-E 3 and gpt-image-1.5 as the default model.
Five things changed: pixel-perfect text rendering across Latin and non-Latin scripts, native Thinking Mode (reasoning + web search + self-verification), 2K standard resolution with 4K via select hosts, up to 8 coherent images per prompt with continuity, and up to 16 reference images for editing.
This masterclass walks every capability end-to-end with the 7-Part Prompt Formula, 100 paste-ready prompts, 8 production templates, 5 reusable skill packages, and the honest decision matrix against Midjourney V8, Nano Banana 2, and Stable Diffusion.
This is a 10-module curriculum plus reference sections. Most readers do not need every module. Pick the persona closest to your situation and the path below points to the modules that pay back fastest.
You want to ship your first usable image today. You have a ChatGPT Free or Plus account and a browser. That is all you need.
You want consistency, character continuity, multi-format outputs, and brand-locked production. You will live in Thinking Mode.
You need the API surface, rate limits, cost math, edit endpoints, and the integration patterns that scale.
The image generation game changed on April 21, 2026. OpenAI shipped gpt-image-2 with no keynote and no hype cycle. It scored 1,512 on the Image Arena leaderboard, the largest lead in Arena history (+242 points over second place). This module is the orientation: what changed, what to use, what to pay.
The most important access distinction in the new model. Both render images. Only one reasons before generating.
| Model | Released | Status | Use For |
|---|---|---|---|
| gpt-image-2 | April 21, 2026 | Current flagship | All new work. Thinking Mode. Text rendering. Multi-image edits. |
| gpt-image-1.5 | December 2025 | Available, legacy | Transparent backgrounds (not yet supported in gpt-image-2). Stable production fallback. |
| gpt-image-1 | April 2025 | Deprecated, shutdown October 23, 2026 | Migrate immediately. Do not start new work here. |
| DALL-E 3 | October 2023 | Retired May 12, 2026 | No longer available. All references in old guides are dead. |
| DALL-E 2 | April 2022 | Retired May 12, 2026 | No longer available. |
The API switched from flat per-image pricing to token-based billing. Estimated per-image costs at 1024×1024 (rough, varies with prompt complexity):
Raw token rates: image input $8/M, cached image input $2/M, image output $30/M, text input $5/M, cached text $1.25/M, text output $10/M. The Batch API halves these rates if you can tolerate 24-hour latency.
For the vast majority of practical work — product photos with labels, marketing graphics, infographics, UI mockups, posters, multilingual content — gpt-image-2 is now the default. Midjourney still wins on pure artistic atmospheric beauty. Nano Banana 2 wins on cost-sensitive bulk and Google Search grounding. Stable Diffusion wins on style fine-tuning and content that gets refused by hosted models. Everything else: gpt-image-2.
If you remember nothing else from this masterclass, remember this formula. Master prompt engineers use the same seven-part structure on every production prompt. Fill every slot in order. Quality scales with specificity, not adjective stacking.
Stack your prompt in this exact order. Every slot has a job. Skip a slot and quality drops.
The "who" or "what" of the image. Be concrete: "a matte black ceramic coffee mug" beats "a coffee cup."
Named visual treatment. "Editorial product photography" beats "professional photo." "Cinematic film still" beats "movie-like."
Framing and layout. "Centered, eye-level, symmetrical" or "rule-of-thirds with subject on the left." Specify negative space.
The single biggest lever for realism. Specify direction, hardness, color temperature: "soft diffused daylight from the upper-left, golden hour temperature."
Photography vocabulary translates directly: "shot on 85mm prime lens, f/1.8, shallow depth of field" or "wide-angle 24mm, deep focus."
The emotional register. "Calm, contemplative, premium." "Energetic, playful, optimistic." Two or three adjectives, not five.
What you do NOT want. Aspect ratio, banned elements, required readability. "16:9 aspect ratio. No watermarks. No extra text. Text must be sharp and legible."
A stunning hyper-realistic photo of a coffee shop in the morning. Make it look amazing and professional with beautiful lighting.
Subject: A neighborhood coffee shop interior with three customers reading at a wooden counter, a barista pulling espresso, and steam rising from a portafilter.
Style: Editorial photojournalism, slightly desaturated, warm undertones.
Composition: Wide shot from the entrance looking in, customers visible in the middle ground, barista in sharp focus in the background, rule-of-thirds.
Lighting: Soft morning daylight streaming through a tall window on the right, warm tungsten accents from pendant lamps, subtle haze in the air.
Camera and lens: Shot on a 35mm prime, f/2.8, slight motion blur on the steam.
Mood: Quiet, focused, unhurried, premium-neighborhood.
Constraints: 16:9 aspect ratio. No visible brand names on cups or signage. Photorealistic. No people facing the camera.
The strong prompt produces a usable image 9 times out of 10. The weak prompt produces a generic AI coffee shop with garbled signage. Same model, same time, different structure.
These adjectives feel like they should help. They do not. They dilute your prompt by adding weight to non-specific tokens. Remove them and watch quality go up.
Banned: stunning, hyper-realistic, breathtaking, amazing, beautiful, professional, high-quality, masterpiece, award-winning, epic, ultra-detailed, 8K, ultra HD, photorealistic (the model is already photoreal — saying it doesn't help).
Replace with: specific lighting, specific lens, specific material, specific lighting temperature, specific composition. Concrete physical descriptors always beat aesthetic adjectives.
Copy this skeleton into a notes app. Fill it in for every prompt. Stop guessing.
Subject: [who or what, with concrete physical descriptors]
Style: [named visual treatment — editorial, cinematic, documentary, illustrative, etc.]
Composition: [framing, where the eye lands, negative space, rule of thirds vs symmetric]
Lighting: [direction, hardness, color temperature, atmospheric notes]
Camera and lens: [focal length, aperture, depth of field]
Mood: [2-3 emotional descriptors, no more]
Constraints: [aspect ratio, banned elements, required readability, file format]
If you have to ask a question like "should I add more detail?" the answer is almost always no. Add concrete specificity to existing slots before adding new ones. A prompt with seven slots filled at high specificity beats a prompt with twelve slots filled at low specificity. Always.
This is where gpt-image-2 produces work indistinguishable from a real photo studio. The key is treating the prompt like a brief to a working photographer, not like a wish to a genie. Photography vocabulary translates directly to model behavior.
Light is the single biggest variable. The phrases below all produce predictable, repeatable results.
1. The plastic skin look: Skin renders too smooth. Fix: add "natural skin texture, visible pores, subtle imperfections, no skin smoothing."
2. The over-edited HDR look: Everything looks vivid and clarity-pushed. Fix: add "subtle contrast, gentle midtones, no HDR processing, color-graded for editorial."
3. The wrong-eye-line gaze: Eyes look slightly off. Fix: specify gaze direction explicitly ("looking directly into the lens" or "looking off-camera to the left").
4. The blank background trap: Pure white feels cut-out. Fix: add "subtle drop shadow, soft gradient, slight floor reflection" — never just "white background."
Studio product photography of a wireless over-ear headphone in matte graphite finish, centered on a clean off-white seamless background, soft even softbox lighting from above and slightly behind, subtle floor reflection and soft drop shadow, 85mm prime lens, sharp focus throughout. Editorial e-commerce style for an Amazon-style listing. 1:1 aspect ratio. No text or watermarks.
Overhead flat lay of skincare bottles, eucalyptus leaves, a folded linen towel, and a small terracotta dish on a pale concrete surface, soft diffused daylight from above-right, muted sage green and warm beige palette, 50mm lens, sharp focus across the plane, editorial product styling. Square aspect ratio. Generous negative space top-left.
Lifestyle product shot of a person's hands in a cream knit sweater holding a ceramic mug of coffee, sitting by a tall window in a sunlit modern kitchen, warm morning light streaming in from the left, shallow depth of field with the window soft in the background, 50mm lens at f/2, lifestyle brand photography. 4:5 aspect ratio. Face not visible. Photorealistic without HDR processing.
Single hero shot of a sculpted glass perfume bottle on a brushed brass plinth, deep navy gradient background, rim lighting from behind separating the bottle from the background, glossy floor reflection, 100mm macro lens at f/8, sharp focus on the bottle's edge highlights, premium fragrance brand aesthetic, soft volumetric haze in the background. 4:5 aspect ratio.
Extreme macro of a single drop of clear serum landing on a glass eyedropper, capturing the moment of impact with crown splash, pure black background, single hard light from above-right, 100mm macro at f/5.6, frozen motion, pristine reflections, beauty editorial style. 1:1 aspect ratio. No text. Photorealistic without color grading.
Overhead three-quarter shot of a rustic ceramic bowl of roasted heirloom tomatoes, torn sourdough, and fresh basil on a weathered wooden table, golden hour side-light from a window out of frame, deep saturated reds and warm wood tones, 50mm prime, f/2.8, shallow depth of field on the bread crust, food editorial style. 4:5 aspect ratio.
Environmental portrait of a master watchmaker working at a wooden bench in a small Geneva atelier, leather apron over a white shirt, magnifying loupe in hand, soft window light from the left, warm tungsten task lamp on the bench, 35mm lens at f/2.8, slight depth of field, documentary editorial style, no posing. 3:2 aspect ratio. Photorealistic with natural skin texture.
Three-quarter front view of a vintage 1960s European sports car parked on a deserted coastal road at dusk, deep cobalt paint with chrome details, blue hour ambient light, single key light from the upper-left simulating residual sunset, 35mm lens at f/5.6, atmospheric haze in the background mountains, cinematic film still aesthetic. 21:9 aspect ratio. No license plate text.
Wide interior shot of a minimalist Scandinavian apartment, white oak floors, off-white walls, a single oxidized brass pendant lamp over a long oak dining table, late afternoon sun streaming through floor-to-ceiling windows on the right, soft shadows on the floor, 24mm wide-angle at f/8, deep focus, editorial architecture photography. 3:2 aspect ratio.
Mid-shot of a model in an oversized camel wool coat standing against a brutalist concrete wall, hands in pockets, looking off to the right, overcast diffused daylight, neutral cool grey palette, 85mm prime at f/2, shallow depth on the face, contemporary fashion editorial style. 4:5 aspect ratio. No visible brand names. Natural skin texture.
A glass jar of premium honey on a rustic kitchen counter next to a folded linen napkin and a wooden honey dipper, warm afternoon side-light from a window out of frame, deep amber liquid catching the light, 85mm lens at f/2.8, shallow depth of field, lifestyle product photography. The jar label reads "HIVE 24" in a serif typeface. 4:5 aspect ratio.
Frozen action shot of a long-distance runner mid-stride on a forest trail at golden hour, captured from a low side angle with the runner crossing left to right, motion blur on the trailing leg, warm directional backlight rim-lighting the silhouette, 200mm telephoto at f/4, heavy background compression, athletic editorial style. 16:9 aspect ratio. No visible brand logos.
Save these 12 as your skeleton library. Swap the subject and material descriptors and the rest of the prompt usually still works. The hardest 80% of the prompt is the lighting + camera + mood combination; that part is reusable across categories.
This is gpt-image-2's biggest superpower. Two years ago, asking any AI image model for a menu with correctly spelled items was a guaranteed failure. Now you can ship a print-ready restaurant menu, a multilingual product label, or a complete poster with display typography on the first try. ~95-99% character accuracy including Japanese, Korean, Chinese, Hindi, Bengali, and Arabic.
Always wrap the exact text you want rendered in quotation marks inside your prompt. The model treats quoted strings as literal targets. Unquoted text in the prompt is treated as description, not content to render.
A vintage poster that says spring sale at the top
A vintage poster with the headline "SPRING SALE" centered at the top in heavy sans-serif type, and the subtitle "30% off all denim through May 31" below in smaller serif text.
A single-page restaurant menu for a modern Mexican kitchen called "BARRIO 7", printed on warm cream paper, headlines in heavy condensed sans-serif, body text in classic humanist serif. Three sections: STARTERS with five items, MAINS with six items, COCKTAILS with four items. Each item has a name, a one-line description, and a price in US dollars. Subtle illustrated chile pepper graphic in the top-right corner. 8.5x11 portrait aspect ratio. Print-ready.
A minimal exhibition poster on a black background with the title "FORM & FOG" centered in massive thin serif type spanning the upper two-thirds, the subtitle "An exhibition of contemporary landscape photography. May 14 to August 30, 2026." in small monospaced type below, and "Boston Museum of Modern Art" in even smaller type at the bottom. Generous negative space. 2:3 aspect ratio.
A close-up product photograph of a small amber glass bottle of essential oil on a marble surface, label in pure white with the brand name "TERRA NOTE" in elegant thin serif at the top, the variant name "BERGAMOT & CEDAR" in heavy sans below, and "30 ml e / 1 fl oz" in tiny type at the bottom. Soft diffused daylight from the left. 4:5 aspect ratio. Photorealistic.
A literary fiction book cover titled "THE LAST QUIET PLACE" by Anna Reyes. Title in tall condensed serif spanning the middle, author name below in classic Trajan-style small caps. Background: a textured pale blue gradient with a single small black silhouette of a lighthouse in the lower-right corner. Publisher imprint "VESPER PRESS" at the very bottom in tiny tracked-out type. 6x9 portrait aspect ratio.
A minimalist Japanese ramen restaurant poster with the kanji title "霧の麺" rendered large in elegant brush calligraphy at the top, the romaji subtitle "Kiri no Men" below in delicate sans-serif, and the English line "FOG NOODLES — TOKYO" at the bottom. Background: a textured washi paper aesthetic in pale grey. Generous negative space. 2:3 aspect ratio.
A photograph of a hand-painted wooden cafe sign hanging above a shop entrance in Seoul. Korean hangul reads "조용한 시간" in confident brushed strokes, with the English translation "QUIET HOURS" in small lowercase serif below. Sign hangs from black iron brackets against a brick wall. Soft afternoon side-light. 1:1 aspect ratio. Photorealistic.
A contemporary architecture magazine cover. The masthead "DWELL" sits across the top in heavy custom display sans. The cover line "INSIDE THE NEW BAUHAUS — Eight homes that rewrite the rulebook" runs across the lower-left in mixed weights. Cover image is a wide architectural shot of a modernist concrete house at golden hour. Issue number "VOL 27 / NO 4 / MAY 2026" in small type at the bottom-right. 8.5x11 portrait aspect ratio.
An Instagram square graphic with the quote "We do not see things as they are. We see them as we are." in confident serif type, centered, broken into four lines across the middle of the canvas. Attribution "— ANAÏS NIN" in small caps below. Background: a soft warm cream gradient with subtle paper grain. 1:1 aspect ratio. No other elements.
A modern airport terminal wayfinding sign mounted on a white pillar, with three lines of information in clean grotesque sans-serif: "GATES A1-A24" with a right arrow, "BAGGAGE CLAIM" with a down arrow, and "RIDESHARE PICKUP" with a left arrow. Black type on white background, soft fluorescent ambient lighting, shallow depth of field. 4:5 aspect ratio. Photorealistic.
A bold gym promotional poster with the massive headline "MOVE BETTER" in heavy condensed sans-serif spanning the full width, the subtitle "Strength & conditioning. Mornings & evenings. No contracts." across the middle in regular weight, and the location "FIELD HOUSE — SOMERVILLE" at the bottom. Background: a deep matte black with a subtle athletic figure silhouette mid-jump. 4:5 aspect ratio.
An infographic header card with the title "THE STATE OF HEADLESS COMMERCE — 2026" in editorial sans-serif at the top, a one-sentence subtitle "How 1,200 brands made the jump in the last twelve months." in serif italic below, and four small statistic teasers across the bottom: "62% report higher conversion", "3.2x average build cost", "$180k median engineering investment", "47% migrated back within 18 months". 16:9 aspect ratio. Editorial magazine aesthetic.
A close-up of a printed concert ticket on a wooden table. The ticket reads, from top to bottom: "ANNA REYES" as the headline artist in heavy serif, "TUESDAY MAY 19 / DOORS 7PM / SHOW 8PM" in mono type, "PARADISE ROCK CLUB / BOSTON MA" in elegant sans-serif, and "GA STANDING / SEC FLOOR / $42" at the bottom. Subtle warm overhead lighting. 16:9 aspect ratio. Photorealistic.
A coffee bag photograph showing a matte kraft pouch with a full label design. Top line: "ORIGIN COFFEE CO." in small caps. Middle: "ETHIOPIA YIRGACHEFFE" in tall condensed serif. Below: "Single Origin / Washed Process / Roasted May 8 2026" in three lines of small mono type. Bottom: "340g e / 12 oz" with a small flag of Ethiopia icon. Studio lighting, 1:1 aspect ratio, photorealistic.
A product launch banner with the same announcement in four languages stacked vertically: English "NEW ARRIVAL", French "NOUVELLE COLLECTION", Japanese "新作登場", Korean "신상품 출시". Each language in its appropriate display typography. Background: a clean cream gradient. Below all four lines, the date "MAY 14 2026" in small monospaced type. 1:1 aspect ratio.
A photograph of a folded heather grey t-shirt laid flat on a wooden surface. The shirt's chest print reads "ASK BETTER QUESTIONS" in confident condensed sans-serif across two lines, with a small minimalist line-drawing of a thinking head icon beneath. Soft diffused daylight. 4:5 aspect ratio. Photorealistic. Print must be sharp and centered.
Three remaining gotchas to manage.
For anything print-bound or brand-critical, generate the image with placeholder text, then composite your real typography in Figma or Photoshop. gpt-image-2 nails the layout 95% of the time, which is the hard part. Final type swap is the easy part.
Thinking Mode is the agentic reasoning layer in gpt-image-2. Before generating a single pixel, the model plans the composition, optionally searches the web for live references, drafts multiple coherent variants, and self-checks the output. It costs more compute and latency. It pays back on complex prompts. This module is the playbook for when to use it.
Four mechanics run before the image renders:
| Prompt characteristic | Recommended mode | Why |
|---|---|---|
| Single subject, single composition | Instant | Reasoning overhead not worth latency |
| 5+ objects with spatial relationships | Thinking | Planning prevents object collisions and miscounts |
| Multi-panel comic, storyboard, batch of 4-8 images | Thinking | Continuity across batch only works in Thinking Mode |
| Reference to real product, place, or date past Dec 2025 | Thinking | Web search grounds factual accuracy |
| Infographic with dense text and data | Thinking | Self-verification catches text errors |
| Hero shot with one product | Instant | Faster and cheaper, quality already strong |
| 50 A/B variants of a headline | Instant | Volume work; reasoning overhead is wasted |
| Complex multilingual layout | Thinking | Composition planning + script accuracy benefits |
gpt-image-2 in Thinking Mode can handle 10-20 distinct objects with their traits and spatial relationships. Use this for explainer infographics, complex marketing scenes, and "everything I sell" hero shots.
A 4-row by 3-column grid on a white background containing these 12 items in order, left to right, top to bottom:
1. A matte black ceramic coffee mug
2. A vintage brass compass
3. A pair of tortoiseshell reading glasses
4. A leather-bound notebook with a fountain pen on top
5. A single ripe Bartlett pear
6. A small terracotta succulent in a hand-thrown pot
7. A folded grey cashmere scarf
8. A polished river stone the size of a fist
9. A glass jar of honey with a wooden dipper
10. A pair of brown leather Oxford shoes
11. A vintage silver pocket watch on a chain
12. A small stack of three hardcover books in earth tones
Soft diffused daylight from above. Subtle drop shadows. Editorial product photography. 4:3 aspect ratio. Items should be visually balanced in size.
An editorial infographic titled "BOSTON MARATHON 2026 — BY THE NUMBERS" with five large stat blocks: total registered runners, fastest finishing time, charity dollars raised, average age of participants, and number of countries represented. Use the actual 2026 figures by searching for them. Layout: title across the top in heavy condensed sans-serif, five stat blocks in a row across the middle each with a giant number and a one-line caption, source citation in tiny type at the bottom. Color palette: deep blue, gold, cream. 16:9 aspect ratio.
Generate 4 coherent images of the same character — a 30-something woman with shoulder-length auburn hair, wearing a cream linen blazer over a charcoal turtleneck, in different scenes:
1. Standing at a podium giving a keynote address, confident pose
2. Sitting at a wooden cafe table reading a book, soft afternoon light
3. Walking down a city sidewalk with a leather tote bag, three-quarter view
4. Laughing with friends at a dinner table, candid lifestyle shot
Maintain exact character likeness, clothing, and hair across all four images. Editorial photojournalism style. 4:5 aspect ratio each. Natural skin texture, no HDR processing.
A 6-panel comic strip in a 2-by-3 grid, drawn in clean line art with limited flat color. Recurring character: a small grey tabby cat. Panels in order:
1. Cat sitting on a windowsill watching rain outside
2. Cat batting at a falling leaf on the glass
3. Cat losing interest, looking inside the room
4. Cat noticing an empty food bowl in the kitchen
5. Cat sitting expectantly next to a human's leg
6. Cat eating from the now-full bowl, looking content
Same character design and color palette in every panel. Clean white gutters between panels. Editorial cartooning style. 1:1 aspect ratio overall.
Most casual users will never need Thinking Mode. Most professional users will use it 30-50% of the time and Instant Mode the rest. The discipline is matching the mode to the prompt complexity, not defaulting to "always Thinking" because it sounds better.
Generating a new image is the easy half. The hard half is fixing a 90%-there image without losing what already works. gpt-image-2's v1/images/edits endpoint supports mask-based inpainting, outpainting, and up to 16 reference images per call. This module is the editing playbook.
The images.edit endpoint handles all three. The mode is determined by what you pass in.
Important to understand: every gpt-image-2 edit is a full redraw under the hood. The model encodes your source image into tokens, generates a new complete token sequence, and decodes it into a new image. Even mask-based inpainting only adds constraints during the redraw — it does not literally overwrite pixels locally.
Four mechanisms preserve apparent consistency: visual token feature abstraction, global self-attention reference to the original, training inductive bias toward minimal change, and explicit planning in Thinking Mode.
Practical implication: micro-details (skin pores, exact fabric weave, precise reflections) can shift even when the model "preserves" a region. For brand-critical pixel-exact regions, composite the original back in via Photoshop.
[Edit endpoint with reference image]
Replace the background with a clean off-white seamless studio backdrop with a subtle drop shadow under the subject. Preserve the subject, its lighting, and its colors exactly. Add a soft floor reflection. No text, no other elements.
[Edit endpoint with reference image]
Remove the person standing on the left side of the frame. Reconstruct the wall, floor, and background that would naturally be visible behind them. Preserve everyone else, the lighting, and the overall composition. Do not add new elements.
[Edit endpoint with reference image]
This is a Spanish menu. Translate every menu item to English while preserving the typography, layout, paper texture, and prices exactly. Keep the section headers in their existing styling. Do not change the visual design of the menu.
[Edit endpoint with reference image]
Change the lighting from harsh midday to soft golden hour. Warm the color temperature, soften the shadows, add long directional shadows from the lower-left, and introduce a gentle haze in the background. Preserve the subject, composition, and all other elements.
[Edit endpoint with reference image and mask covering the jacket]
Change the masked jacket from olive green to deep navy blue. Match the original lighting, fabric texture, and material. Preserve every other element in the image including the face, hair, and background.
[Edit endpoint with reference image and mask extending the canvas]
Extend the canvas to a 16:9 aspect ratio by adding scene to the right of the existing image. Continue the wooden table, add a window in the upper-right, and place a small terracotta plant in the lower-right. Match the existing lighting and color grade exactly.
[Edit endpoint with 2 reference images: image A = a portrait, image B = a Wes Anderson film still]
Generate a new portrait of the subject in image A, restyled in the visual aesthetic of image B. Match the color palette, framing, and symmetry. Preserve the subject's facial identity and clothing. Do not literally composite the two — render a new image inspired by both.
[Edit endpoint with 4 reference images of the same character from different angles]
Generate a new image of this character in a different scene: walking through a Tokyo train station at night, wearing the same outfit shown in the references, captured in a wide candid documentary style. Preserve the character's facial features, hair, and clothing exactly. 3:2 aspect ratio.
[Edit endpoint with 3 reference images: image A = a blank coffee bag template, image B = a brand logo, image C = a typography sample]
Generate a finished coffee bag mockup using the bag shape from image A, the logo from image B placed in the upper-third, and the typography style from image C for the product name "EQUINOX BLEND" in the lower-third. Render as a photograph on a marble counter with soft daylight. 4:5 aspect ratio.
[Edit endpoint with reference image and mask covering existing text]
Replace the headline text in the masked region with "SPRING ARRIVALS" in the same typography style, weight, and color as the original. Preserve the surrounding image, layout, and any other elements exactly.
Important pricing note: edit requests always process reference images at high fidelity regardless of your quality parameter. This means edit requests cost more per call than pure text-to-image generation.
Two practical implications:
The "same character across multiple scenes" claim works well for simple subjects and gets less reliable as scene complexity grows. For series work (product line, character storyboard, recurring brand mascot), stack 4-8 reference images of the character from different angles and pass them all on every edit. Quality jumps significantly.
This is where gpt-image-2 produces the biggest practical ROI for working brands. Multi-format batch generation, brand-locked product photography at scale, 8-image carousels with character continuity, UGC-style ads, and before/after pairs. This module is the production playbook.
Generate one hero image, then request platform-specific variants from the same source. Each variant preserves the headline text and brand color cues.
Generate the same marketing campaign in four aspect ratios, with the same product, the same headline, and the same brand identity across all four:
Subject: A premium glass water bottle in matte sage green with the brand name "WELL & FLOW" on a minimal white label.
Headline text: "DRINK BETTER. EVERY HOUR."
Brand palette: Sage green, off-white, deep charcoal.
Tone: Calm, premium, intentional.
Four outputs needed:
1. 1:1 square for Instagram feed
2. 9:16 vertical for Instagram Story / TikTok
3. 16:9 horizontal for YouTube thumbnail
4. 4:5 portrait for Pinterest / shopping ad
Maintain identical typography, color, and product appearance across all four. Composition should adapt to each aspect ratio — do not just crop the same image.
[Thinking Mode] Generate 8 coherent product photography images of the same skincare line for a launch carousel. Brand: "FOREST ROOM". Three products: a cream cleanser, a botanical serum, and a moisturizing balm — all in matching amber glass with cream labels.
Image 1: All three products together on a marble surface, flat lay with eucalyptus leaves
Image 2: Cleanser in use, hands lathering at a sink with warm morning light
Image 3: Serum dropper close-up, a single drop captured mid-air
Image 4: Balm jar open showing texture, top-down macro
Image 5: A model's reflection in a bathroom mirror, holding the serum
Image 6: Hero shot of all three products vertically aligned on a wooden shelf
Image 7: Texture macro of all three formulas swatched side by side on skin
Image 8: Lifestyle shot of the products on a bedside table with morning sun
Maintain exact brand identity across all 8 images: same amber glass, same cream labels, same color grading, same warm minimalist aesthetic. 4:5 aspect ratio each.
A UGC-style smartphone photograph (not professional studio) of a person holding up a wireless earbud case at arm's length in front of a casual home setting. Slight phone camera imperfections — minor lens flare, slightly over-exposed, casual framing. The earbuds case is white with a small brand logo. The person is mid-30s, in a casual sweatshirt, smiling. Natural daylight from a window. Looks like a real customer post, not an ad. 4:5 aspect ratio.
A single 1:1 image split into two halves with a thin vertical dividing line. Left half labeled "BEFORE" in small caps at the top — a cluttered home office: papers stacked, tangled cables, sticky notes everywhere, dim lighting. Right half labeled "AFTER" at the top — the same office, now organized: clean desk, single monitor, hidden cables, a single plant, soft daylight. Same room, same furniture, same window — only the organization changes. Photorealistic with natural lighting. Maintain a consistent perspective and time of day across both halves.
A wide editorial shot of a single ceramic pour-over coffee dripper sitting on a wooden counter in a quiet morning kitchen. Steam rising from the carafe below. Soft window light streaming from the left, warm color temperature. A simple cream-tile backsplash, no clutter. 3:2 aspect ratio. The coffee dripper is the unmistakable focal point but the scene reads as a lived-in moment, not a product catalog. Editorial lifestyle photography. Photorealistic with natural skin texture if any hands are visible.
A 1:1 Instagram social card with a deep forest green background and the headline "SUSTAINABILITY IS A SUPPLY CHAIN, NOT A SLOGAN." in large confident serif, broken into three lines, centered, cream-colored type. Below the headline, the small attribution "— OUR Q1 LETTER" in tracked-out monospace small caps. Generous negative space top and bottom. Brand mark "FOREST ROOM" in tiny lowercase at the very bottom center.
A 21:9 ultrawide horizontal hero banner. Left two-thirds: a cinematic editorial photograph of a couple walking on a foggy New England coastline at golden hour, shot from behind, wide angle. Right one-third: a clean off-white panel containing the headline "MAY ARRIVALS — COASTAL ESSENTIALS" in heavy condensed serif, the subhead "Hand-knit sweaters, weatherproof outerwear, and waxed cotton bags. Now in stock." in regular body, and a "SHOP THE COLLECTION" CTA button rendered as a black rounded rectangle with cream type. Photorealistic on the image side, clean editorial layout on the text side.
A 600x300 pixel email header for a brand newsletter. Left half: a single product photograph of a leather-bound notebook on a desk with a pen. Right half: cream solid panel with "ISSUE 17" in tiny mono, "ON WRITING" as the main headline in tall serif, and "Weekly notes on craft, process, and the long game." as a one-line subhead. Editorial newsletter aesthetic. Newsletter brand "PARAGRAPH 7" in tiny type at the bottom-left.
Generate the same product — a brushed aluminum water bottle in three colorways: graphite, sage, and bone — as three separate 1:1 product-on-white images. Each image: bottle centered, soft even softbox lighting, subtle drop shadow, slight floor reflection, off-white background. Maintain identical lighting, camera angle, and overall composition across all three so they look like a consistent product catalog set. No text or branding visible. Photorealistic.
For agency or production work where 100% brand consistency matters, prepend this brand lock to every prompt in the session. Pair it with the Brand Voice Skill in the Skills Library section.
BRAND LOCK — apply these to every output in this session:
Brand: [BRAND NAME]
Palette: [hex codes — primary, accent, neutral, background]
Typography style: [display family character + body family character]
Photography style: [editorial / documentary / studio / lifestyle / catalog]
Lighting bias: [warm / neutral / cool] + [hard / soft / mixed]
Color grading: [subtle / saturated / desaturated / warm / cool]
Tone: [3 adjectives max]
Banned: [generic phrases, banned visual cliches, banned compositions]
Required: [logo placement rules, required negative space, required readability]
The hardest test for any image model: keep a single character looking like themselves across many scenes. gpt-image-2 is the strongest model on the market for this as of May 2026, with two limits: simple subjects work better than complex ones, and exact facial likeness of a real person remains a hard wall. This module is the character pipeline.
Step 1 of every character pipeline: generate a clean reference sheet. Step 2: re-use that reference sheet as input to every subsequent prompt.
A character reference sheet for an animated story. Single character, displayed in five views on a neutral grey background:
1. Front view, neutral expression
2. Three-quarter view from the left, neutral expression
3. Side profile facing right
4. Back view
5. Front view with a smile
Character details:
- Late-20s woman
- Shoulder-length wavy auburn hair, parted on the left
- Hazel eyes
- Pale freckled skin, natural texture
- Wearing a cream linen blazer over a charcoal turtleneck and dark indigo jeans
- Tan leather ankle boots
- Small gold hoop earrings, no other jewelry
Each pose at the same height, evenly spaced, full body visible in each. Clean concept-art style. Soft even studio lighting. 16:9 aspect ratio with all five poses in a single horizontal row.
An expression sheet for the same character (use the reference sheet as input). Display the character's face from the same three-quarter angle in 8 different expressions, arranged in a 2-by-4 grid:
1. Neutral
2. Smiling warmly
3. Laughing
4. Concentrating
5. Concerned
6. Surprised
7. Skeptical
8. Tired
Maintain exact facial features and hair across all 8. Clean character design style. Soft studio lighting. 16:9 aspect ratio.
A costume design sheet for the same character in 4 different outfits, full body, front view, same neutral pose across all 4:
1. Business: tailored navy blazer, white silk shell, grey trousers, low black heels
2. Casual: oversized grey hoodie, dark jeans, white sneakers
3. Outdoor: olive waxed cotton jacket, plaid flannel shirt, cuffed denim, leather hiking boots
4. Evening: deep emerald slip dress, gold drop earrings, strappy heels
Maintain exact facial features and hair across all 4 outfits. Each outfit labeled in small mono caps below the figure. Clean concept-art style. 16:9 aspect ratio with all 4 in a horizontal row.
[Thinking Mode, with character reference sheet as input]
Generate 6 photographic scenes of the same character, maintaining exact facial features, hair, and the Business outfit (navy blazer, white shell, grey trousers, low black heels) throughout:
1. Arriving at an office building entrance, morning, three-quarter view
2. In a glass-walled conference room presenting at a whiteboard, mid-shot
3. Walking through an open-plan office carrying a coffee cup, candid
4. Sitting at a desk with a laptop, mid-afternoon, side profile
5. Stepping out of the building at dusk with a coat over her arm
6. At a casual after-work bar with two colleagues, laughing
Editorial corporate photography style. Natural skin texture, no HDR processing. 4:5 aspect ratio each. Same lighting feel throughout the day's arc.
[Thinking Mode]
A 9-panel storyboard in a 3-by-3 grid, drawn in clean black-and-white ink wash. Recurring protagonist: a young man with short dark hair, round glasses, wearing a wool overcoat. Story:
Panel 1: Wide establishing shot, the protagonist walking down a foggy street at night
Panel 2: Close-up on his face, looking at something off-frame
Panel 3: Reveal of what he sees — a glowing window in an old library
Panel 4: He approaches the library door, hand on the brass handle
Panel 5: Interior: shelves of books, single lamp, an elderly librarian
Panel 6: The librarian looks up, smiles, gestures to a chair
Panel 7: Close-up of an open leather-bound book on a wooden table
Panel 8: The protagonist reading, candle nearby, deep in thought
Panel 9: Wide pull-back through the window, library lit warm against the foggy night
Maintain exact character design across all 9 panels. Cinematic graphic novel style. 1:1 aspect ratio overall.
Three concept art views of a fictional handheld GPS device for outdoor explorers. Industrial design aesthetic: titanium body, matte black bezel, single circular monochrome e-ink display, two physical buttons. Brand mark: a small embossed compass icon. Three views in a single horizontal row on a clean white background:
1. Front view, screen on, displaying a topographic map
2. Three-quarter view from the right, showing the side button and lanyard loop
3. Back view, showing the textured grip and a small data plate
Studio product photography, soft directional light. Editorial industrial design portfolio aesthetic. 16:9 aspect ratio.
gpt-image-2 will refuse to render exact likenesses of real public figures or named celebrities. Workarounds for legitimate use cases:
Character consistency is excellent for simple subjects (one person, clear outfit, predictable scene). It degrades as scene complexity grows. For brand-critical character work across 20+ assets, plan to do final brand-pass cleanup in Photoshop. The model gets you 90% of the way; the last 10% is hand work.
Everything above runs in ChatGPT.com. For volume, automation, integrations, and pipelines, you need the API. This module is the engineering reference: authentication, request shape, rate limits, cost discipline, error handling, and the patterns that scale.
Four endpoints touch gpt-image-2:
POST /v1/images/generations — text-to-imagePOST /v1/images/edits — image-to-image with optional mask and up to 16 referencesPOST /v1/responses — newer unified endpoint with reasoning trace accessPOST /v1/chat/completions — image generation called as a tool from a chatfrom openai import OpenAI
client = OpenAI()
response = client.images.generate(
model="gpt-image-2-2026-04-21", # pin to snapshot for production
prompt="Studio product photography of a wireless over-ear headphone in matte graphite, centered on a clean off-white seamless background, soft even softbox lighting from above, subtle drop shadow, 85mm prime lens.",
size="1024x1024",
quality="medium", # low | medium | high
n=1, # 1 to 10
)
# Image returned as a URL or base64 depending on response_format
image_url = response.data[0].url
print(image_url)
with open("product.png", "rb") as image_file, \
open("mask.png", "rb") as mask_file:
response = client.images.edit(
model="gpt-image-2",
image=image_file,
mask=mask_file, # transparent areas define edit region
prompt="Replace the masked background with a clean seamless off-white studio backdrop. Preserve the subject exactly.",
size="1024x1024",
# Quality is fixed at high for image inputs regardless of this parameter
)
print(response.data[0].url)
# Up to 16 reference images
reference_files = [open(p, "rb") for p in [
"char_front.png",
"char_three_quarter.png",
"char_side.png",
"char_back.png",
]]
response = client.images.edit(
model="gpt-image-2",
image=reference_files, # list of file handles
prompt="Generate a new image of this character walking through a foggy Tokyo train station at night, in the same outfit shown in the references. Editorial cinematography. 3:2 aspect ratio.",
size="1536x1024",
)
for f in reference_files:
f.close()
print(response.data[0].url)
import OpenAI from "openai";
const client = new OpenAI();
const response = await client.images.generate({
model: "gpt-image-2-2026-04-21",
prompt: "Editorial photojournalism of a master watchmaker at a wooden bench in a Geneva atelier, soft window light, 35mm lens, f/2.8, natural skin texture.",
size: "1024x1536",
quality: "high",
n: 1,
});
console.log(response.data[0].url);
Some third-party hosts (fal.ai, WaveSpeed, etc.) wrap gpt-image-2 in an async queue. Pattern:
// Submit
const { request_id } = await fal.queue.submit("fal-ai/gpt-image-2", {
input: { prompt: "..." },
webhookUrl: "https://your-server.com/webhook",
});
// Poll for status
const status = await fal.queue.status("fal-ai/gpt-image-2", {
requestId: request_id,
logs: true,
});
// Fetch result when ready
const result = await fal.queue.result("fal-ai/gpt-image-2", {
requestId: request_id,
});
Plan the tier ramp before you launch. Limits are per minute for image generation:
| Tier | Images/Min | Unlock Threshold |
|---|---|---|
| Tier 1 | 5 | Default for new accounts |
| Tier 2 | 20 | $50 spent + 7 days since first payment |
| Tier 3 | 50 | $100 spent + 7 days |
| Tier 4 | 100 | $250 spent + 14 days |
| Tier 5 | 250 | $1,000 spent + 30 days |
from openai import OpenAI, APIError, RateLimitError, BadRequestError
import time
client = OpenAI()
def generate_with_retry(prompt: str, max_retries: int = 3) -> str:
for attempt in range(max_retries):
try:
response = client.images.generate(
model="gpt-image-2-2026-04-21",
prompt=prompt,
size="1024x1024",
quality="medium",
)
return response.data[0].url
except RateLimitError as e:
wait = (2 ** attempt) + 1
print(f"Rate limited, waiting {wait}s")
time.sleep(wait)
except BadRequestError as e:
# Likely content policy refusal - no retry will help
print(f"Refused: {e.message}")
return None
except APIError as e:
if attempt == max_retries - 1:
raise
time.sleep(2)
return None
Log every request with at minimum:
Without these, you cannot diagnose incidents or forecast monthly burn rate accurately.
Before you flip the switch on a real pipeline: pin to snapshot, validate rate limit headroom at peak traffic, implement retry with exponential backoff, log every request, set a daily spend cap on your OpenAI account, and run a 1-week pilot at your real workload before forecasting your monthly bill. Multiply the weekly burn by 4.3 and add a 20% buffer.
gpt-image-2 is the strongest general-purpose image model in May 2026. It is not the strongest at everything. This module is the honest decision matrix: where each competitor wins, where gpt-image-2 hits walls, and how to build a multi-model workflow.
| Axis | gpt-image-2 | Nano Banana 2 | Midjourney V8 | Stable Diffusion XL+ |
|---|---|---|---|---|
| Text rendering | Best (~99%) | Mid (~70%) | Weak | Weak (better with LoRA) |
| Artistic style | Strong | Strong | Best | Strong (with fine-tunes) |
| Instruction following | Best | Strong | Weakest (interprets liberally) | Mid |
| Reference-driven editing | Best (16 refs) | Strong | Weak | Strong (ControlNet) |
| Web search grounding | Yes (Thinking Mode) | Yes (Google Search) | No | No |
| Cost at scale | Moderate | Best | Subscription only | Best (self-host) |
| Content policy strictness | Strict | Strict | Moderate | Permissive (self-host) |
| Transparent backgrounds | No (yet) | Yes | No | Yes (with workflow) |
Be honest about where the model still fails. Pretending otherwise wastes time.
Most professional shops in May 2026 use 2-3 models in combination:
If you can only learn one model in 2026, learn gpt-image-2. The text rendering, the instruction following, the multi-image edits, the 8-image batches — these alone replace 80% of what working teams were doing across multiple models a year ago. Keep Midjourney for atmospheric work, keep Stable Diffusion for self-hosted needs, but make gpt-image-2 your daily driver. The rest of this masterclass — the 100 prompts, the 8 templates, the 5 skill packages — assumes that decision is already made.
One hundred production-tested prompts across ten categories. Each follows the 7-Part Prompt Formula from Module 2. Replace bracketed placeholders with your specifics. Save the ones that work for your brand as your personal prompt library.
Studio-quality photoreal output. Pair with explicit lighting and lens vocabulary from Module 3.
Editorial portrait of a [age] [gender] [profession] in [setting], shot on 85mm prime at f/2, soft window light from the [direction], natural skin texture with visible pores, slight desaturation, no HDR processing. 4:5 aspect ratio.
Wide environmental shot of [location] at [time of day], 24mm wide-angle at f/8, deep focus, [warm/cool/neutral] color temperature, subtle atmospheric haze, documentary photography style. 3:2 aspect ratio.
Macro detail shot of [subject] showing [specific texture or surface], 100mm macro lens at f/5.6, single hard light source from above-right, sharp focus on [focal point], frozen motion, beauty-editorial aesthetic. 1:1 aspect ratio.
Three-quarter front view of a [object] on a [surface] in [environment], 50mm prime at f/2.8, golden hour side light, shallow depth of field, editorial commercial photography. 4:5 aspect ratio.
Frozen action moment of [subject] mid-[action], 200mm telephoto at f/4, heavy background compression, warm rim backlight, slight motion blur on [moving element], athletic editorial style. 16:9 aspect ratio.
Candid documentary moment of [scene description] in [location], 35mm lens at f/2.8, available light only, slight grain, no posing, photojournalism aesthetic. 3:2 aspect ratio. Natural skin texture if humans visible.
Symmetric architectural interior of [space], 24mm tilt-shift lens corrected verticals, soft north-facing daylight, neutral cool palette, generous negative space, editorial architecture photography. 3:2 aspect ratio.
Top-down flat lay of [items list] arranged on [surface] with [filler elements], 50mm directly overhead, soft diffused daylight, muted [palette] palette, editorial product styling. 1:1 aspect ratio.
Cinematic film still of [character] in [location] at [time of day], anamorphic 2.39:1, soft directional key light, deep contrast, color-graded for editorial film aesthetic. 21:9 aspect ratio.
Hands-only lifestyle shot of [action with product] on [surface], 50mm at f/2, warm morning side-light from the left, shallow depth on the hands, lifestyle brand photography, face not visible. 4:5 aspect ratio.
E-commerce-ready, catalog-grade product shots. Default to clean lighting and subtle drop shadows.
Studio product photography of [product] in [color/material], centered on a clean off-white seamless background, soft even softbox lighting from above, subtle drop shadow, soft floor reflection, 85mm prime lens. 1:1 aspect ratio. No text, no watermarks.
Hero shot of a single [product] floating against a deep [color] gradient background, rim lighting from behind, 100mm macro at f/8, glossy floor reflection, premium [category] aesthetic, soft volumetric haze. 4:5 aspect ratio.
Three-quarter view of [product] on a [surface material] surface with [contextual prop], soft window light from the upper-left, 85mm at f/2.8, shallow depth on the product. Editorial lifestyle context, not pure catalog. 4:5 aspect ratio.
Same [product] shown in [number] colorways arranged in a horizontal row, identical lighting and camera angle across all variants for consistent catalog look. Soft softbox lighting, off-white background, subtle drop shadows. 16:9 aspect ratio.
Macro detail of [product]'s [specific component or texture], 100mm macro at f/5.6, single directional light source, sharp focus on [detail], premium product photography aesthetic. 1:1 aspect ratio.
Lifestyle context shot of [product] being used by [hands or person] in [environment], 35mm at f/2.8, warm golden hour light, candid moment not posed. Photorealistic with natural skin texture. 4:5 aspect ratio.
Exploded view of [product] showing its [number] components separated and floating in space, clean white background, soft even lighting, technical product illustration aesthetic with photorealistic rendering. 1:1 aspect ratio.
In-context packaging shot of [product] sitting on [contextual surface] with [supporting props], soft afternoon side-light, 85mm at f/2.8. The [label/box] is clearly readable with [brand name] in [typography style]. 4:5 aspect ratio.
Hands-in-frame demonstration shot of [product action], shallow depth of field on the action, soft window light, editorial product photography. Hands visible but face not visible. 4:5 aspect ratio.
Product line family portrait of [product family with variants] arranged in [composition — pyramid, row, grid], identical brand identity across all items, clean studio lighting, premium brand catalog aesthetic. 16:9 aspect ratio.
gpt-image-2's strongest superpower. Always wrap exact text in quotation marks.
A minimal poster with the headline "[YOUR HEADLINE]" centered in massive thin serif type spanning the upper two-thirds, the subtitle "[YOUR SUBTITLE]" in small monospace below, and "[YOUR FOOTER]" in tiny tracked-out type at the bottom. [Background color] background, generous negative space. 2:3 aspect ratio.
A single-page menu for [restaurant name and concept]. Headlines in [type style], body in [type style]. Three sections: [section 1], [section 2], [section 3]. Each item has name, description, and price. Subtle illustrated motif in the corner. 8.5x11 portrait. Print-ready.
A book cover titled "[BOOK TITLE]" by [author name]. Title in [typography style] spanning the middle, author below in [secondary typography]. Background: [visual treatment] with [single graphic element]. Publisher "[IMPRINT]" at the bottom in tiny type. 6x9 portrait.
A product label photographed on the product itself. Top line "[BRAND NAME]" in [style]. Middle "[PRODUCT NAME]" in [style]. Bottom "[VARIANT / WEIGHT / DATE]" in small mono. Studio lighting, photorealistic, 1:1 aspect ratio.
A magazine cover for "[MAGAZINE NAME]". Masthead across the top in heavy display sans. Cover line "[HEADLINE]" running across the lower-third. Hero image: [image description]. Issue indicator "[VOL/NO/DATE]" in small type. 8.5x11 portrait.
A multilingual launch banner with the same announcement stacked vertically in [number] languages: [language 1] "[TEXT]", [language 2] "[TEXT]", [language 3] "[TEXT]". Each language in appropriate display typography. Background: [palette]. 1:1 aspect ratio.
A wayfinding sign mounted on a [material] post or wall, reading "[DESTINATION]" with a [direction] arrow, in clean grotesque sans-serif. [Color combination] type. Soft ambient lighting, photorealistic. 4:5 aspect ratio.
A 1:1 Instagram quote graphic with "[QUOTE TEXT IN QUOTES]" in [typography style], broken into [number] lines across the middle. Attribution "— [ATTRIBUTION]" below in small caps. Background: [color/gradient]. No other elements.
A printed ticket close-up on a [surface]. Reads top to bottom: "[HEADLINE]" in heavy serif, "[DATE/TIME]" in mono, "[VENUE/LOCATION]" in elegant sans, "[SEATING/PRICE]" at bottom. Soft overhead lighting. Photorealistic. 16:9 aspect ratio.
A bold promotional poster with massive headline "[HEADLINE]" in heavy condensed sans spanning the full width, subtitle "[SUBHEAD]" across the middle in regular weight, and location "[VENUE/LOCATION]" at the bottom. Background: [visual treatment]. 4:5 aspect ratio.
Campaign-ready hero shots, social cards, and brand-locked visuals.
A hero campaign image for [brand and product]. Headline "[HEADLINE]" overlaid in [typography]. Subject: [product/scene] in [setting] with [lighting]. Tone: [3 adjectives]. Brand palette: [palette]. 16:9 aspect ratio.
A 1:1 Instagram square with [product or scene] centered, headline "[HEADLINE]" in upper-third, CTA "[CTA TEXT]" rendered as a [button style] at the bottom. Brand palette: [palette]. Photorealistic.
A 9:16 vertical Story/TikTok creative with [product] as the focal point, headline "[HEADLINE]" stacked across the top, social handle "[@HANDLE]" in the bottom-right. Mobile-first composition. Brand palette: [palette].
A 21:9 ultrawide hero banner. Left two-thirds: cinematic editorial photo of [scene]. Right one-third: clean panel with "[HEADLINE]" in heavy serif, "[SUBHEAD]" below, and "[CTA]" as a button. Brand palette: [palette].
A before-and-after split image. Left "BEFORE": [problem state description]. Right "AFTER": [improved state description]. Same setting, same perspective, same time of day, only [variable] changes. Photorealistic with consistent lighting.
A UGC-style smartphone photo of [person profile] holding [product] at arm's length in [casual setting]. Slight phone camera imperfections, natural daylight, candid framing. Looks like a real customer post not an ad. 4:5 aspect ratio.
An email header 600x300 pixel. Left half: [product photograph]. Right half: cream panel with "[ISSUE NUMBER]" tiny mono, "[HEADLINE]" tall serif, "[SUBHEAD]" body. Brand "[NEWSLETTER NAME]" at the bottom.
A square social card with [background palette] and the headline "[STATEMENT TEXT]" in large [type style], broken into [number] lines, centered. Attribution "— [SOURCE]" below in small caps. Brand mark "[BRAND]" at the bottom.
A launch announcement creative for [product] showing the product hero shot with "[NEW]" or "[LAUNCH DATE]" badge in upper-right corner, headline "[HEADLINE]" overlaid, brand palette throughout. Editorial e-commerce aesthetic. 4:5.
A seasonal campaign image evoking [season] for [brand]. Subject: [season-appropriate scene] with [product] integrated naturally. Color palette: [seasonal palette]. Mood: [3 adjectives]. 16:9 aspect ratio.
Character pipelines. Reference sheets first, scene generation second.
A character reference sheet showing [character description] in 5 views in a single row: front, three-quarter left, side, back, smiling front. Neutral grey background, soft even studio lighting. 16:9 aspect ratio.
An expression sheet for [character], same three-quarter angle, 8 expressions in 2x4 grid: neutral, smiling, laughing, concentrating, concerned, surprised, skeptical, tired. Maintain exact facial features across all. 16:9.
A costume design sheet for [character] in [number] outfits, full body front view, neutral pose across all. Each outfit labeled in small mono caps below the figure. Maintain exact facial features. 16:9 horizontal row.
[Thinking Mode with reference sheet] Generate [number] scenes of [character], maintaining exact facial features and [outfit]: scene 1: [description], scene 2: [description], scene 3: [description]. Editorial photography style. 4:5 each.
[Thinking Mode] A 9-panel storyboard 3x3 grid in [style — ink wash / clean line / watercolor]. Recurring character: [character]. Story panels: [9 panel descriptions]. Maintain exact character design across all. 1:1 overall.
Concept art views of a fictional [object type] showing [number] views in a horizontal row on a clean white background: [view 1], [view 2], [view 3]. Industrial design portfolio aesthetic. Studio product photography. 16:9.
A character key art piece showing [character] in a dramatic [setting] with [dramatic lighting]. Mid-shot composition, [emotion] expression, [costume]. Cinematic film poster aesthetic. 2:3 portrait aspect ratio.
A side-by-side comparison of [character] in [number] different lighting setups for a film lookbook: [setup 1 — e.g., golden hour], [setup 2 — e.g., blue hour], [setup 3 — e.g., harsh midday]. Maintain identical character and costume.
A brand mascot character design sheet for [brand]. Character traits: [physical descriptors], personality: [3 adjectives], color palette: [palette]. Show the character in 4 poses: hero pose, action pose, friendly pose, thinking pose. Clean illustration style.
A multi-character group portrait showing [number] characters from a story side-by-side, each with distinct [physical features and costume]. Editorial group photo aesthetic, soft even studio lighting, neutral background. 21:9.
Use Thinking Mode for fact-grounded infographics. The model can web-search for current data.
[Thinking Mode] An editorial infographic titled "[TITLE]" with [number] large stat blocks across the middle, each with a giant number and one-line caption. Web-search for current [topic] figures. Source citation in tiny type at the bottom. 16:9.
A process diagram showing [number] steps in a horizontal flow. Each step has a number, a one-line label, and a small icon. Connecting arrows between steps. Brand palette: [palette]. 21:9 aspect ratio.
A side-by-side comparison infographic of [option A] vs [option B] with [number] comparison rows. Each row labeled with a category, with A and B columns showing values or checkmarks. Editorial design aesthetic. 4:5 portrait.
A timeline infographic showing [topic] across [date range] with [number] key milestones plotted along a horizontal axis. Each milestone has a date, label, and brief description. Brand palette: [palette]. 21:9.
A pie chart or donut chart infographic showing [topic breakdown] with [number] segments. Each segment labeled with category and percentage. Clean editorial design. Brand palette: [palette]. 1:1 aspect ratio.
A bar chart comparing [number] [items] across a single metric. Y-axis labeled "[metric]", X-axis labeled with items. Tallest bar in accent color. Clean editorial chart aesthetic. 16:9 aspect ratio.
A geographic infographic showing [data] mapped across [region]. Heat map style with darker shades indicating higher values. Legend in the bottom-right corner. Brand palette: [palette]. 4:3 aspect ratio.
A how-it-works diagram of [system/process] showing components labeled with arrows pointing to their location. Technical illustration style with clean line art. Background: cream paper texture. 4:3 aspect ratio.
A decision tree infographic for "[QUESTION]" with [number] branches and leaf outcomes. Clean flowchart style, each node a rounded rectangle, connecting lines clearly visible. Editorial diagram aesthetic. 16:9.
A statistic hero card with a single massive number "[NUMBER]" centered, label "[METRIC NAME]" below in smaller type, and one-line context "[CONTEXT]" at the bottom. Background: bold solid color. 1:1.
Mobile and web UI mockups with readable interface text. Thinking Mode helps with layout precision.
A mobile app screen mockup for [app concept], shown on a [device model] mockup. Screen shows [screen description with specific UI elements]. Clean modern design language, [palette]. 9:16 aspect ratio.
A web dashboard mockup shown in a browser frame. Sidebar nav on left with [items], main panel showing [content], header with [user info]. Clean SaaS design aesthetic. 16:9 aspect ratio.
A mobile app onboarding flow showing 3 screens side-by-side: screen 1 [welcome], screen 2 [feature highlight], screen 3 [sign up CTA]. Each on a phone mockup. Brand: [name]. 16:9 horizontal layout.
An e-commerce product detail page mockup shown in a browser. Product image left, details right with title "[PRODUCT]", price "[PRICE]", description text, "ADD TO CART" button. Clean modern e-comm aesthetic. 16:9.
A SaaS pricing page mockup with [number] tiers side-by-side. Each tier card has tier name, price, feature checklist, CTA button. Recommended tier highlighted. Clean editorial design. 16:9 aspect ratio.
A mobile messaging interface mockup showing a conversation between [two parties] about [topic]. Show last 5 messages with timestamps. Phone status bar visible at top. 9:16.
A settings screen UI for [app name] showing toggleable preferences: [setting 1], [setting 2], [setting 3]. Each row has label, current value, and chevron. Clean iOS or Android-style design. 9:16.
A web checkout flow mockup showing 3 steps: cart review, shipping, payment. Each step header with step indicator. Form fields clearly labeled. Brand: [name]. Clean e-comm design. 16:9.
An analytics dashboard mockup showing [number] chart widgets: [chart 1 description], [chart 2 description], [chart 3 description]. Side panel with filters. Header with date range picker. 16:9.
A device mockup hero shot showing [app name] running on [phone + laptop + tablet] in a creative composition. App's hero screen visible on each device. Soft studio lighting, premium tech product photography. 16:9.
Scroll-stopping social content for Instagram, TikTok, LinkedIn, Pinterest.
An Instagram carousel cover slide with bold question "[QUESTION]" in heavy display type, intriguing visual hook in background, "SWIPE →" indicator in bottom-right. Brand: [name]. 1:1.
A LinkedIn post graphic with editorial business aesthetic. Headline "[INSIGHT]" in clean sans, supporting subhead, author byline with role at bottom. Brand palette: [palette]. 4:5 portrait.
A TikTok thumbnail 9:16 with explosive visual hook, massive text overlay "[HOOK COPY]" in white outlined heavy sans, subject [description] dominating frame. Mobile-optimized composition.
A Pinterest pin 2:3 portrait, recipe or how-to card style. Top image: [hero shot]. Bottom panel: title "[TITLE]" in heavy sans, source URL "[URL]" in small mono. Brand: [name].
An Instagram Reel cover 9:16 with subject [description] in dynamic action pose, text overlay "[HOOK]" massive at top, brand handle "@[HANDLE]" at bottom-right. High contrast for thumbnail visibility.
A Twitter/X header banner 3:1 ultrawide with editorial brand mood image, brand mark "[BRAND]" in upper-left, optional tagline "[TAGLINE]" across the lower-third.
A YouTube thumbnail 16:9 with subject's face dominating left half showing strong emotion, massive text "[HOOK]" filling right half in bold outlined sans, high contrast palette for click-through optimization.
An Instagram Story 9:16 quote card with text "[QUOTE]" in serif, attribution "— [SOURCE]" below in small caps, background: [palette gradient], swipe-up indicator at bottom.
A 1:1 testimonial graphic with customer quote "[QUOTE]" in serif, customer photo circle at bottom-left, customer name "[NAME]" and role "[ROLE]" right of photo, brand mark top-right.
A series of 5 connected Instagram Story frames each 9:16, telling a sequential story about [topic]. Frame 1: hook. Frames 2-4: content. Frame 5: CTA. Consistent visual style across all 5.
Print-ready posters, event flyers, packaging. Use Thinking Mode for complex multi-element layouts.
An event poster for "[EVENT NAME]" on [date] at [venue]. Massive title in [type style], date and venue below in smaller mono, hero illustration or photo dominating the upper half. 2:3 portrait. Print-ready.
A movie-poster-style key art for fictional film "[TITLE]". Main character mid-frame, dramatic [lighting] background, title in massive display below subject, tagline above title, credit block at bottom. 27x40 portrait.
A vintage-style travel poster for [destination] in 1960s mid-century illustration style. Title "[DESTINATION]" in tall thin serif at top, illustrated landscape in middle, subtitle "[TAGLINE]" at bottom. 2:3.
A minimal exhibition poster on [background color]. Title "[EXHIBITION TITLE]" massive thin serif spanning upper two-thirds. Subtitle "[VENUE / DATE]" small mono below. Generous negative space. 2:3.
A product packaging design for [product] in [container type]. Brand "[BRAND]" prominent, product name "[PRODUCT]" below, ingredient or feature callouts on side panel, regulatory text at bottom. Photorealistic mockup. 4:5.
A vinyl record album cover for fictional artist "[ARTIST]" album "[ALBUM]". Style: [music genre aesthetic]. Cover art: [description]. Artist name and album title in [typography]. 1:1 aspect ratio.
A business card design for [name] at [company], [role]. Front: name, role, contact details. Back: brand mark dominating. Clean editorial design, [palette]. Standard 3.5x2 inch card mockup.
A wine label design for "[VINEYARD]" [variety] [vintage]. Vineyard name in elegant serif at top, illustrated motif of [landscape element], variety name and vintage below, regulatory text in tiny type. Print-ready. 4:5.
A magazine spread mockup showing two facing pages. Left page: full-bleed photograph. Right page: editorial article layout with headline "[HEADLINE]", deck "[DECK]", body copy in 2 columns, pull quote. 16:9 spread.
A typographic poster with only text — no images. Massive quote "[QUOTE]" in display serif filling most of the canvas, attribution in small caps at bottom. Background: solid [color]. Editorial design. 2:3 portrait.
For the images.edit endpoint. Each assumes one or more reference images attached.
[Edit endpoint, 1 reference] Replace the background with a clean off-white seamless studio backdrop with subtle drop shadow under the subject. Preserve subject, lighting, and colors exactly.
[Edit endpoint, 1 reference] Remove the [object/person] in the [location] of the frame. Reconstruct the background that would naturally be visible. Preserve everything else exactly.
[Edit endpoint, 1 reference + mask] Change the masked [element] from [current color] to [new color]. Match original lighting, texture, and material. Preserve every other element including face, hair, and background.
[Edit endpoint, 1 reference] Change the lighting from [current condition] to [target condition]. Adjust color temperature, shadow direction, and atmospheric quality accordingly. Preserve subject and composition.
[Edit endpoint, 1 reference + outpaint mask] Extend the canvas to [new aspect ratio] by adding scene to the [direction]. Continue the existing environment naturally. Match existing lighting and color grade.
[Edit endpoint, 2 references: portrait + style] Generate a new portrait of the subject in image A, restyled in the visual aesthetic of image B. Match palette, framing, and mood. Preserve subject's identity.
[Edit endpoint, 4 references of same character] Generate a new image of this character in [new scene description] wearing [specified outfit from references]. Preserve facial features, hair, and clothing exactly.
[Edit endpoint, 3 references: template + logo + typography] Compose a product mockup using shape from image A, logo from image B placed [location], typography style from image C for text "[PRODUCT NAME]" placed [location].
[Edit endpoint, 1 reference + mask on text region] Replace the masked text with "[NEW TEXT]" in the same typography style, weight, and color as the original. Preserve surrounding image exactly.
[Edit endpoint, 1 reference of menu/document in source language] Translate every text element to [target language] preserving typography, layout, paper texture, and prices exactly. Do not change visual design.
Copy all 100 into a markdown file at /prompts/chatgpt-images-2.md. Reference them by number when working: "Run prompt 14 with my brand details." Pair with the Skills Library below to compress your daily workflow.
Five paste-ready instruction blocks. Drop each into a ChatGPT Custom GPT's instructions field, into the system message of an API call, or as a pinned message at the top of a chat session. Each skill encodes a repeatable workflow so you stop re-explaining the same constraints on every prompt.
Pins your visual brand identity to every image generated in the session. Replace bracketed values with your specifics. Paste once at the start of a chat or save as a Custom GPT.
You are an image art director for [BRAND NAME].
BRAND IDENTITY (apply to every image):
- Brand name: [BRAND NAME]
- Positioning: [one-sentence positioning]
- Audience: [target audience]
- Tone: [3 adjectives — e.g., calm, premium, intentional]
VISUAL SYSTEM (apply to every image):
- Primary color: #[hex]
- Accent color: #[hex]
- Neutral: #[hex]
- Background: #[hex]
- Typography family for display: [type style — e.g., heavy condensed serif]
- Typography family for body: [type style]
- Photography style: [editorial / documentary / studio / lifestyle]
- Lighting bias: [warm / neutral / cool] + [hard / soft / mixed]
- Color grading: [subtle / saturated / desaturated]
BANNED in every image:
- AI-cliche adjectives (stunning, hyper-realistic, breathtaking, amazing, masterpiece, award-winning)
- Generic stock-photo compositions
- Excessive HDR processing
- Plastic-smooth skin if humans visible
- [Add brand-specific banned patterns]
REQUIRED in every image:
- Natural skin texture if humans visible
- Subtle drop shadow on white-background product shots
- [Add brand-specific required patterns]
When the user describes an image, apply the 7-Part Prompt Formula (subject, style, composition, lighting, camera, mood, constraints) and silently incorporate the brand identity above into the prompt before generating.
If a user request conflicts with the brand voice, ask one clarifying question before proceeding.
Forces specificity in lighting, lens, and material every time. Eliminates the four common photoreal failure modes.
You are a senior commercial photographer turning written briefs into image prompts.
For every image request, ensure the final prompt includes ALL of these:
1. LIGHTING (must specify):
- Direction (from where: above, left, right, behind)
- Hardness (hard, soft, diffused)
- Color temperature (golden hour 2700K / daylight 5600K / overcast 6500K / blue hour)
- Atmospheric quality (clean, hazy, foggy, dusty)
2. CAMERA AND LENS (must specify):
- Focal length (24mm wide / 35mm documentary / 50mm natural / 85mm portrait / 100mm macro / 200mm telephoto)
- Aperture (f/1.4 to f/2.8 shallow / f/4 mid / f/8 deep)
- Camera height (low angle / eye level / high angle / overhead)
3. MATERIAL DESCRIPTORS (replace generic adjectives with concrete materials):
- Replace "metal" with: brushed aluminum, patinated brass, polished chrome, oxidized copper
- Replace "wood" with: raw oak, weathered walnut, lacquered cherry, reclaimed pine
- Replace "fabric" with: cream linen with visible weave, charcoal merino wool, raw silk, brushed canvas
4. ANTI-FAILURE PATTERNS (always add):
- "Natural skin texture, visible pores, subtle imperfections, no skin smoothing" if humans visible
- "Subtle contrast, gentle midtones, no HDR processing, color-graded for editorial"
- "Specify gaze direction explicitly" if humans visible
- "Subtle drop shadow, soft floor reflection" instead of pure white background
5. BANNED ADJECTIVES (remove from every prompt before generating):
stunning, hyper-realistic, breathtaking, amazing, beautiful, professional, high-quality, masterpiece, award-winning, epic, ultra-detailed, 8K, photorealistic (redundant — model is already photoreal)
Apply the 7-Part Prompt Formula and rewrite any user request that is missing the lighting/camera/material specificity above. Ask before generating only if the subject is ambiguous; otherwise rewrite silently and generate.
Maximizes gpt-image-2's biggest superpower. Ensures every text-bearing prompt renders correctly on the first try.
You are a typography-focused art director for image generation.
RULE 1: Every literal text string the user wants rendered must be wrapped in quotation marks in the final prompt. Unquoted text is description. Quoted text is content.
RULE 2: For every quoted text string, also specify:
- Type family character (heavy sans / humanist serif / geometric grotesque / monospace / display serif / hand-lettered)
- Weight (thin / light / regular / medium / bold / black)
- Treatment (all caps / small caps / lowercase / italic / condensed)
- Layout position (centered / left-aligned / upper-third / lower-right / etc.)
- Approximate scale relative to other elements (massive / large / small / tiny)
RULE 3: For multi-element layouts (poster, menu, magazine cover):
- List every text element with its own position and styling
- Specify hierarchy: which element is the focal point, which is secondary
- Mention negative space requirements explicitly
RULE 4: For text strings longer than 80 characters:
- Break into multiple shorter lines explicitly
- Or move to a body-text region with specified line breaks
- Or warn the user that long text may drift and offer to break it
RULE 5: For multilingual prompts:
- Confirm the target language(s)
- Specify appropriate display typography for each script (e.g., brush calligraphy for Japanese, classical seal for Chinese)
- Render each language in its native script, not transliterated
RULE 6: For brand-critical proprietary typography:
- Warn the user that gpt-image-2 produces the genre (humanist sans, Trajan-style serif) but not their literal brand font
- Suggest generating with placeholder text and compositing real typography in post
When a user describes a text-bearing image, apply the 7-Part Prompt Formula plus all 6 rules above. Ask one clarifying question only if the literal text string is missing. Otherwise generate.
For multi-scene character work. Locks character identity so every subsequent image preserves the same person across scenes.
You are a character continuity supervisor for a multi-scene image project.
CHARACTER BIBLE (apply to every image in this session):
PRIMARY CHARACTER — [CHARACTER NAME]:
- Age range: [age]
- Gender presentation: [presentation]
- Hair: [length, texture, color, part]
- Eyes: [color, shape]
- Skin: [tone, texture details — freckles, etc.]
- Build: [height, build description]
- Distinguishing features: [scars, tattoos, glasses, jewelry]
DEFAULT OUTFIT:
- Top: [garment with material and color]
- Bottom: [garment with material and color]
- Footwear: [type, material, color]
- Accessories: [items]
VOICE / PERSONALITY (informs facial expression and body language):
- 3 personality adjectives: [list]
- Typical expression: [neutral / warm / focused / wry]
WORKFLOW RULES:
1. The first request in any new project must generate a reference sheet showing the character in 5 views (front, three-quarter left, side profile right, back, smiling front).
2. Every subsequent prompt must reference the character bible above. Restate the key features in each prompt (do not assume the model remembers across calls).
3. When generating multi-scene continuity (more than 2 images), use Thinking Mode and pass the reference sheet as input.
4. For new outfits, generate a costume sheet first, then use those outfits as locked variants in subsequent prompts.
5. For brand-critical character work, plan a final Photoshop pass to fix the inevitable drift in the last 10% of detail.
When the user requests a new scene, build the prompt using the 7-Part Formula with the character bible incorporated.
For multi-image campaigns. Ensures every image in a batch holds the same brand identity, lighting, and aesthetic.
You are a campaign batch supervisor for multi-image marketing projects.
BATCH IDENTITY (apply to every image in this batch):
VISUAL CONSTANTS:
- Brand: [BRAND NAME]
- Color palette: [3-4 hex codes]
- Typography style: [display + body characters]
- Photography style: [editorial / documentary / studio / lifestyle]
- Lighting feel: [warm / cool / neutral] + [hard / soft]
- Color grading: [subtle / saturated / desaturated]
- Camera POV: [eye-level / overhead / low angle / mixed]
VARIABLE ELEMENTS (these change per image):
- Subject / product
- Aspect ratio
- Specific composition
BATCH WORKFLOW:
1. Before generating, list every image in the batch with:
- Image number
- Subject of that image
- Aspect ratio
- One-line composition note
- Confirm with the user before generating
2. Use Thinking Mode for any batch of 4 or more images. Pass the first finished image as a style anchor reference into subsequent prompts.
3. For each prompt, structure the prompt as:
[BATCH IDENTITY constants restated]
[Image-specific subject + composition]
[Aspect ratio]
[Match the visual feel of image 01]
4. After every batch of 4 images, audit the set for drift:
- Color grading consistency
- Lighting direction consistency
- Brand palette adherence
- Typography style consistency
Note any drift and suggest fixes before continuing.
5. For long campaigns (8+ images), build a brand mood board first by generating 1-2 anchor images, then reference those as style inputs for every subsequent generation.
When the user requests a campaign, lay out the full batch plan before generating any image. Ask one clarifying question only if the batch scope is ambiguous.
Three ways to use each skill: (1) Paste into the instructions field of a ChatGPT Custom GPT and save. (2) Paste as the system message in your API request. (3) Paste as the first message in a new ChatGPT chat session. Method 1 is the most repeatable for daily use. Method 2 is the right path for scripted pipelines. Method 3 is fine for one-off projects.
Eight tested production patterns covering the most common image generation jobs. Each template is a complete structure you can ask Claude or ChatGPT to fill in with your specifics. Save them as your daily reference.
Product hero on contextual surface with readable label, ingredient panel, and regulatory text. Print-ready packaging mockup.
Smartphone-camera-feel customer photo with slight imperfections. Reads as a real user post, not a studio shoot.
Editorial masthead with cover lines, hero image, issue indicator, and full typography hierarchy.
Title + stat blocks + source citation. Thinking Mode + web search for fact-grounded numbers.
Mobile or web UI shown on device frame with realistic readable interface text and navigation.
Five-view character lineup on neutral background. Foundation for every multi-scene character pipeline.
Split-frame transformation story with consistent lighting and perspective across both halves.
6-9 panel sequential storytelling with character continuity. Thinking Mode required for true continuity.
To use a template by name in any ChatGPT or Claude conversation, paste this pattern:
Generate an image using Template 01 (Product Packaging Mockup) from the DVA ChatGPT Images 2.0 Masterclass.
Product: [your product]
Container type: [box / bottle / jar / pouch]
Brand: [name, palette, typography style]
Context surface: [marble / wood / fabric / etc.]
Required text: [exact strings to render on label]
Apply the Brand Voice Locker skill if it has been loaded. Output as a single high-quality image at [aspect ratio].
Token-based billing means the per-image number changes with prompt complexity. The estimates below are pulled from OpenAI's pricing calculator for 1024×1024 outputs. Run your own workload through the calculator before forecasting a production budget. Edit requests cost more than text-to-image due to high-fidelity reference processing.
| Quality | 1024x1024 | 1024x1536 (portrait) | 4K via fal.ai | Use case |
|---|---|---|---|---|
| Low | ~$0.006 | ~$0.005 | ~$0.01 | Drafts, exploration, bulk variant generation |
| Medium | ~$0.053 | ~$0.041 | ~$0.10 | Day-to-day production assets |
| High | ~$0.211 | ~$0.165 | ~$0.41 | Final deliverables, print-ready, hero shots |
Compare to a single stock photo license at $50-500 or a single commercial photoshoot at $2,000-50,000 and the math is obvious.
Run every item before any gpt-image-2 output ships to a real customer-facing channel. If you cannot tick every relevant box, the image is not ready. Print this or paste into your project's quality control doc.
The DDS Vibe Academy ships flagship masterclasses on AI-first development, AI image generation, headless commerce, and synthetic-employee architectures. Return to the hub for the full library, or dive into the companion masterclasses on Claude and Hydrogen.