GPT Image 2.5
GPT Image 2.5 is an image model by OpenAI. The current default image family in Vini. Use generation for a new composition and editing for a precise change to an existing scene. The quality selector changes the request; inspect the result before increasing quality.
| Setting | Value |
|---|---|
| Model ID | gpt_image_25 |
| Name | GPT Image 2.5 |
| Availability | available |
| Default for new nodes | yes |
| Earlier IDs resolving to this record | gpt_image_2 |
Common declared limits
Section titled “Common declared limits”| Setting | Value |
|---|---|
| Capabilities | text to image, reference edit, multi reference |
| Maximum prompt length | 32000 |
| Controls | quality: draft, light, balanced, premium, max; fallback balanced |
| Element input form | sheet |
| Maximum image references | 16 |
Limits apply together. initial is the studio choice for a new node; default is the fallback for an existing node without that value. Declared language or input support does not guarantee a successful result.
Prompting playbook
Section titled “Prompting playbook”State the change first and list the source properties to preserve. Assign each reference a role. Quote exact lettering and specify its position rather than asking for a generic poster.
Studio example
Section titled “Studio example”On an Image node (Картинка), choose GPT Image 2.5. Supply the input image or Element when the selected operation requires a reference. Use this original example as a starting point:
Change only the mug on the table to pale blue. Preserve the person’s face, hand position, table grain, camera angle and warm window light.
Review the prompt and expected format before running. Compare the output against the source, including hands, face, object count, lettering and crop. Use the resulting scene image as a video first frame only after approving it. For automation, follow the supported MCP workflow.
Known failure modes
Section titled “Known failure modes”Small text and complex edits still need inspection. Acceptance of one realistic portrait does not define the model’s entire moderation policy. Higher quality is not evidence that a failed composition will be repaired.
Evidence and limits
Section titled “Evidence and limits”The 18 September 2026 study generated a realistic portrait previously refused by the older 2.0 model and measured generation and editing requests. The upper quality levels were not rendered in that study; no output-quality comparison across all five levels is claimed.
An available reference input is not a guarantee of identity preservation. See reference workflows and image style samples. Generation and editing are different tasks; keep their evidence separate.
Reviewed 14 October 2026 against Vini’s image registry and dated research. Official model-author documentation checked on that date. Recheck the guidance when the model version or connected operation changes.