OpenAI: GPT Image 1

GPT Image 1 is OpenAI’s natively multimodal image model, accepting text and image inputs and producing image outputs. It can generate a new scene, use reference images to shape a composition or edit an existing asset through a new instruction. Practical work includes illustration, product imagery, design concepts and local changes guided by a mask. This is the named GPT Image 1 edition, separate from Image 1 Mini, Image 1.5 and later GPT Image versions.

Prepare a brief that states the subject, setting, composition, lighting and visual style. If the task needs visible text, include the exact wording and explain its intended hierarchy and position. For a reference composition, assign each image a clear role, such as a product, an environment or a visual treatment. The direct generation and edit workflows permit a prompt up to 32,000 characters and one to ten image results. Compare the returned variations against the same brief and keep the chosen asset for subsequent work.

Editing can take up to sixteen reference images. GPT Image 1 supports low and high input_fidelity, with low as the default. This control affects the effort to preserve the source image’s features and style; it is separate from rendering quality and output size. A high-fidelity edit is useful when source details matter, but it is not a guarantee that every identifying detail will remain identical. State what should be retained and what may change rather than relying on the fidelity setting alone.

Masked editing identifies an area to modify. Prepare the source and mask in the same format and dimensions, below 50 MB in the guide’s requirements, with an alpha channel in the mask. With multiple source images, the mask is applied to the first. A useful edit brief can ask for a different wall colour while retaining furniture, camera angle and lighting. The model uses the mask as guidance rather than promising an exact boundary, so inspect both the changed area and its surroundings.

The documented standard output sizes are 1024 × 1024, 1536 × 1024 and 1024 × 1536 pixels, with automatic size selection where supported. Quality is low, medium, high or auto. GPT Image 1 uses these standard frames and quality levels up to high. Choose the standard frame around the intended display and the detail level around the asset’s purpose.

PNG is the default output, with JPEG and WebP alternatives. Compression from 0 to 100 is available for JPEG and WebP. Transparent, opaque and automatic background choices support different delivery needs; transparency requires PNG or WebP. For an isolated graphic, use a format that retains its alpha channel before placing it in a later layout. For a complete scene, keep the background and the subject’s relationship to it in the brief.

OpenAI documents generation and editing directly, and the model card also includes conversational image use. For iterative work, retain the chosen result as the source for a follow-up change instead of choosing a different variation implicitly. The image reference documents streaming with up to three partial images; these show progress rather than extra finished assets. Keep the completed image and the brief together so later edits refer to the actual composition. The manufacturer interface also supports batch processing.

The model can still have difficulty with precise text placement, tightly structured layouts and consistency across recurring subjects. Inspect lettering, object boundaries, colour treatment and the details that identify source material at the intended size. If an edit needs refinement, describe the visible mismatch and the elements to preserve. GPT Image 1 is useful for a concrete sequence of generation, reference-led editing and asset preparation, with its own standard sizes and selectable input fidelity kept distinct from other editions.