Prompt-generated images include the Wan 3.0 site mark when removal is not available on your account.
View paid options for watermark control
Image generation and editing
GPT Image 2.5 is an image generation and editing model family released on September 8, 2026. It is built for sharper detail, more natural lighting and texture, stronger preservation of reference subjects, and more reliable local changes across repeated revisions.
Use this page to choose a creation or editing path, write a prompt the model can follow, and continue in the Wan 3.0 image workspace without treating the first draft as a finished brand asset.
Core capabilities
The model family is designed for both first-pass generation and controlled revision: keep the details that already work, then change only the part the brief requires.
01When a face, product silhouette, room layout, or camera angle already matters, GPT Image 2.5 is positioned to keep those anchors while the scene, wardrobe, or style changes. Name the details that define recognition first, then describe the transformation separately so the reference acts as a constraint rather than a loose mood board.
Open reference editing
02Targeted edits are useful when the rest of the frame is already approved. Ask for one change—background, color, material, object removal, or copy space—then list what must stay: subject, crop, camera height, lighting direction, and product geometry. A narrow first revision is easier to judge than a request that rebuilds the entire image.
Edit from a source image
03Later instructions should build on earlier decisions instead of drifting the subject, layout, or finish. Save the last accepted still, repeat the protected details, and change one variable per turn. If several unwanted shifts appear, return to the last approved image rather than stacking corrections.
Start a refinement loop
04Layered briefs can combine multiple subjects, spatial relationships, style, hierarchy, and on-image wording. Supply the final copy, language, and reading order. Treat generated letters, prices, dates, and legal lines as a visual draft and rebuild anything that must stay exact in an editable design layer.
Draft from a text brief
05The family supports opaque or transparent backgrounds and common output tiers from 1K through 4K, with a wide set of aspect ratios. Ultra-wide ratios such as 27:16 and 16:27 are documented at 1K only. Inspect hair, glass, shadows, and edge fringing before placing a cutout over a final background.
Set the output format
06A headline quality change is more natural light and richer surface detail, not only a sharper thumbnail. Check reflections, contact shadows, fabric weave, metal brushing, and whether the light direction agrees across the subject. Flare is the faster everyday option; Sunburst is the slower, precision-led option for demanding generation and editing.
Generate a first stillPractical workflow
Start from a prompt or an uploaded still, then generate or edit toward one clear visual target.
Use text-to-image for a new composition, or image-to-image when a source must stay recognizable. Set aspect ratio and resolution to match the channel before you write a long prompt.
Name the subject, scene, required text, format, ratio, and details that must not change. For an edit, write the change first and the preservation list second.
Review the first still at a usable size, then make one focused correction for type, color, crop, or a protected region. Keep the accepted version with its prompt so later turns stay comparable.
Model comparison
Match the task to a documented strength. These are workflow differences, not ranked scores.
| Capability | GPT Image 2.5 | Nano Banana 2 | Seedream 5.0 Pro |
|---|---|---|---|
| Text-rich visuals | Layout, posters, labels, and instruction-led type | Multilingual text and localized layouts | Dense infographics and multilingual design |
| Text to image | Supported | Supported | Supported |
| Image to image editing | Local edits with surrounding composition held | Conversational edits and multi-reference blending | Region-aware edits, layers, and fusion |
| Reference fidelity | Subject, product, and scene anchors | Up to 10 object and 4 character references | Geometry and identity held during regional edits |
| Production fit | Posters, product stills, ads, UI concepts, and edits | Fast iteration, 4K, and knowledge-heavy scenes | Infographics, localized campaigns, and layered comps |
Model library
Open another model page when the brief needs a different balance of speed, layout control, or editing style.
Fast image drafts
Structured production
Live image workspace
Connected tools
Move from a model page to the task you need: create a still, revise a source, or prepare a frame for video.
Create from a prompt
Revise a source
Product stills
FAQ
Direct answers about GPT Image 2.5 generation, editing, text, variants, and output options.
GPT Image 2.5 is an image generation and editing family released on September 8, 2026. It accepts text and image inputs and returns images. Documented improvements include sharper detail, more natural lighting and texture, stronger reference-subject preservation, more reliable local edits, and better consistency across repeated revisions.
It is a strong fit when a still must stay close to a brief: campaign concepts, product scenes, posters, UI studies, and reference-led edits. Flare is the faster everyday choice. Sunburst is the slower option when editing precision and finish matter more than the first-pass speed.
Yes. Upload a source, state the change, and list what must remain: identity, product geometry, crop, lighting, or label area. Review the unchanged regions as carefully as the new area, because a useful edit is the one that does not disturb already approved decisions.
The model family is documented for complex layouts that include text. Provide the exact wording and language. Then check every character, number, and line break at full size. Rebuild prices, legal copy, and accessibility-critical text in a normal design file before publishing.
Flare is positioned as the default for most everyday generation, with lower latency than the previous Images 2.0 generation path. Sunburst is the more capable generation and editing option for precision-led work, with longer generation times. Start with Flare when you need many directions quickly; test Sunburst when a protected subject or product must survive a demanding edit.
Text-to-image and image-to-image are the core modes. Documented controls include prompt, optional reference images, aspect ratio, 1K / 2K / 4K resolution, and opaque, transparent, or automatic background. Some ultra-wide ratios are limited to 1K. For 2K or 4K transparent output, the brief should describe an isolated subject without scenery or drop shadow.
On Wan 3.0
Write a production brief, generate or edit in the current image workspace, then keep the accepted still with its prompt so the next crop, localization, or image-to-video test stays on the same visual system.
Define channel, ratio, protected marks, and the one change you need. Attach only files you are allowed to transform.
Check identity, product geometry, type, crop safety, and contrast. Rebuild critical copy in an editable layer.
Adapt desktop, mobile, and social crops from the same checkpoint, or carry the frame into image-to-video with a separate motion brief.
Prompt-generated images include the Wan 3.0 site mark when removal is not available on your account.
View paid options for watermark controlImage generation and editing
GPT Image 2.5 is an image generation and editing model family released on September 8, 2026. It is built for sharper detail, more natural lighting and texture, stronger preservation of reference subjects, and more reliable local changes across repeated revisions.
Use this page to choose a creation or editing path, write a prompt the model can follow, and continue in the Wan 3.0 image workspace without treating the first draft as a finished brand asset.
Core capabilities
The model family is designed for both first-pass generation and controlled revision: keep the details that already work, then change only the part the brief requires.
01When a face, product silhouette, room layout, or camera angle already matters, GPT Image 2.5 is positioned to keep those anchors while the scene, wardrobe, or style changes. Name the details that define recognition first, then describe the transformation separately so the reference acts as a constraint rather than a loose mood board.
Open reference editing02Targeted edits are useful when the rest of the frame is already approved. Ask for one change—background, color, material, object removal, or copy space—then list what must stay: subject, crop, camera height, lighting direction, and product geometry. A narrow first revision is easier to judge than a request that rebuilds the entire image.
Edit from a source image03Later instructions should build on earlier decisions instead of drifting the subject, layout, or finish. Save the last accepted still, repeat the protected details, and change one variable per turn. If several unwanted shifts appear, return to the last approved image rather than stacking corrections.
Start a refinement loop04Layered briefs can combine multiple subjects, spatial relationships, style, hierarchy, and on-image wording. Supply the final copy, language, and reading order. Treat generated letters, prices, dates, and legal lines as a visual draft and rebuild anything that must stay exact in an editable design layer.
Draft from a text brief05The family supports opaque or transparent backgrounds and common output tiers from 1K through 4K, with a wide set of aspect ratios. Ultra-wide ratios such as 27:16 and 16:27 are documented at 1K only. Inspect hair, glass, shadows, and edge fringing before placing a cutout over a final background.
Set the output format06A headline quality change is more natural light and richer surface detail, not only a sharper thumbnail. Check reflections, contact shadows, fabric weave, metal brushing, and whether the light direction agrees across the subject. Flare is the faster everyday option; Sunburst is the slower, precision-led option for demanding generation and editing.
Generate a first stillPractical workflow
Start from a prompt or an uploaded still, then generate or edit toward one clear visual target.
Use text-to-image for a new composition, or image-to-image when a source must stay recognizable. Set aspect ratio and resolution to match the channel before you write a long prompt.
Name the subject, scene, required text, format, ratio, and details that must not change. For an edit, write the change first and the preservation list second.
Review the first still at a usable size, then make one focused correction for type, color, crop, or a protected region. Keep the accepted version with its prompt so later turns stay comparable.
Model comparison
Match the task to a documented strength. These are workflow differences, not ranked scores.
| Capability | GPT Image 2.5 | Nano Banana 2 | Seedream 5.0 Pro |
|---|---|---|---|
| Text-rich visuals | Layout, posters, labels, and instruction-led type | Multilingual text and localized layouts | Dense infographics and multilingual design |
| Text to image | Supported | Supported | Supported |
| Image to image editing | Local edits with surrounding composition held | Conversational edits and multi-reference blending | Region-aware edits, layers, and fusion |
| Reference fidelity | Subject, product, and scene anchors | Up to 10 object and 4 character references | Geometry and identity held during regional edits |
| Production fit | Posters, product stills, ads, UI concepts, and edits | Fast iteration, 4K, and knowledge-heavy scenes | Infographics, localized campaigns, and layered comps |
Model library
Open another model page when the brief needs a different balance of speed, layout control, or editing style.
Fast image drafts
Structured production
Live image workspace
Connected tools
Move from a model page to the task you need: create a still, revise a source, or prepare a frame for video.
Create from a prompt
Revise a source
Product stills
FAQ
Direct answers about GPT Image 2.5 generation, editing, text, variants, and output options.
GPT Image 2.5 is an image generation and editing family released on September 8, 2026. It accepts text and image inputs and returns images. Documented improvements include sharper detail, more natural lighting and texture, stronger reference-subject preservation, more reliable local edits, and better consistency across repeated revisions.
It is a strong fit when a still must stay close to a brief: campaign concepts, product scenes, posters, UI studies, and reference-led edits. Flare is the faster everyday choice. Sunburst is the slower option when editing precision and finish matter more than the first-pass speed.
Yes. Upload a source, state the change, and list what must remain: identity, product geometry, crop, lighting, or label area. Review the unchanged regions as carefully as the new area, because a useful edit is the one that does not disturb already approved decisions.
The model family is documented for complex layouts that include text. Provide the exact wording and language. Then check every character, number, and line break at full size. Rebuild prices, legal copy, and accessibility-critical text in a normal design file before publishing.
Flare is positioned as the default for most everyday generation, with lower latency than the previous Images 2.0 generation path. Sunburst is the more capable generation and editing option for precision-led work, with longer generation times. Start with Flare when you need many directions quickly; test Sunburst when a protected subject or product must survive a demanding edit.
Text-to-image and image-to-image are the core modes. Documented controls include prompt, optional reference images, aspect ratio, 1K / 2K / 4K resolution, and opaque, transparent, or automatic background. Some ultra-wide ratios are limited to 1K. For 2K or 4K transparent output, the brief should describe an isolated subject without scenery or drop shadow.
On Wan 3.0
Write a production brief, generate or edit in the current image workspace, then keep the accepted still with its prompt so the next crop, localization, or image-to-video test stays on the same visual system.
Define channel, ratio, protected marks, and the one change you need. Attach only files you are allowed to transform.
Check identity, product geometry, type, crop safety, and contrast. Rebuild critical copy in an editable layer.
Adapt desktop, mobile, and social crops from the same checkpoint, or carry the frame into image-to-video with a separate motion brief.
