live
MODEL IDideogram:4.6@0

Ideogram 4.6

Ideogram
by
Available with Zero Data Retention

Ideogram 4.6 is Ideogram's image generation and editing model, with native output up to 4096x4096. It generates images from text at any size on a 32-pixel grid, with an optional transparent background, and edits a source image using up to four more images as references. It keeps characters recognizable across scenes and matches exact colors more reliably than Ideogram 4.5, which suits product shots, brand recoloring, and campaign variations.

Ideogram 4.6

Prompting

How to write prompts for Ideogram 4.6: the order of a good prompt, sizes up to 4096 × 4096 and as wide as 6:1, Magic Prompt and the four quality tiers.

Introduction

Ideogram 4.6 generates an image from a text prompt of up to 10,000 characters, natively at up to 4096 × 4096 pixels. The prompt carries almost all of the control. Three settings sit around it. The size can be any width and height on a 32-pixel grid, Magic Prompt can rewrite the prompt before generation, and the quality tier sets speed and cost.

This guide covers the order that makes a prompt work, choosing a size, what 4K adds, what Magic Prompt does to short and long prompts, and choosing a quality tier.

Building the prompt

Ideogram 4.6 follows a long, specific prompt closely, so write the details in a fixed order. Every prompt in this guide uses the same one: what the image is for, then the subject, the setting, where things sit, the light and the camera.

Commercial product photograph for a 16:9 website hero:, a matte terracotta smart speaker with a woven fabric front, sits on a pale oak sideboard against a warm plaster wall., The speaker sits in the left third of the frame, and the right half is calm, softly lit plaster wall with room for copy., Late afternoon sun through a window casts a soft leaf shadow across the wall., Shot on an 85mm lens at eye level, shallow depth of field.
FormatSubjectSettingCompositionLightCamera
Try in Playground
import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'ideogram:4.6@0',
  positivePrompt: 'Commercial product photograph for a 16:9 website hero: a matte terracotta smart speaker with a woven fabric front sits on a pale oak sideboard against a warm plaster wall. The speaker sits in the left third of the frame, and the right half is calm, softly lit plaster wall with room for copy. Late afternoon sun through a window casts a soft leaf shadow across the wall. Shot on an 85mm lens at eye level, shallow depth of field.',
  width: 4096,
  height: 2304
})
import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "ideogram:4.6@0",
            "positivePrompt": "Commercial product photograph for a 16:9 website hero: a matte terracotta smart speaker with a woven fabric front sits on a pale oak sideboard against a warm plaster wall. The speaker sits in the left third of the frame, and the right half is calm, softly lit plaster wall with room for copy. Late afternoon sun through a window casts a soft leaf shadow across the wall. Shot on an 85mm lens at eye level, shallow depth of field.",
            "width": 4096,
            "height": 2304
        })


asyncio.run(main())
curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "imageInference",
      "taskUUID": "b7e2d4a1-6c93-4f05-8a1e-3d9c7f2b5e60",
      "model": "ideogram:4.6@0",
      "positivePrompt": "Commercial product photograph for a 16:9 website hero: a matte terracotta smart speaker with a woven fabric front sits on a pale oak sideboard against a warm plaster wall. The speaker sits in the left third of the frame, and the right half is calm, softly lit plaster wall with room for copy. Late afternoon sun through a window casts a soft leaf shadow across the wall. Shot on an 85mm lens at eye level, shallow depth of field.",
      "width": 4096,
      "height": 2304
    }
  ]'
runware run ideogram:4.6@0 \
  positivePrompt="Commercial product photograph for a 16:9 website hero: a matte terracotta smart speaker with a woven fabric front sits on a pale oak sideboard against a warm plaster wall. The speaker sits in the left third of the frame, and the right half is calm, softly lit plaster wall with room for copy. Late afternoon sun through a window casts a soft leaf shadow across the wall. Shot on an 85mm lens at eye level, shallow depth of field." \
  width=4096 \
  height=2304
{
  "taskType": "imageInference",
  "taskUUID": "b7e2d4a1-6c93-4f05-8a1e-3d9c7f2b5e60",
  "model": "ideogram:4.6@0",
  "positivePrompt": "Commercial product photograph for a 16:9 website hero: a matte terracotta smart speaker with a woven fabric front sits on a pale oak sideboard against a warm plaster wall. The speaker sits in the left third of the frame, and the right half is calm, softly lit plaster wall with room for copy. Late afternoon sun through a window casts a soft leaf shadow across the wall. Shot on an 85mm lens at eye level, shallow depth of field.",
  "width": 4096,
  "height": 2304
}
Response
[
  {
    "taskType": "imageInference",
    "taskUUID": "b7e2d4a1-6c93-4f05-8a1e-3d9c7f2b5e60",
    "imageUUID": "e4c19a7d-2b56-4e83-9f0a-7d1b3c5e8f24",
    "imageURL": "https://im.runware.ai/image/os/a14d18/ws/2/ii/e4c19a7d-2b56-4e83-9f0a-7d1b3c5e8f24.jpg"
  }
]

Lead with the format. "Commercial product photograph for a 16:9 website hero" names what the image is for first, and the composition follows from it. The empty half of the frame is there because a sentence of its own asked for it and said what the empty part looks like: calm plaster wall. Leave it undescribed and the model fills it with its own choice.

The camera line comes last. An 85mm lens and a shallow depth of field keep the background soft, so the leaf shadow reads as texture behind the copy that goes there later instead of competing with it.

Choosing a size

width and height take any multiple of 32 from 32 to 4096, as long as the ratio stays between 1:6 and 6:1. There is no list of presets to match, so the size comes straight from the placement: a website strip, a story, a feed post or a print. Leave both out and the model picks the size.

Name the format in the prompt as well as in the size. "Ultra-wide website banner" and "for a 9:16 story" tell the model what frame it is composing for, and one sentence on how the subject fills it ("stretching from the left edge to the right edge", "full figure from head to toe") does the rest.

What 4K adds

Rendered from the same prompt and seed at 2048 × 1152 and at 4096 × 2304, the shoe faces the other way at 4K, because the size is part of the composition. Below the full 4K frame, the toe of the shoe is cropped from each result and shown at the same width.

At the same width, the 2K toe goes soft while the 4K crop holds the double row of stitching and the edge of every droplet on the toe cap. That detail is generated, not upscaled, so it survives a print or a crop for a detail tile. Because the size changes the picture as well as the detail, settle the prompt at 2K and choose the final from renders at the size that ships.

Magic Prompt

settings.magicPrompt lets Ideogram rewrite your prompt before generating, expanding a short description into a fuller one. It takes auto, on or off. The three images below come from the same six-word prompt and the same seed.

On a short prompt, the setting decides who makes the choices the prompt left open: the clinic's name, the people in the room, the camera position. All three are usable, and from the same seed each is a different room.

On a long prompt every detail is already written, so rewriting can only add its own choices.

The on version kept the desk, the orchid and the sage chairs and rewrote the rest: four chairs where the prompt asked for three, an abstract print in place of the lake, a console table and a floor lamp, and a lettered glass door framing the shot. The off version kept the three chairs and the lake print, though it hung the print on the back wall rather than above the chairs.

With off, the model gets the prompt exactly as written, which is what you want when the prompt is a brief a client signed off on.

Magic Prompt only works without reference images: a request that sends settings.magicPrompt together with inputs.referenceImages is rejected. See Editing with reference images.

Choosing a quality tier

settings.quality takes very_low, low, medium or high, and medium is the default. The lower tiers return faster and cost less. very_low is limited to 4,194,304 pixels, such as 2048 × 2048, so it is a drafting tier rather than a delivery one. The four results below share a prompt and a seed.

From the same seed, each tier composed the flat lay differently, so a lower tier is not a cheap preview of the high render. What holds across all four is the content: every item the prompt lists, and the stamped word NORTHBOUND spelled right at every tier. Use very_low to check that the items and the wording land, and judge the composition on the tier you will ship.

Tips

  1. Lead with the format. "Product photograph for a 16:9 website hero" sets up every composition decision that follows.

  2. Say where the subject sits and what the empty space looks like. Name the third, then describe the empty part instead of leaving it to the model.

  3. Size from the placement. Any multiple of 32 up to 4096, within 1:6 to 6:1, works, so match the frame to where the image runs and name that format in the prompt.

  4. Render the final at the size it ships. 4K is generated detail, so a print or a detail crop needs it from the start.

  5. Turn Magic Prompt off for a finished brief. Rewriting adds details the brief never asked for.

  6. Check the content on a cheap tier, judge the layout on the one you ship. The same seed composes differently at each tier, so a very_low draft confirms the items and the wording, not the arrangement.

  7. Put exact copy in quotes. Text rendering has its own guide: Text and design.