Prompt Patterns for gpt-image-2 and seedance-2, and How to Run Them Through One API

Amany Syori — Wed, 04 Nov 2026 03:55:16 +0000

Most prompt guides stop at the prompt. This one goes a step further. It covers prompt patterns for an image model (gpt-image-2) and a video model (seedance-2), then shows how to run them from code so you can iterate systematically instead of copy-pasting into a UI.

Both models are reachable through ToAPIs, an OpenAI-compatible gateway with one API key for text, image, video and audio models. You can test prompts in its playground first, then script them.

Part 1: Image prompts (gpt-image-2)

A reliable structure for image prompts:

[Subject] + [Setting] + [Style/Medium] + [Lighting] + [Composition] + [Details/Constraints]

Pattern A: Product shot

A ceramic coffee mug with a matte sage-green glaze on a light oak table,
soft studio photography, diffused window light from the left,
centered composition with generous negative space, no text, no logos.

Pattern B: Editorial illustration

Flat vector illustration of a developer reading documentation beside a window,
limited palette of navy, coral and cream, subtle paper texture,
wide 16:9 banner composition, friendly and calm mood.

Pattern C: Style transfer by description

A city street at night rendered as a 1970s travel poster:
bold geometric shapes, grainy print texture, warm sunset colors,
large empty area at the top for a headline.

Tips that tend to help:

State what you don't want ("no text", "no watermark") explicitly.
Name the medium ("watercolor", "35mm photo", "isometric 3D").
Describe the light. It changes the mood more than almost any other word.
Keep one idea per prompt. Split complex scenes into separate generations.

Part 2: Video prompts (seedance-2)

Video adds time, so the prompt needs motion and camera:

[Subject] + [Action] + [Setting] + [Camera movement] + [Lighting/Mood] + [Pacing]

Pattern D: Establishing shot

A small fishing village at dawn, mist over the harbor, boats gently rocking,
slow aerial drone shot moving forward, soft golden light, calm pacing.

Pattern E: Character action

A barista pours latte art into a white cup, close-up on the hands,
static camera with shallow depth of field, warm café lighting,
steam rising slowly.

Pattern F: Product reveal

A pair of running shoes rotates slowly on a pedestal,
camera orbits 180 degrees, clean white background,
crisp studio lighting, smooth and steady motion.

Video-specific tips:

Describe one main motion. Several competing actions often look chaotic.
Always specify the camera: static, pan, orbit, dolly or drone.
Words like "slow", "gentle" and "steady" make it easier to get coherent movement.

Part 3: Running prompts from code

Here's a small harness to batch-test the image patterns:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["TOAPIS_API_KEY"],
    base_url="https://toapis.com/v1",
)

prompts = {
    "product_mug": "A ceramic coffee mug with a matte sage-green glaze on a light oak table, soft studio photography, no text.",
    "editorial": "Flat vector illustration of a developer reading documentation beside a window, navy, coral and cream palette.",
}

for name, prompt in prompts.items():
    result = client.images.generate(model="gpt-image-2", prompt=prompt)
    item = result.data[0]
    print(name, "->", item.url or "base64 image returned")

And a single video request:

import requests

resp = requests.post(
    "https://toapis.com/v1/videos/generations",
    headers={"Authorization": f"Bearer {os.environ['TOAPIS_API_KEY']}"},
    json={
        "model": "seedance-2",
        "prompt": "A barista pours latte art into a white cup, close-up on the hands, static camera, warm café lighting.",
    },
    timeout=60,
)
resp.raise_for_status()
print(resp.json())

Video generation is a longer job. Check the docs for the response format and how to poll for the finished result.

Part 4: Let a text model write the first draft

Use a chat model to expand short ideas into the structures above:

template = (
    "Rewrite the idea as a video prompt with: subject, action, setting, "
    "camera movement, lighting/mood, pacing. One paragraph."
)
r = client.chat.completions.create(
    model="gpt-5.6-terra",
    messages=[{"role": "system", "content": template},
              {"role": "user", "content": "dog surfing at sunset"}],
)
print(r.choices[0].message.content)

A simple iteration loop

Draft with a pattern.
Generate two or three variations, changing one variable at a time (light, camera or style).
Log prompt → result, so you learn what each model responds to.
Promote winners into a reusable prompt library.

Docs for the endpoints: https://docs.toapis.com

PromptZone - AI Prompts, Guides and Tools for Builders: Amany Syori