GPT Image 2.5 Prompt Guide: Sketch, Edits, and Video

GPT Image 2.5 follows a brief more like a production spec than a mood board. OpenAI’s own prompting guide is the best primary source: name the result, describe what is visible, quote exact text, and — for edits — say what must not change.

This article turns that guidance into prompts you can reuse in ChatGPT Images 2.5 or the API, then hand the still to PixVerse as a first frame. For the model overview, see What is GPT Image 2.5?. For API routing, see Flare vs Sunburst.

PixVerse now includes GPT Image 2.5 Flare, GPT Image 2.5 Sunburst, and GPT Image 2. The prompt patterns below work on those OpenAI image models. You can generate a 2.5 still in the PixVerse workspace and continue with image-to-video there. If you generate in ChatGPT or the OpenAI API, download the frame and upload it. A 14-day launch offer made Flare and Sunburst available at no credit cost to eligible overseas PixVerse members. As of September 27, 2026, do not assume that offer is still active, and do not treat either model as free or unlimited. Check the credit price shown in your workspace. Availability follows the GPT Image 2.5 model page and PixVerse’s September 20, 2026 review.

The Brief That Works

OpenAI’s fundamentals, compressed into a reusable block:

Result: [poster / pack shot / UI mock / comic / video first frame] for [use case]. Subject: [who or what, visible details, pose, scale]. Exact text: “[copy in quotes]” — appear once, no extra words. Composition: [framing, placement, negative space, aspect ratio]. Style and light: [medium, lighting, materials, what “photoreal” means here]. Constraints: [do not change X; no watermark; no extra logos]. Output: [size, transparent or opaque, video-ready / print / social].

The important shift from GPT Image 2 prompting: separate the change from the lock. “Make it winter” is incomplete. “Make it a winter evening with snowfall. Keep the billboard copy, product, camera angle, and layout unchanged” is a 2.5-style edit.

Pick the Surface Before the Prompt

Surface	Use it when	Prompt implication
ChatGPT Sketch (`@Sketch`)	Layout is easier to draw than describe	Sketch the geometry; put style, materials, and exclusions in text
ChatGPT templates	Poster, merch, product photo, icons	Answer the template questions; do not fight the format
ChatGPT comments	One local change on an existing still	Click the region; write a short instruction; do not restated the whole scene
API Flare	Volume and latency	Keep quality explicit (`medium`/`high`); iterate fast
API Sunburst	Finals and hard edits	Same prompt as Flare so you can compare; raise quality only if it fails
PixVerse Flare / Sunburst	Stills that should stay in the video workspace	Same structured brief on Flare or Sunburst, then image-to-video without exporting; GPT Image 2 remains in that workspace

TechRadar’s product-photo template is a good example of letting the UI collect the brief: upload the product, pick Clean Studio / Editorial / Lifestyle / Dramatic, pick a setting, then generate. Use a custom prompt when the template cannot express a constraint (exact SKU label, locked camera, no extra type).

Generation Prompts

These follow OpenAI’s documented techniques. Adapt the nouns; keep the structure.

Photoreal product hero (video first frame)

Create a photorealistic 16:9 product photograph for an image-to-video ad.
Main subject: a matte black wireless earbud case, lid half open, one earbud hovering just above the cradle.
Exact text: none. No logos, no watermarks, no pack copy.
Composition: product centered, large in frame, generous left-side negative space for later captions, tabletop at chest height.
Style and lighting: clean studio, soft key from camera left, faint rim light, shallow depth of field, real plastic and metal texture, no glamor retouching.
Constraints: both earbuds identical; hinge aligned; no extra props.
Output: 16:9, opaque background, still camera, ready to animate with a slow orbit.

A first frame for PixVerse should keep the subject large, unobstructed, and lit consistently. Motion prompts later describe camera and material movement, not a second product description. See how to turn product photos into video ads.

Poster with quoted type

Create a concert poster for a live jazz night.
Exact text (verbatim, once each):
Title: "AFTER HOURS"
Subhead: "Friday 10pm — Room 12"
Footer: "All ages"
Typography: bold condensed sans for the title, smaller grotesque for the subhead, generous tracking, high contrast, fully legible.
Scene: a single trumpet in warm spotlight, smoke catching the beam, deep navy background.
Constraints: no extra copy, no sponsor marks, no QR codes, no watermarks.
Output: 2:3 portrait, print-like layout with clear margins.

Quote the strings. Ask for them once. Then check spelling in the file. OpenAI recommends medium or high quality for small or dense type.

Transparent asset

In the API, transparency is a parameter, not a metaphor. Set background="transparent", output_format="png" or webp, and say so in the prompt:

Extract the kettle from the input photo and isolate it on a fully transparent background.
Centered product, crisp silhouette, no halos or fringing.
Preserve geometry and label lettering exactly.
Do not add a solid backdrop, checkerboard, drop shadow, or scenery.

OpenAI’s examples use the same pattern for logos and pack cutouts. Inspect the alpha channel on hair, glass, and edges. A drawn checkerboard means the model painted a background.

Sketch to render

In ChatGPT, draw the layout with @Sketch, then:

Use this sketch as the layout and perspective reference.
Turn it into a photorealistic modern living room: light oak floor, sage accent wall, daylight from the window shown in the sketch.
Match the sketch's room proportions and furniture positions.
Do not add furniture, people, or text that is not in the sketch.

In the API, attach the drawing as an input image and use the same lock language. OpenAI’s valley-drawing example is the canonical version: preserve layout, add plausible materials and light, do not invent new elements.

Edit Prompts

2.5’s selling point is local control. Write edits as one change + a lock list.

Change one object

In this kitchen photo, replace ONLY the white chairs with light oak chairs.
Preserve camera angle, room lighting, floor shadows, and surrounding objects.
Keep all other aspects of the image unchanged.
Photorealistic contact shadows and wood grain.

Preserve identity, change clothes

Dress the person using the provided clothing images.
Do not change face, facial features, skin tone, body shape, pose, or identity.
Preserve likeness, expression, hairstyle, and proportions.
Replace only the clothing; fit garments to the existing pose with realistic fabric.
Match lighting and color temperature to the original photo.
Do not change background, camera angle, or framing. No extra accessories, text, or logos.

Translate layout text

Translate the infographic text to Spanish.
Do not change any other aspect of the image: icons, colors, arrows, spacing, or diagram structure.

Multi-turn without drift

OpenAI’s billboard example is the right rhythm:

Build the scene with quoted copy: "Fresh and clean".
Next turn: Make it look like a winter evening with snowfall. — pass the previous output as input.
Restate critical locks if anything starts to slide (copy, product, camera).

If a region must stay pixel-identical, composite it in an editor. Prompting is not a lossless patch tool, even with 2.5.

In ChatGPT, prefer comments for this class of edit: drop a pin on the sky and write “snowfall, keep the billboard type.” TechRadar and The Verge both found comments faster than rewriting the scene.

From Still to PixVerse Video

A 2.5 still is only half of a social or ad workflow. Once the frame is locked:

Export at the delivery aspect ratio (16:9, 9:16, or 1:1). Avoid a later crop that changes composition.
In PixVerse, open image-to-video on the still if it was made there. Upload the file only when it came from ChatGPT or another tool.
Prompt motion, not the product again: camera move, material motion, lighting shift.
Keep duration short for the first test, then extend or restyle if the action holds.

Example motion prompt on top of the earbud still:

Slow 30-degree orbit around the case, lid already open.
The hovering earbud rotates in place with a faint specular glide.
Soft studio light stays consistent. No new props, no on-screen text, no camera shake.

For character-led clips, generate a consistent character sheet with 2.5 first (same outfit, same face locks across poses), then animate in PixVerse. See how to create consistent characters with AI.

If the job does not need 2.5, GPT Image 2, Nano Banana 2, or Seedream inside PixVerse still avoid an extra export. If it does need 2.5, Flare and Sunburst are in that same workspace, so an export is optional rather than required.

Quality, Size, and Cost Knobs

Set these outside the prompt in the API:

model: gpt-image-2.5-flare or gpt-image-2.5-sunburst
quality: auto, low, medium, high, xhigh, max
size: auto or WIDTHxHEIGHT
background: auto, opaque, transparent

OpenAI’s advice: choose the model first, keep quality fixed while comparing models, then raise quality only if the job still fails. xhigh and max are new versus GPT Image 2; they are not a default “make it better” switch.

Listed token rates match GPT Image 2, but 2.5 token counts are not estimated by the GPT Image 2 calculator. Time and cost your own sizes. Details: image generation pricing.

Check the File Before You Animate It

OpenAI’s acceptance list is also a good video gate:

Required text accurate and legible?
Identity, product shape, and labels intact?
Edit changed only what you asked?
If you needed transparency, is there a real alpha channel?
Is the subject large enough to survive compression and motion?

If the still fails that list, do not spend video credits on it.

FAQ

Should prompts be longer for 2.5?

They should be more structured, not necessarily longer. Named sections beat a paragraph of adjectives. OpenAI says short prompts, tagged blocks, and JSON-like structures can all work; pick the format you can maintain.

Can I prompt Sketch from the API?

Not as @Sketch. Pass a drawing as a reference image and say it controls layout and perspective.

Does this replace the GPT Image 2 prompt guide?

No. The GPT Image 2 prompt guide still applies to GPT Image 2, including on PixVerse. Use this article for Images 2.5, Flare, or Sunburst, including when those models run in the PixVerse workspace, or when you want Sketch- and comment-aware briefs.