live
MODEL IDideogram:4.6@precise-edit

Ideogram 4.6 Precise Edit

Ideogram
by
Available with Zero Data Retention

Ideogram 4.6 Precise Edit is the precision editing variant of Ideogram 4.6. It applies a natural-language instruction to a source image, optionally limited by a mask and guided by up to four reference images, and returns the result at the source's exact dimensions. It accepts sources up to 8192x8192, and an edit limited to one region keeps a 4K or 8K source at full resolution. It suits retouching, exact color changes, product detail repair, and masked edits.

Ideogram 4.6 Precise Edit

Editing images

How to edit an image with Ideogram 4.6 Precise Edit: the request, one instruction per change, the output size and the four quality tiers.

Introduction

Ideogram 4.6 Precise Edit applies one natural-language instruction to an existing image and returns the edited image at the source's dimensions. It is built for the edits a product team runs every day: a stain on a packshot, a price on a sign, a prop on a shelf, a lawn that needs to look like spring.

A pair of white leather sneakers with gum soles on a light gray background, the toe of the left one smeared with a green grass stain

Square e-commerce product photograph of a pair of white leather low-top sneakers with white laces and gum rubber soles, side by side at a three-quarter angle on a plain light gray background. The toe of the left sneaker has a visible green grass stain smeared across the leather. Soft even studio light, a gentle shadow below, crisp detail.

SourceResult

Drag the slider across the left toe: the leather grain and the stitching around the stain line up with the source.

This guide covers the request, the instruction, the output size and the quality tiers.

The request

An edit takes two required fields: the image in inputs.seedImage and the instruction in positivePrompt. The source sets the size of the output, so the request has no width or height.

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@precise-edit',
  positivePrompt: 'Remove the green grass stain from the toe of the left sneaker, so the leather there is clean white like the rest of the shoe. Keep everything else in the image exactly the same.',
  inputs: {
    seedImage: 'https://assets.runware.ai/guides/ideogram-4-6-precise-edit/editing-images/source-sneakers.jpg'
  }
})
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@precise-edit",
            "positivePrompt": "Remove the green grass stain from the toe of the left sneaker, so the leather there is clean white like the rest of the shoe. Keep everything else in the image exactly the same.",
            "inputs": {
                "seedImage": "https://assets.runware.ai/guides/ideogram-4-6-precise-edit/editing-images/source-sneakers.jpg"
            }
        })


asyncio.run(main())
curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "imageInference",
      "taskUUID": "1f6c3a85-7e29-4d40-9b13-5a8e2c7f4d06",
      "model": "ideogram:4.6@precise-edit",
      "positivePrompt": "Remove the green grass stain from the toe of the left sneaker, so the leather there is clean white like the rest of the shoe. Keep everything else in the image exactly the same.",
      "inputs": {
        "seedImage": "https://assets.runware.ai/guides/ideogram-4-6-precise-edit/editing-images/source-sneakers.jpg"
      }
    }
  ]'
runware run ideogram:4.6@precise-edit \
  positivePrompt="Remove the green grass stain from the toe of the left sneaker, so the leather there is clean white like the rest of the shoe. Keep everything else in the image exactly the same." \
  inputs.seedImage=https://assets.runware.ai/guides/ideogram-4-6-precise-edit/editing-images/source-sneakers.jpg
{
  "taskType": "imageInference",
  "taskUUID": "1f6c3a85-7e29-4d40-9b13-5a8e2c7f4d06",
  "model": "ideogram:4.6@precise-edit",
  "positivePrompt": "Remove the green grass stain from the toe of the left sneaker, so the leather there is clean white like the rest of the shoe. Keep everything else in the image exactly the same.",
  "inputs": {
    "seedImage": "https://assets.runware.ai/guides/ideogram-4-6-precise-edit/editing-images/source-sneakers.jpg"
  }
}
Response
[
  {
    "taskType": "imageInference",
    "taskUUID": "1f6c3a85-7e29-4d40-9b13-5a8e2c7f4d06",
    "imageUUID": "b4e8d1a7-6c35-4f92-8a0e-2d7b9c5f1e63",
    "imageURL": "https://im.runware.ai/image/os/a14d18/ws/2/ii/b4e8d1a7-6c35-4f92-8a0e-2d7b9c5f1e63.jpg"
  }
]

Three optional fields shape the edit further, each with a guide of its own: reference images to take a product or a logo from, a mask to aim the edit at one area, and an edit region for large sources and small, exact changes.

Writing the instruction

Precise Edit works best with one change per request, stated as an action on a named target. Three verbs cover most of the work: remove, change and add. Each instruction below names the target by what it is and where it is, then closes with what has to stay.

A black food truck menu board listing TACOS, Carne Asada $4, Al Pastor $4, Grilled Fish $5 and Veggie $3.50 in white lettering, with the edit region drawn around the $5

Photograph of the menu board on the side of a food truck: a black board with white hand-lettered style text listing "TACOS" as a heading, then "Carne Asada $4", "Al Pastor $4", "Grilled Fish $5" and "Veggie $3.50", each on its own line, with a small drawing of a chili pepper in the corner. Bright daylight, shot straight-on, the board filling most of the frame with a strip of the red truck around it.

Source and regionChange

Box the text you change. The instruction quotes the old price and the new one, and settings.editRegion draws a box around the "$5", in pixels of the source. Nothing outside a box can change, and here the result matches the source pixel for pixel outside it. "In the same hand-lettered style" keeps the new number in the board's lettering instead of a clean font pasted on top. Without the box, the same kind of change went wrong in three of four runs on this board, adding a second price beside the first or rewriting another line. Boxes are covered in Region edits.

A hotel bathroom with a white floating vanity, a vessel sink, a round mirror on pale green tiles and an oak shelf holding only a rolled towel

Photorealistic hotel bathroom photograph: a white floating vanity with a rectangular vessel sink and a matte black faucet under a round mirror, against pale green zellige tiles, a floating oak shelf to the right of the mirror holding a rolled white towel at one end and empty along the rest of its length. Soft daylight, shot straight-on at standing height with a 35mm lens.

SourceAdd

Place an addition against something already there. "On the empty part of the oak shelf, next to the rolled towel" gives the fern a surface and a neighbor, so its position and its scale both come from the scene. To pin the spot exactly, draw it with a mask.

A white single-story house with a covered porch behind a patchy, dry brown lawn and a concrete path between two hedges

Real estate listing photograph of a single-story suburban house with white siding, a dark gray roof and a covered front porch, seen from the street. The front lawn is patchy and dry, mostly yellowed brown grass with bare spots, and a concrete path leads to the porch between two low hedges. Clear blue sky, midday sun, shot on a 24mm lens at standing height.

SourceChange

List the neighbors in the keep clause when the change covers a large area. The lawn touches the path, the hedges and the porch, so the instruction names each of them: "Keep the house, the porch, the path, the hedges and the sky exactly the same."

The output size

At low, medium and high, the result comes back at the source's exact dimensions, from a phone photo to a 4096 × 4096 master. The edit can be diffed against the original or dropped into the layout the original came from without a resize.

very_low is the exception. It returns at most about 4.2 megapixels, so a larger source comes back scaled down: the 3072 × 2048 driveway below came back at 2496 × 1664.

Above 4096 × 4096, the size of the result depends on settings.editRegion, and only a box keeps the full size. See Region edits.

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. The four results below share the source, the instruction and the seed.

Every tier removed the car. The tiers differ in how cleanly the scene continues where it stood. At very_low the siding behind it carries a lighter patch where the boards do not line up, while low and up continue the boards and the paving joints without a seam. A removal also says what goes in the gap, "continue the paved driveway and the edge of the lawn", so the model fills it with the scene instead of a guess.

Check the instruction at very_low, which returns large sources at a reduced size, then render the one you keep at a higher tier.

Tips

  1. One change per request. Run a second edit on the result for the next change.

  2. Name the target by what and where. "The toe of the left sneaker" points at one spot on one shoe.

  3. Box small text changes. Quote the old and the new text, and draw settings.editRegion around it so nothing else on the layout can change.

  4. Anchor additions to the scene. A surface and a neighbor set both the position and the scale.

  5. Say what fills a removal. "Continue the paved driveway" is the instruction for the gap.

  6. Close with a keep clause. List the neighbors of a large change by name.

  7. Draft at very_low, ship at a higher tier. It is the cheapest way to check that an instruction lands, and it returns large sources at a reduced size.