# Prompting GPT Image GPT Image reads a prompt like a spec sheet. Name the result, lay out the frame, quote every word of text and list what must not change. Last updated: 2026-10-05 **Key facts** - Sloosh has three GPT Image models: GPT-image-2, GPT Image 2.5 Flare (fast) and GPT Image 2.5 Sunburst (careful), and all three take the same prompts. - GPT Image models in Sloosh take up to 16 reference images and prompts of up to 32,000 characters. - GPT Image reads a prompt as a spec: the result, the subject, the setting and light, the camera, any text in quotes, and what must not change. - GPT Image quality in Sloosh is auto, low, medium or high on GPT-image-2, and the GPT Image 2.5 models add xhigh and max. Web page: https://sloosh.ai/docs/prompting-gpt-image Sloosh has three GPT Image models: GPT-image-2, GPT Image 2.5 Flare (fast) and GPT Image 2.5 Sunburst (careful). They take the same prompts. Reach for them when the image carries words, when an edit has to be exact, or when several references go into one shot. ## How it reads a prompt GPT Image treats your prompt as a spec. Concrete, visible details land. Mood words mostly don't: "cinematic" tells it very little, while "low sun from the left, long shadows, teal and orange" tells it what to draw. It also follows a prompt better when each kind of instruction sits in its own sentence or line: what the image is, how it's framed, what it looks like, what it says and what it keeps. ## A prompt that works 1. Result: what the image is and where it goes, such as "a product photo for a Shopify listing" or "a 4:5 Instagram ad". 2. Subject: the product or person, with the details that make it yours. 3. Setting and light: surface, background, where the light comes from. 4. Camera: angle, distance, lens, depth of field. 5. Text: the exact words in quotes, where they sit and the type style. 6. Keep: what must stay as it is from the references or the last result. **Template** ```text Result: [what it is and where it will be used] Subject: [the product or person, with the details that matter] Setting and light: [surface, background, light source and direction] Camera: [angle, distance, lens] Text: "[exact words]" in [type style], [position]. Render it once. Keep: [what must not change] ``` The labels are optional. Plain sentences in the same order work as well. Use whatever is easiest to edit on the next try. ## Using references GPT Image takes up to 16 reference images. Give each one a job and call it by its position: "the bag in image 1", "the model in image 2", "the light in image 3". Then say how they combine: what goes where, and what comes from which image. In Studio, references keep the order you added them in. In Spacelab, wire them into an [Image Model](https://sloosh.ai/docs/nodes-generate.md) node and set Input labels (Garment, Model, Background) so the prompt can name them. More in [References](https://sloosh.ai/docs/references.md). An edit prompt has two halves: the change, and everything else. Write "change only" the one thing, then list what stays: face, pose, camera angle, light, the label. Edits drift over a few rounds, so repeat the keep list every time. ## Before and after ### Product on white **Before** ```text nice photo of my serum bottle on white background, high quality, 4k ``` **After** ```text Result: a clean e-commerce photo of the serum bottle in image 1, for a product page. Setting and light: seamless white background, soft light from above and to the left, a faint shadow under the bottle. Camera: straight on at bottle height, the bottle filling about two thirds of the frame. Keep: the bottle's shape, the dropper and the label text exactly as in image 1. ``` The first prompt asks for a quality it can't measure. The second says where the light comes from, how big the bottle sits and that the label stays readable. Set 4K in the size chip, not in the prompt. ### Ad with a headline **Before** ```text summer sale ad for sunglasses with text ``` **After** ```text A 4:5 Instagram ad. The sunglasses from image 1 rest on a striped beach towel in hard midday sun, crisp shadows, bright saturated colours. The top third is empty pale sand. In that space, the headline "SUMMER SALE" in heavy black sans serif, and under it "Up to 40% off" in smaller type. Render each line once and add no other text. ``` ### Garment on a model **Before** ```text make the woman wear this jacket ``` **After** ```text Image 1 is the jacket. Image 2 is the model. Put the jacket from image 1 on the model in image 2, zipped halfway. Change only her top. Keep her face, hair, pose, the background and the light exactly as in image 2. Keep the jacket's colour, the chest logo and the stitching as in image 1. ``` ### New background **Before** ```text change background to kitchen ``` **After** ```text Change only the background of image 1 to a bright kitchen counter in white stone, with a plant out of focus behind. Keep the blender, its size, position and the camera angle. Match the new light to the shadow already falling to the right. ``` ## Common mistakes - Mood words in place of details. Swap "premium", "stunning" and "cinematic" for light, colour and materials. - Text without quotes. Quote every word you want, say it appears once, and spell out brand names that are easy to get wrong. - Edits without a keep list. The model changes whatever you didn't protect. - Details that don't matter. Every extra clause is one more thing the model trades against the ones that do. - Treating lens names as physics. "85mm" gives the look of a portrait lens. It doesn't simulate one. ## Settings and limits in Sloosh - Aspect ratio: Auto, 1:1, 3:2, 2:3, 16:9, 9:16, 4:5 or 5:4. - Resolution: 1K, 2K or 4K. - Quality: low, medium or high on GPT-image-2. The 2.5 models add xhigh and max. Higher settings are slower, and they don't help every prompt. Try them when small text or fine detail won't come out right. - Up to 16 reference images. - Start on Flare for drafts and volume. Move to Sunburst when a detail won't hold. OpenAI's guide covers transparent backgrounds. Sloosh doesn't offer that setting on GPT Image, so generate on a plain background and use Remove BG when you open the result. ## Further reading [OpenAI: Image prompting guide](https://developers.openai.com/api/docs/guides/image-prompting). In these docs: [Prompting Nano Banana](https://sloosh.ai/docs/prompting-nano-banana.md) and the general [Prompt guide](https://sloosh.ai/docs/prompt-guide.md). ## Frequently asked questions ### How do I get GPT Image to spell text correctly? Put every word in quotes, say where it sits and in what type style, and say it appears once, as in "SUMMER SALE" in heavy black sans serif. Spell out brand names that are easy to get wrong. If small text still won't come out, try a higher quality setting. ### How do I edit an image with GPT Image without changing everything? Write the edit in two halves: "change only" the one thing, then a keep list of what stays, such as the face, pose, camera angle, light and label. GPT Image edits drift over a few rounds, so repeat the keep list in every prompt. ### How do I combine several reference images in a GPT Image prompt? Give each reference a job and call it by position: "the bag in image 1", "the model in image 2", "the light in image 3". Then say how they combine: what goes where, and what comes from which image. GPT Image in Sloosh takes up to 16 references. ### Can GPT Image make a transparent background in Sloosh? No. Sloosh doesn't offer a transparent background setting on GPT Image. Generate on a plain background instead, then open the result in Sloosh and use Remove BG from the edit toolbar to cut the subject out.