Edit-first model for precise changes to uploaded photos. Edit Live credit cost
Upload the image you want to editThe source image anchors identity, product details, and composition.
Model Guide · GPT Image 2
GPT Image 2 — AI Photo Editor for Uploaded Images
GPT Image 2 is the edit-first model for changing an image you already have. Upload a portrait, product photo, or campaign asset, then describe the exact result you want. It can replace backgrounds, refine lighting, clean up distractions, and preserve recognizable facial or product details through complex edits.
Best for editing uploaded imagesPreserves identity and product detailsHandles detailed instructionsControlled, natural-looking edits
Best fitPortrait, product, and background editing
WorkflowUpload first, then describe the result
Prompt styleDetailed instructions with constraints
AlternativeSeedream 4.0 for generation from text
Complete GPT Image 2 editing guide
How to Edit Photos with GPT Image 2 Without Losing the Subject
One complete workflow for defining the change, locking identity or product details, diagnosing weak instructions, and producing a usable final image.
01 · Editing capabilities and instruction structure
Separate What Changes From What Must Stay Fixed
Reliable GPT Image 2 edits create a clear boundary between the target area, the desired new state, and protected source details.
01Core Editing StrengthsSubject-aware control
GPT Image 2 understands the relationship between the subject, foreground, background, lighting, and local image details.
- Identity-aware editing: maintains facial structure, expression, and skin tone
- Precise isolation: handles hair, fur, glasses, jewelry, and irregular edges
- Instruction following: supports multiple edits and preservation constraints
- Product preservation: retains packaging, proportions, labels, and materials
02Edit Instruction FormulaAction + target + constraints
[Edit action] + [Target area] + [Desired result] + [What must stay unchanged] + [Output style]
Describe the new state
- Say “replace with a white background”
- Do not stop at “remove background”
Lock important details
- Preserve identity for portraits
- Maintain labels and proportions for products
Separate complex edits
- Use clear sentences
- Avoid vague requests like “make it better”
02 · GPT Image 2 prompt templates
Use Complete Edit Instructions, Not Vague Commands
Each template names the action, desired result, and protected details. Replace only the target and output state for your own image.
01
Background Editing
02
Portrait Refinement
03
Product Enhancement
04
Controlled Changes
03 · Before-and-after anatomy
Turn a Raw Product Photo Into a Controlled Marketplace Asset
This case shows how one instruction can change environment, color balance, and shadow while explicitly protecting packaging.
Source problem
Product photographed on a cluttered desk
The packaging is usable, but the environment, color cast, and hard shadow do not meet marketplace standards.
Edit instruction
Replace the desk with a pure white studio background. Correct the warm color cast and add a soft contact shadow. Keep the bottle shape, label text, cap, colors, and proportions unchanged.Expected result
- Changed: background, white balance, and shadow quality
- Preserved: product geometry, packaging, label, and brand color
- Ready for: marketplace listing and catalog consistency
04 · Troubleshooting GPT Image 2 edits
Fix the Instruction Before You Repeat the Generation
Most unwanted changes come from one of three prompt failures: missing constraints, an undefined replacement state, or too many dependent edits in one pass.
“Make this a professional photo”
The model may rebuild the face, outfit, or product because nothing is locked.
Better: name identity, expression, packaging, colors, and proportions as preserved details.“Remove the background”
The desired output state is ambiguous: transparent, white, or a new environment?
Better: specify the exact background and whether natural shadows should remain.Background, wardrobe, pose, and lighting at once
A long chain of structural changes makes it harder to verify which instruction failed.
Better: finish structural edits first, then run a second pass for color or retouching.Before / after workflows
Four Edits Where the Source Image Still Matters
Unlike generation from scratch, each result keeps a recognizable subject or product.
BackgroundBackground replacement
Busy environment → controlled studio backdrop with hair edges preserved.
PortraitNatural portrait refinement
Uneven lighting → balanced headshot without changing identity.
ProductMarketplace cleanup
Raw product capture → consistent listing image with packaging intact.
CampaignCampaign adaptation
Base asset → channel-ready composition with planned negative space.
FAQ
GPT Image 2 Questions
It is best when you already have an image and need to modify it while keeping the subject recognizable: portrait refinement, background replacement, product cleanup, and controlled object removal.
Uploading an image is recommended because this is an edit-first workflow. For a completely new photorealistic image from text, start with Seedream 4.0.
Explicitly ask the model to preserve identity, facial structure, skin tone, expression, and hairstyle. Clear preservation constraints reduce unnecessary reconstruction.
Yes. Identify those elements as locked details and request that label text, colors, packaging design, proportions, and product materials remain unchanged.
You can request an isolated subject intended for transparent PNG output. Explicitly request clean edges around hair, fur, glass, or other difficult boundaries.
Use GPT Image 2 to adapt a real product photo. Use Seedream 4.0 when no source photo exists and the scene should be generated from text.