GPT Image 2.0: The Complete Guide
Jul 2026
GPT Image 2.0 is OpenAI's image model for creators who need more than one beautiful frame. It generates up to 2K images, renders multilingual text more accurately than earlier OpenAI image models, and in Thinking mode can produce up to eight coherent images from one prompt with the same character across the batch.
That batch consistency is the reason this guide exists. For AI filmmakers, GPT Image 2.0 is not just an image generator — it is a fast character-sheet machine. Generate the face, wardrobe, angles, expressions, and pose references once, then use that sheet inside Flick to keep the same character consistent across Kling, Seedance, Veo, FLUX, and whatever model you use next.
GPT Image 2.0, at a glance
- What it is: GPT Image 2.0, API name
gpt-image-2, is OpenAI's image generation and editing model, released April 21, 2026. - Best use case: layout-heavy images, text in images, multilingual posters, product/scene stills, and character sheets for AI video.
- Key filmmaking feature: Thinking mode can generate up to eight coherent images from one prompt, which makes it ideal for front/side/back/expression reference packs.
- Modes: Thinking mode for paid tiers; Instant mode for faster exploration and free-tier access.
- Resolution: up to 2K, with experimental 2560×1440 output.
- API: supports image generation, edits, multi-image inputs,
nbatches, quality levels, and PNG/JPEG/WebP outputs. - Workflow: design a character in GPT Image 2.0, turn that character into a reference sheet, upload it to Flick, then animate the same identity with reference-driven video models.
- Verdict: use GPT Image 2.0 as the art department and Flick as the production canvas.
Turn your character sheet into shots
Upload the sheet once. Keep the same face in every generation, every model.
GPT Image 2.0 Spec
| Spec | GPT Image 2.0 |
|---|---|
| Release date | April 21, 2026 |
| API model name | gpt-image-2 |
| Resolution | Up to 2K; experimental 2560×1440 |
| Aspect ratios | 3:1 ultra-wide through 1:3 ultra-tall |
| Batch generation | 1–8 images per prompt with the n parameter |
| Best continuity feature | Coherent 8-image Thinking-mode batches |
| Modes | Thinking and Instant |
| Thinking mode access | Plus, Pro, Business, Enterprise |
| Instant mode access | All tiers, including free |
| Text rendering | Strong, including non-Latin scripts like Hindi and Bengali |
| Output formats | PNG, JPEG, WebP with compression control |
| Knowledge cutoff | December 2025 |
| Best Flick use | Character sheets, scene stills, poster/key art, reference packs |
Thinking mode vs Instant mode
GPT Image 2.0 has two practical personalities.
Thinking mode reasons before it renders. It can plan the layout, search for references where available, verify details, and keep characters or objects coherent across a batch. That makes it the mode to use when a character will appear more than once.
The tradeoff is speed. Thinking mode can take 15–30 seconds per generation, so it is not the fastest way to rough out ideas.
Instant mode is for exploration. It keeps the same core visual quality but skips the deeper reasoning pass. It is faster, available on every ChatGPT tier, and useful for finding a look before spending time or budget on a character sheet.
The simple rule:
- Use Instant for moodboards, first looks, thumbnails, and visual exploration.
- Use Thinking for characters, products, posters with text, and any image that needs continuity.
Why GPT Image 2.0 matters for AI video
Most AI video consistency problems start before the video model. The model drifts because the character was never defined clearly enough.
A single portrait tells the video model one angle, one expression, and one lighting condition. A reference sheet tells it the full person:
- face shape
- hair
- wardrobe
- proportions
- accessories
- front view
- side views
- back view
- full body
- close-up expressions
- hands
- speech shapes
- lighting baseline
GPT Image 2.0's 8-image batch gives you that material in one coherent run. Instead of prompting eight separate images and getting eight similar-but-different people, you can ask for the whole reference pack at once. The batch is planned as one set, so it tends to agree with itself.
That agreement is exactly what reference-driven video models reward.
The character-sheet method
This is the workflow for turning GPT Image 2.0 into a production tool.
Step 1: design the identity frame
Start in Instant mode. Generate one clear portrait or full-body frame until the character is worth preserving.
Do not prompt vaguely. Include concrete identity details:
- age range
- face shape
- hair color and cut
- skin details
- distinctive marks
- exact wardrobe
- accessory placement
- color palette
- body type
- emotional baseline
- world/style
When you get a design that works, save it. This is the identity frame. The exact identity-frame prompt example is collected in the prompt examples section at the bottom.
Step 2: generate the 8-image Thinking-mode sheet
Feed the identity frame back in as a reference image. Then ask GPT Image 2.0 to produce the angles and expressions your video models need.
This is the core GPT Image 2.0 trick: use the model's batch coherence to create a contact sheet that can survive downstream video generation. The full 8-image sheet prompt is collected at the bottom with the other examples.
Step 3: export the reference pack
Save the batch as a clean reference pack. Keep the files organized by angle:
front-neutralleft-profileright-profilethree-quarter-backfull-bodyseated-handscloseup-smilecloseup-speech
If you want a stronger sheet, combine the images into one contact sheet and keep the individual frames too. Video tools may prefer either individual references or a single sheet depending on the model.
Step 4: bring the sheet into Flick
In Flick, GPT Image 2.0 becomes the art department and the canvas becomes the production board.
A practical Flick workflow:
- Open a fresh canvas or the project canvas for your film.
- Upload the GPT Image 2.0 identity frame and reference sheet.
- Use the sheet as your character reference.
- Generate or upload a scene still.
- Select the still and choose Generate Video.
- Choose the model that fits the shot:
- Kling for motion-heavy action and physical detail.
- Seedance for longer clips, multi-reference workflows, and dialogue-heavy scenes.
- Veo for polished cinematic realism where available.
- FLUX or other models for model-specific looks.
- Reference the same character pack for every shot.
- Generate variations, compare drift, and keep the clip that preserves identity best.
- Save strong frames back to the canvas as future references.
- Cut, grade, and assemble the sequence on the same canvas.
The goal is not to make GPT Image 2.0 do everything. The goal is to make it produce the reference material that every video model needs.
How to use GPT Image 2.0 in Flick-style production
A complete still-to-video pipeline looks like this:
1. Build the character before the scene
Do not start with a complex action shot. Start with a clean identity.
Prompt GPT Image 2.0 for a plain-background portrait or full-body reference. Keep the lighting simple. Avoid props that may confuse the character's silhouette.
2. Create the reference sheet
Use Thinking mode and generate the eight-angle batch. This gives Flick's Character Reference workflow more than a single face to work with.
3. Generate the first scene still
Once the character is locked, generate the scene still. Use the same identity-preservation language from your character sheet and apply it to the new setting. The full scene-still prompt is included in the prompt examples section at the bottom.
4. Upload the still and sheet to Flick
On the Flick canvas, keep the character sheet and the scene still near each other. Treat them as your shot kit.
5. Animate with a reference-driven video model
Select the scene still, generate video, and attach the character reference. If the model supports image references, use the strongest identity images from the sheet: front, profile, full body, and expression.
If the model supports named mentions or element references, describe the shot using stable handles. The exact Flick-style prompt is included with the prompt examples at the bottom.
6. Review drift
After each generation, check:
- face shape
- eyes
- hairline
- outfit color
- accessories
- body proportions
- hands
- side-profile accuracy
- voice/speech alignment if applicable
Save the best frames back to the canvas. Strong output frames can become better references for later shots.
Prompt formula for GPT Image 2.0
GPT Image 2.0 responds best to structured prompts. Use this order:
- Artifact type: what the output is.
- Subject: who or what appears.
- Layout or camera: how the image is composed.
- Important details: the details that must survive.
- Style and lighting: the visual finish.
- Constraints: what to preserve or avoid.
The order matters because the model plans early. If you begin with vague praise, the planner wastes attention. If you begin with the artifact and layout, it knows what to build. All copy-paste prompt examples are collected at the bottom of the guide.
GPT Image 2.0 vs the field
| Feature | GPT Image 2.0 | Midjourney v8 | Nano Banana 2 |
|---|---|---|---|
| Text in images | Best in class | Weak | Improved |
| Layout control | Strong | Medium | Strong |
| Aesthetic ceiling | Strong | Still the single-frame benchmark | Comparable |
| Consistent batches | Up to 8 per prompt | No | No |
| API | Public | No public API | Available |
| Speed | Medium in Thinking mode | Fast | 1–3s |
| Best use | Text, layout, character sheets | Beautiful one-off frames | Cheap volume and fast iteration |
The simplest split:
- Use Midjourney for one beautiful still.
- Use Nano Banana 2 for cheap volume and fast variants.
- Use GPT Image 2.0 for text, layout, editable references, and character sheets.
- Use Flick to turn those stills into consistent video shots.
Best practices
- Start with one clean identity frame before building a sheet.
- Use Thinking mode for continuity-sensitive work.
- Generate the sheet as one batch, not eight separate prompts.
- Repeat invariants exactly: same person, same outfit, same lighting, same proportions.
- Use plain backgrounds for references.
- Keep accessories simple and consistent.
- Label every input image by role.
- Use medium quality for reference packs and high quality for final hero images.
- Save strong frames back to your Flick canvas as future references.
- Treat every good output as reusable production material, not a disposable image.
Common mistakes
Mistake 1: using one portrait as the entire reference
A portrait is not enough for multi-shot video. Add profile, back, full-body, and expression references.
Mistake 2: generating references one at a time
Separate prompts drift. Use one 8-image batch so the sheet agrees with itself.
Mistake 3: changing the preservation language
If one prompt says "navy bomber" and the next says "blue jacket," you invite drift. Paste the same identity block every time.
Mistake 4: asking for accurate text without layout
For posters, title cards, and UI, define the layout before the copy. The model needs to plan where the text goes.
Mistake 5: animating before locking identity
If you send a weak still into a video model, you get a weak lock. Build the reference pack first, then animate.
Limitations
GPT Image 2.0 is powerful, but it is not magic.
- Knowledge cutoff: December 2025, so newer products, people, events, and logos may not be accurate.
- Logos: precise brand marks are still unreliable.
- Thinking latency: 15–30 seconds can feel slow for rapid ideation.
- Character consistency across separate sessions: strong in a batch, weaker across unrelated prompts unless you use references and identical preservation language.
- Tiny text: large titles and labels work better than legal-copy-sized text.
- Architecture: OpenAI has not disclosed the full architecture, so workflow claims should be treated as empirical production guidance.
Prompt examples
Identity frame prompt
Real photograph, studio reference portrait of a courier in their late 20s, buzzed copper hair, freckles across the nose, silver hoop in the left ear, navy bomber jacket with a yellow zipper, black cargo pants, scuffed white sneakers, alert but tired expression, plain grey background, soft even light, no extra text.
8-image character sheet prompt
Character reference sheet of the person in Image 1: same person, same face, same outfit, same lighting in every image. Generate 8 coherent reference images in one batch: 1) front view, neutral expression; 2) left profile; 3) right profile; 4) three-quarter back view; 5) full body standing; 6) seated with hands visible; 7) close-up smiling; 8) close-up mid-speech. Plain grey background, soft even light, photorealistic. Do not redesign the character. Do not change outfit, hair, face, proportions, or accessories.
General character sheet prompt
Character reference sheet of [CHARACTER]. Same person, same outfit, same lighting in every image. Generate 8 coherent images in one batch: front neutral, left profile, right profile, three-quarter back, full body standing, seated with hands visible, close-up smiling, close-up mid-speech. Plain grey background, soft even light, photorealistic. Do not redesign the character.
Prompt formula example
Character reference sheet, photorealistic. A lighthouse keeper in their 60s with a short white beard, broken nose, cream cable-knit sweater, dark oilskin coat, wool cap, heavy boots. Eight coherent images in one batch: front, left profile, right profile, three-quarter back, full body standing, seated with hands visible, close-up smile, close-up speaking. Plain grey background, soft even studio light. Same person, same outfit, same proportions, same lighting in every image. No extra text.
Expression grid prompt
Expression grid of the character in Image 1. Same face, same hair, same outfit, same lighting in every panel. 16 panels in a 4x4 grid: neutral, smile, laugh, worry, fear, anger, shouting, crying, exhausted, relieved, suspicious, focused, surprised, calm, mid-speech, listening. Plain background, bust framing, consistent identity.
Scene still prompt
Image 1 is the character reference sheet. Create a cinematic film still of the same character entering a rain-soaked subway platform at midnight. Preserve the same face, hair, jacket, proportions, and accessories. 35mm feel, motivated fluorescent overhead light, wet tile reflections, no extra text.
Film still prompt
Cinematic film still of the character in Image 1 crossing a wet neon alley at 2am. Preserve the same face, hair, jacket, proportions, and accessories. 35mm feel, shallow depth of field, blue shadows, orange practical light from a noodle stall, steam in the background, no extra text.
Flick video prompt
Use @CharacterSheet as the identity reference and @SceneStill as the start frame. The same courier steps onto the platform, looks over their shoulder, and slows as the train lights flicker behind them. Preserve face, jacket, freckles, earrings, proportions, and color palette.
Poster prompt
Minimal film poster, vertical. The character from Image 1 stands in silhouette at the end of a flooded hallway, one overhead light reflected in the water. Title "LAST EXIT" in thin white serif letters, centered lower third. Small credit block beneath. No extra text, verbatim title rendering.
Product/layout prompt
Three-column festival poster. Left column: dates in bold condensed type. Center column: title "ONE LAST SUMMER" in distressed serif. Right column: lineup text in small caps. Sunset beach photo background, faded print texture, warm orange and cream palette. No extra text beyond the specified copy.
Multi-image edit prompt
Image 1 is the character reference sheet. Image 2 is the base scene. Place the character from Image 1 into Image 2, matching Image 2's lighting and camera angle. Preserve the character's face, outfit, proportions, and accessories. Change nothing else in the scene.
Frequently Asked Questions
What is GPT Image 2.0?
GPT Image 2.0 is OpenAI's image generation and editing model, released April 21, 2026. Its API model name is gpt-image-2.
Is GPT Image 2.0 free?
Instant mode is available on every ChatGPT tier, including free. Thinking mode requires a paid tier such as Plus, Pro, Business, or Enterprise. API usage is priced by image/output settings.
Can GPT Image 2.0 keep characters consistent?
Yes, especially inside one Thinking-mode batch. It can generate up to eight coherent images from one prompt, which makes it useful for character sheets. Across separate prompts, use reference images and repeat the same identity constraints verbatim.
How do I make a character sheet with GPT Image 2.0?
Design one identity frame, then use Thinking mode to generate eight reference views in one batch: front, left profile, right profile, three-quarter back, full body, seated hands, close-up smile, and close-up speech. Use the same-person/same-outfit/same-lighting constraint.
How much does a GPT Image 2.0 character sheet cost?
Based on the article's working estimates, an 8-image 1024×1024 sheet costs about $0.42 at medium quality and about $1.69 at high quality.
Is GPT Image 2.0 better than Midjourney?
For text, layout, API access, and character sheets, yes. For a single beautiful frame, Midjourney may still have the higher aesthetic ceiling.
Can I use GPT Image 2.0 images for AI video?
Yes. The strongest workflow is to generate a character reference sheet with GPT Image 2.0, upload it to Flick, and use the same sheet as a character reference for video models like Kling, Seedance, Veo, and FLUX.
What should I use GPT Image 2.0 for in Flick?
Use it for character sheets, scene stills, title cards, posters, product references, and style frames. Then use Flick's canvas to animate the stills, compare models, save references, and assemble the sequence.