GPT-Image 2.5
Generate and edit images with GPT Image 2.5 on Mitte. Precise edits, references that stay recognizable, real text, up to 4K, transparent backgrounds.
GPT Image 2.5: Generate and Edit Images That Stay True to Your References
GPT Image 2.5 is OpenAI’s new image model, and you can use it on Mitte right now. Describe an image and get it. Attach a photo and say what to change. It produces natural light, real texture, and legible type, and it edits the one thing you asked for without repainting the rest.
The big difference from earlier models is how well it holds on to what you give it. Put a person from a reference photo into a new scene and they still look like themselves. Swap a product’s label and the hands holding it don’t change. Ask for a poster with specific copy and the words come out spelled right, in the layout you described.
It’s also fast. Variations come back in seconds, so you can iterate the way you’d sketch: many quick takes, then commit.
What can you do with it?
- Generate from a brief. Describe the subject, style, lighting, and composition, up to 4 images per run. Long prompts are welcome; write it like a creative brief, not a search query.
- Edit with plain language. Attach an image and type “make it golden hour”, “remove the second chair”, “change the jacket to red”. Everything you didn’t mention stays put.
- Keep people and products recognizable. Use reference photos of a face, a product, or a place, and move them into new settings, seasons, or art styles.
- Combine up to 16 references. Product from one shot, model from another, environment from a third. Name each one in the prompt and say what it contributes.
- Mask an exact area. Add a mask to lock the edit to one region: a headline, a sleeve, a window.
- Design with real text. Posters, packaging mockups, menus, social tiles, thumbnails. Put the copy in quotes and it renders it.
- Cut out on transparent. Generate logos, stickers, and product shots on a transparent background, ready to drop into a layout.
- Stack edits. Run an edit on the result, then another. Each pass builds on the last without the image going soft.
How to use it
- Write the prompt. Subject, style, lens, light, and any copy in quotes. The more specific, the less it drifts.
- Attach references to edit (optional). One image for a straight edit, several when you’re combining elements. Your prompt becomes the instruction for what changes.
- Add a mask (optional). With a reference attached, upload a black-and-white mask where white marks the area to change.
- Choose ratio and resolution. Square, landscape, portrait, or cinematic 21:9, at 1K, 2K, or 4K. Editing? Pick Auto and the output matches your source image.
- Pick a quality level. Low for quick explorations, High for finished work, Extra High or Max when every detail counts.
- Generate. Download the results, share them, or keep editing right on Mitte.
What you get
Up to 4 images per run in JPEG, PNG, or WebP, at 1K, 2K, or 4K, in your chosen ratio. Transparent backgrounds come through in PNG and WebP.
Tips
- Explore low, finish high. Run a batch of 4 at Low quality to find the direction, then rerun the winner at High or Extra High with the same prompt.
- Say what stays. “Replace the label with the attached artwork. Keep the lighting, hands, and background.” Naming what shouldn’t move is the fastest way to a clean edit.
- One change per pass. Three small edits in a row beat one paragraph of simultaneous changes.
- Address references by number. “The person in image 1, wearing the jacket from image 2, on the street in image 3.”
- Quote your copy. Put headlines and labels in quotation marks and tell it the font feel: “bold condensed sans”, “hand-lettered script”.
- Use Auto ratio when editing. It keeps the output the same shape as your input, so nothing gets cropped or stretched.
- Transparent needs PNG or WebP. JPEG can’t hold transparency. Switch the output format before you generate a cutout.
- Save 4K for the final. Lock the composition at 1K or 2K, then rerun the keeper at 4K.