# Ideogram 4: layout by bounding boxes, JSON prompts and open weights — what Ideogram 4.0 and 4.5 can do

> Ideogram 4 is the current generation of image models from the Canadian company Ideogram. Ideogram 4.0 (June 3, 2026) is the company's first model with open weights: 9.3B parameters, 2K output, text and objects placed by bounding boxes from a JSON prompt. Ideogram 4.5 (September 30, 2026) edits finished images by copying untouched pixels. API prices start at $0.03. Ideogram is available in the AI2Face generator.

Developer: Ideogram · Released: 2026-06-03 · Updated 2026-10-28
URL: https://ai2face.com/models/ideogram-4

## Key facts

| | |
|---|---|
| Developer | Ideogram (Toronto, San Francisco) |
| Versions | Ideogram 4.0 (June 3, 2026), 4.0 Transparent, Ideogram 4.5 (September 30, 2026) |
| Resolution | 1K or 2K; 4.0 Transparent up to 8K via upscaling |
| Prompt | Text or JSON with bounding boxes and a palette of up to 16 colors; up to 10,000 characters in 4.5 |
| References | 4.0 — none; 4.5 — up to 5 images |
| Open weights | Ideogram 4.0, 9.3B parameters, fp8 and nf4, non-commercial license |
| API price | 4.0 — $0.03 / 0.06 / 0.10 (Turbo / Default / Quality) |
| Arena rank | 4.0 — 20th in generation (Oct 26, 2026) |
| In AI2Face | Yes, in the generator |

## What is Ideogram 4

Ideogram 4 is the current generation of image models from Ideogram, a company founded in 2023 by former Google Brain researchers. Since its first version, Ideogram has specialized in text inside images. An overview of the company and all versions, including Ideogram 3.0 with its styles, is on the [Ideogram](/models/ideogram) page.

The generation includes three models:

- **Ideogram 4.0** — generation from scratch, released June 3, 2026;
- **Ideogram 4.0 Transparent** — the same model, with no background out of the box;
- **Ideogram 4.5** — precise editing of finished images, released September 30, 2026.

Ideogram 4.0 is the company's first open-weight model. On the Arena leaderboard as of October 26, 2026, 4.0 in Quality mode ranks 20th of 82 in image generation.

## What Ideogram 4.0 can do

**Text that looks designed.** 4.0 renders headlines, packaging copy and signage, including small type and multiple languages. In a blind ContraLabs test, ten professional designers picked the best typography from four models: Ideogram 4.0 won 47.9% of the time, [Nano Banana 2](/models/nano-banana-2) 30%.

**A bounding box for every element.** Instead of guessing the composition, 4.0 places the headline, logo, caption and main subject where you drew boxes for them. Boxes are set in a JSON prompt — see the example below.

![A BREAD SCHOOL workshop poster with croissants: red dashed boxes around the headline, logo, photo and caption, labeled title, logo, subject, caption](https://ai2face.com/_astro/layout.BZXbuXxF_139R8D.jpg "What a box-based layout looks like: each element has its own place. Diagram made with GPT Image 2.5 Sunburst")

**JSON prompts.** Ideogram trained 4.0 on structured JSON captions, so that's what the model understands best. A plain text prompt works too: the service converts it into JSON itself (the Magic Prompt feature).

**2K and wide formats.** Output is 1K or 2K, with very wide aspect ratios like 23:9 and 3:1 — for banners and headers.

**Transparent background.** 4.0 Transparent renders with no background — stickers, icons, prints. Output goes up to 8K; anything above 2K is upscaled.

![A sheet of six vinyl stickers with white borders: COFFEE FIRST, GOOD VIBES, STAY SHARP, LET IT FLY, NAP TIME, KEEP GOING](https://ai2face.com/_astro/stickers.zljADHn2_Z2jFl5x.jpg "Stickers with short slogans — a job for 4.0 Transparent. Made with GPT Image 2.5 Sunburst")

**Three speed modes.** Turbo, Default and Quality — from draft to final.

## Ideogram 4.0 vs 4.5

| | Ideogram 4.0 | Ideogram 4.0 Transparent | Ideogram 4.5 |
|---|---|---|---|
| Released | Jun 3, 2026 | 2026 | Sep 30, 2026 |
| Purpose | generation from scratch | images with no background | editing and generation with references |
| Modes | Turbo, Default, Quality | Turbo, Default, Quality | Very low, Low, Medium, High |
| Resolution | 1K, 2K | 1K, 2K, 4K, 8K (upscaled) | 1K, 2K or source size |
| References | none | none | up to 5 |
| Background | standard | transparent | auto, transparent or opaque |
| API price | $0.03 / 0.06 / 0.10 | — | from a prompt $0.03 / 0.06 / 0.10; editing $0.008 / 0.03 / 0.06 / 0.22 |
| Open weights | yes | — | promised |

**Ideogram 4.5** is for precise editing. The model returns an image of the same size and copies the pixels an edit doesn't touch instead of redrawing them. That's why the tenth edit in a row doesn't pile up artifacts, pixel shifts or color drift. You can limit an edit with a mask. Ideogram also shows text translation on a finished layout — say, a banner into French — with the design preserved, plus editing crops of large images that drop back into place without seams. On Arena's image editing leaderboard as of October 24, 2026, 4.5 ranks 21st.

## Ideogram 4.0 open weights

The Ideogram 4.0 weights are on Hugging Face and the code is on GitHub:

- architecture — a 9.3B-parameter DiT trained from scratch;
- the prompt is read by a Qwen3-VL-8B encoder;
- weight variants — fp8 and nf4; nf4 runs on a single 24 GB GPU and works in Diffusers;
- inference code — Apache 2.0; weights — the Ideogram 4 Non-Commercial license.

For research and personal projects the weights are free. Running it commercially on your own server requires a separate agreement with Ideogram. Through the API, you can use the model commercially without one.

## What Ideogram 4 is good for

- **Posters and flyers** with several text blocks.
- **Banners and headers** in wide formats.
- **Packaging and labels** — accurate text on the product.

![A jar of honey on a wooden table with a kraft label reading WILD HONEY, Raw and unfiltered, 350 g, and a bee drawing](https://ai2face.com/_astro/label.Dz3V_3D-_ZyzWXG.jpg "A label with three levels of text — a typical Ideogram 4.0 job. Made with GPT Image 2.5 Sunburst")

- **Stickers and prints** — 4.0 Transparent.
- **Editing and translating text** on a finished layout — 4.5.
- **Your own server** for non-commercial projects — the 4.0 open weights.

## Ideogram 4 vs alternatives

| | Ideogram 4.0 | [GPT Image 2.5](/models/gpt-image-2-5) | [Qwen Image](/models/qwen-image) | [Recraft](/models/recraft) V4.1 |
|---|---|---|---|---|
| Arena rank (Oct 26, 2026) | 20 | 1 (Sunburst), 2 (Flare) | 12 (3.0 Pro) | 26 (Utility Pro) |
| Strength | layout by boxes | overall quality and text | long prompts, small text | SVG vectors |
| Resolution | up to 2K | up to 3840 px | up to 2048 px | 1024 / 2048 px |
| Transparent background | yes (Transparent) | yes | yes (2.1) | vector |
| Open weights | yes, non-commercial | no | 2.1 — yes | no |
| API price | from $0.03 | per token | — | — |

On overall quality, Ideogram 4.0 trails GPT Image 2.5 and Qwen Image 3.0 Pro. People choose it for predictability: text and objects land where you specify, not where the model decides.

## How to write prompts for Ideogram 4

An Ideogram 4.0 JSON prompt has three fields:

- `high_level_description` — what the image is overall;
- `style_description` — style, lighting, palette;
- `compositional_deconstruction` — the required list of elements.

Each element has a type (`text` or `obj`), a description `desc` and a box `bbox` in the format `[y_min, x_min, y_max, x_max]` on a 0–1,000 grid. For text, `desc` holds the exact string. You can add a palette of up to 16 HEX colors.

An example for the poster above:

```json
{
  "high_level_description": "A bakery workshop poster, cream background, a photo of croissants",
  "style_description": "Editorial minimalism, dark brown serif type, soft daylight",
  "compositional_deconstruction": [
    {"type": "text", "desc": "BREAD SCHOOL", "bbox": [40, 60, 280, 760]},
    {"type": "obj", "desc": "a round logo with a croissant silhouette", "bbox": [40, 790, 180, 940]},
    {"type": "obj", "desc": "two croissants on a light plate", "bbox": [300, 60, 840, 940]},
    {"type": "text", "desc": "Saturday 10 am", "bbox": [870, 250, 950, 750]}
  ]
}
```

If you write plain text:

- **Start with the layout type:** "poster," "jar label," "podcast header."
- **Put text verbatim in quotes,** each line separately.
- **Lead with what matters:** earlier words carry more weight.
- **Don't paste other tools' parameters** like `--ar 16:9` — Ideogram will treat them as text.

## How to use Ideogram 4 in AI2Face

Ideogram is in the AI2Face generator next to [GPT Image 2.5](/models/gpt-image-2-5) and [Nano Banana Pro](/models/nano-banana-pro).

1. Sign up — new accounts get starter credits.
2. Open the generator and pick Ideogram.
3. Write a prompt with the exact text or upload an image to edit.
4. The credit cost is shown before you start. Credits don't expire.

It's handy to generate a poster from the same prompt in Ideogram and GPT Image and keep the version with the cleanest text.

## Limitations

- **Not the quality leader.** 20th on Arena: in photorealism and complex scenes Ideogram trails [GPT Image](/models/gpt-image) and [Nano Banana](/models/nano-banana).
- **Non-Latin scripts are weaker than Latin** — check Cyrillic and other scripts letter by letter.
- **No references in 4.0** — for a character or a product from a photo you need 4.5.
- **Styles only in 3.0.** Presets, style references and negative prompts didn't carry over to 4.0 and 4.5.
- **Weights aren't for business use.** Running 4.0 commercially yourself requires an agreement with Ideogram.
- **The 4.5 weights** aren't open yet.

## Bottom line

Ideogram 4 is built for designers who need predictability: 4.0 puts the headline, logo and product exactly into the boxes you set, and 4.5 edits a finished layout without damaging the rest. On top of that, 4.0 is one of the few models with open weights and strong typography. Use it for posters, labels, banners and stickers. If overall image quality matters more, go with [GPT Image 2.5](/models/gpt-image-2-5); for vectors, [Recraft](/models/recraft). You can compare them on the same prompt in AI2Face.

## FAQ

### What can Ideogram 4.0 do?

It generates 1K and 2K images with accurate text and places headlines, logos and objects in set bounding boxes from a JSON prompt. It's Ideogram's first open-weight model: 9.3 billion parameters, and the nf4 version runs on a GPU with 24 GB of memory.

### How is Ideogram 4.5 different from 4.0?

4.5 is an editing model. It changes only what you specify and copies untouched pixels from the source, so repeated edits don't build up artifacts. 4.5 also takes up to 5 reference images and prompts up to 10,000 characters.

### How do you write a JSON prompt for Ideogram 4?

Three fields: an overall description, a style, and a required list of elements. Each element has a type (text or object), a description and a bounding box in coordinates from 0 to 1,000. For text, the description holds the exact string to render.

### Can I download Ideogram 4?

Ideogram 4.0 — yes: the weights are on Hugging Face in fp8 and nf4, and the code is on GitHub under Apache 2.0. The weights are free for non-commercial use only. Ideogram has said it will open the 4.5 weights later.

### How much does Ideogram 4 cost?

In the API, Ideogram 4.0 costs $0.03 in Turbo, $0.06 in Default and $0.10 in Quality. Ideogram 4.5 from a prompt is $0.03, $0.06 and $0.10; editing with references ranges from $0.008 to $0.22.

### What's the latest version of Ideogram?

Ideogram 4.5 from September 30, 2026 — for editing and characters from references. For generating from scratch with bounding boxes, Ideogram 4.0 is current.

## Sources

- https://ideogram.ai/news/ideogram-4.0/
- https://ideogram.ai/models/4.0/
- https://ideogram.ai/models/4.5/
- https://github.com/ideogram-oss/ideogram-4
- https://docs.ideogram.ai/using-ideogram/generation-settings/available-models
- https://developer.ideogram.ai/api-reference/images/generate/ideogram-4-5
- https://arena.ai/leaderboard/text-to-image
- https://arena.ai/leaderboard/image/image-edit