
Ideogram 4.0 Official Specs: Open Weights, JSON Prompts, and Layout Control
- Ideogram
- Image Generation
- AI
Ideogram 4.0 is an image-generation model that was officially announced on June 3, 2026. This article does not assume an unreleased successor. It compares, side by side, the architecture, prompt format, and evaluation scores published in the official technical blog; an independent leaderboard checked on November 11, 2026 (Asia/Shanghai); and how the public workspace distinguishes 4.0 from 4.5.
According to the official blog, Ideogram 4.0 is a 9.3-billion-parameter, 34-layer Diffusion Transformer, and its weights are publicly available. Prompts can be structured in JSON, with bounding boxes specified as [y_min, x_min, y_max, x_max] in a 0–1000 coordinate system. Text elements have types, and hexadecimal color values can be supplied for up to 16 colors. The output resolution is 256–2048 pixels.
Key Points First
- The official announcement was on June 3, 2026. The open-weight 4.0 and the editing-focused 4.5, announced on September 30, 2026, are separate official model cards.
- The official scores (the figures in the blog body) are 0.69 for layout (7Bench mIoU), 0.97 for English text (X-Omni OCR), 0.76 for spatial understanding, and 0.89 for prompt adherence.
- On AA-Image-T2I v2.1 (checked on 2026-11-11), Ideogram 4.0 Quality has an Elo of 1027 ± 10 (10,937 samples, rank range 26–40), while Ideogram 4.5 High has 1016 ± 14 (3,199 samples, 31–48). The ranges overlap.
- The public workspace presents 4.0 as being for text and composition generation and 4.5 as being for precise, continuous editing, while stating that it does not run the native 4.5 on this page.
Conclusion of This Article
- The official basis for the text and layout figures is the body-text data in the 4.0 technical blog. No additional scores are inferred from the radar chart’s tick marks.
- 4.5 is officially the editing-focused flagship. The fact that its T2I Elo is slightly lower than 4.0 Quality’s and that the ranges overlap is a reason not to call 4.5 a lower-tier version based on generation rankings alone.
- Do not treat the workspace’s “Ideogram 4” label as evidence that 4.5 is running on the same screen.
Official Specifications List
| Item | What can be confirmed from official sources | Source |
|---|---|---|
| Announcement | 2026-06-03 | Ideogram 4.0 technical blog |
| Scale | 9.3 billion parameters, 34-layer DiT, open weights | Same blog |
| Prompt | JSON, bbox [y_min, x_min, y_max, x_max] (0–1000), styled text, up to 16 colors | Same blog |
| Resolution | 256–2048 px | Same blog |
| Official scores | Layout 0.69, text 0.97, spatial 0.76, prompt 0.89 | Same blog body |
| 4.5 | 2026-09-30. Fine-grained multi-turn editing, native 2K, up to 5 reference images for generation, open weights “coming soon” | Official 4.5 page |
| AA-Image-T2I v2.1 | 4.0 Quality 1027 ± 10, 4.5 High 1016 ± 14 | Verified on 2026-11-11 |
Source: June 3, 2026, Ideogram's official Ideogram 4.0 technical blog, Figure 2. The 0.69 / 0.97 / 0.76 / 0.89 values in the body are quoted from the blog's prose and are not digitized readings of the figure's scales.
What the official blog gives designers
The official account of 4.0 focuses less on whether it can “generate an image” than on whether text-bearing designs such as headlines, logos, and posters can be controlled as layouts. JSON prompts, bboxes, and color specifications are the official interfaces for exercising that control without relying on natural-language instructions alone. The 7Bench layout mIoU of 0.69 and X-Omni English OCR score of 0.97 are figures the official source itself supplied in support of that claim. They differ from the independent leaderboard's Elo scores in both units and tasks. When citing both, keep their sources separate.
The radar chart in the same blog compares 4.0 with 3.0, GPT Image 2, and Nano Banana 2. Because its axes do not have readable scales, it would be more accurate not to create a second score table from the figure.
Why cover 4.5 in the same article
Official 4.5 (September 30, 2026) emphasizes precise multi-turn editing, native 2K, up to five references during generation, and in-place typography and translation. The official wording for open weights is “coming soon”; it does not state that they have already been released, as they were for 4.0. In AA-Image-T2I v2.1, 4.5 High scores 1016 ± 14, slightly below 4.0 Quality at 1027 ± 10, and the intervals overlap. Since 4.5 is the official flagship focused on editing, T2I generation Elo alone cannot determine which is better overall.
How to read the public workspace
Ideogram 4 Free Online AI Image Generator: Logos & Posters guides you through Ideogram 4.0 typography and layout for logos, posters, advertisements, and product images. The model list includes 4.0 and 4.5, while the FAQ describes 4.5 as being for precise continuous editing. The page states that it is a guide using official examples and that the linked service must run the native Ideogram 4.5 model. Even if 4.5 appears in the workspace name, that is not evidence that the official 4.5 endpoint is working.
If you are using official 4.0 as the starting point for production, a workflow close to the intended use of the official specifications is to first lock in headings, colors, and text positions in JSON or bounding boxes, then perform a final check of the results for typos and hierarchy in a design tool. If precise 4.5 editing is required, you need to separately verify a channel through which 4.5 is actually officially offered.
Frequently asked questions
Has Ideogram 4.0 been officially released?
Yes. It was announced in a technical blog post and press release on June 3, 2026, and its weights have also been released.
Where are the official scores cited from?
They are cited from the body of the technical blog post: Layout 0.69, English OCR 0.97, Spatial 0.76, and Prompt Adherence 0.89. No additional values were read from the chart scales.
Since 4.5 is newer, is it also better at generation?
The official team positions 4.5 as a flagship focused on editing. In the independent T2I Elo, 4.0 Quality is slightly higher, and the intervals overlap. Generation and editing should not be collapsed into a single ranking.
Does 4.5 work in this workspace?
The page itself states that it requires the native version of 4.5 to run.
Public records to watch next
The key official records to watch are whether the open weights for 4.5 actually appear and whether an independent leaderboard breaks out editing tasks specifically for 4.5. What is currently established is that the official typography and layout specifications are for 4.0, while 4.5 has a separate official card.
Go to list of users who liked
Go to list of comments
Register as a new user and use Qiita more conveniently
- You get articles that match your needs
- You can efficiently read back useful information
- You can use dark theme What you can do with signing up

