Ideogram 4 is the first open-weight text-to-image model from Ideogram, released on June 3, 2026, and it solves the problem that has embarrassed image generators for years: text. Posters with garbled words, logos with misspelled names, and signs that read like a dream are the norm for most models. Ideogram 4 renders text with near-perfect accuracy, and it does it while giving you precise control over layout, color, and composition. It has 2,914 stars on GitHub as of November 9, 2026, and it is the top-ranked open-weight model on the Design Arena leaderboard, according to the project's README.
Watch the 30-second repo signal
Key Takeaways
Ideogram 4 is a 9.3B parameter model trained from scratch, not a fine-tune, with native 2K output and best-in-class text rendering among open-weight models.
Instead of plain prompts, it uses structured JSON captions with bounding boxes and color palettes, so you can place text and subjects exactly where you want them.
In a blind evaluation by 10 professional designers, Ideogram 4 won first place on typography 47.9% of the time, ahead of Gemini 3.1 Flash Image Preview at 30.0%.
The inference code is Apache-2.0, but the model weights are under a non-commercial license, so commercial use needs a separate Ideogram license.
The nf4 build runs on a single 24 GB GPU, which makes studio-grade image generation genuinely runnable at home.
The Problem It Solves
Every image model can paint a pretty picture. Very few can spell. Ask most generators for a poster and you get a beautiful layout with "COFEEE" on the sign, or a logo where the brand name dissolves into squiggles. Designers learned to work around this by generating the image, then fixing the text in Photoshop, which defeats the point of generating the design in the first place.
The second problem is control. A plain text prompt is a vague instruction, and vague instructions produce vague layouts. You ask for a headline at the top and a product shot on the left, and the model puts the headline wherever it feels like it. Professional design needs precision: this element here, that color there, this size, that spacing.
Ideogram 4 attacks both problems directly. It was trained exclusively on structured JSON captions that describe every element of an image in detail, including bounding-box coordinates and color palettes. That training makes it unusually good at rendering text, and unusually obedient about where things go. (Ideogram)
What Makes Ideogram 4 Different
The headline difference is the prompting interface. Ideogram 4 expects a structured JSON caption, not a sentence. In that JSON you can name every object, give it a bounding box in normalized 0 to 1000 coordinates, and specify a color palette of up to 16 hex colors. The model was trained on captions exactly like that, so mirroring the format at inference time is the most reliable way to get every requested object rendered in the right place.
If writing JSON by hand sounds tedious, that is what the magic prompt is for. It uses a language model to expand a plain-text prompt into the full structured caption before generation. By default it calls Ideogram's hosted magic-prompt API, which is free and needs only an API key from the Ideogram platform. The magic-prompt system prompt is open source, so you can also run the expansion through your own language model provider. (GitHub)
Under the hood, the architecture is a 9.3B parameter single-stream Diffusion Transformer with 34 layers, trained entirely from scratch. Text and image tokens are concatenated into one unified sequence and processed through the same transformer, with no separate text or image branches. Instead of a text-only encoder like CLIP or T5, Ideogram 4 uses Qwen3-VL-8B-Instruct, a full vision-language model, as its text encoder, with hidden states taken from 13 intermediate layers. The model supports any resolution from 256 to 2048 pixels per side, with aspect ratios up to 6:1, so one model covers square thumbnails, ultrawide banners, and everything in between. (Ideogram)
The benchmark numbers back up the design claims. On X-Omni English OCR, which measures text rendering accuracy, Ideogram 4 scores 0.97. On 7Bench layout control it scores 0.69 mIoU, on SpatialGenEval 0.76, and on Prism prompt alignment 0.89. At 9.3B parameters it delivers the best text rendering of any open-weight release the team benchmarked, ahead of much larger models like Qwen-Image at 20B, FLUX.2 dev at 32B, and HunyuanImage 3.0 at 80B MoE. (Ideogram)
Ideogram 4 at a Glance
Aspect | Detail |
|---|---|
Parameters | 9.3B |
Architecture | Single-stream Diffusion Transformer, 34 layers |
Text encoder | Qwen3-VL-8B-Instruct, 13 intermediate layers |
Trained from | Scratch, not a fine-tune |
Resolution | 256 to 2048 pixels per side, up to 6:1 aspect ratio |
Prompting | Structured JSON captions, bounding boxes, up to 16 hex colors |
GitHub stars | 2,914, as of November 9, 2026 |
GitHub forks | 288, as of November 9, 2026 |
Repo created | May 28, 2026 |
Released | June 3, 2026 |
Code license | Apache-2.0 |
Weights license | Ideogram 4 Non-Commercial |
Variants | nf4 for CUDA with Diffusers support, fp8 for all hardware |
Local GPU need | nf4 fits on a single 24 GB GPU |
How It Stacks Up
The strongest independent evidence comes from ContraLabs, which ran a blind typography evaluation judged by 10 professional designers from Contra's top-earning talent. The test covered 4 models and 240 images, and Ideogram 4 won first place on typography 47.9% of the time, clearing 50% in the second round at 51.1%.
Model | First-place win rate, typography | Would use in real client work |
|---|---|---|
Ideogram 4 | 47.9% | 3.55 out of 5 |
Gemini 3.1 Flash Image Preview (Nano Banana 2) | 30.0% | 2.84 out of 5 |
FLUX.2 max | 15.5% | 2.49 out of 5 |
Grok Imagine 1.0 | 15.0% | 2.61 out of 5 |
Raw text accuracy was nearly tied at the top, with Ideogram at 4.59, Gemini at 4.55, and Grok at 4.54. Where Ideogram pulled away was typographic craft, scoring 3.92 against 3.50 for Gemini, which is the difference between text that is spelled right and text that looks designed. (ContraLabs)
On Design Arena, a third-party image Elo leaderboard focused on design-oriented generation, Ideogram 4 is the top-ranked open-weight model overall, trailing only proprietary GPT and Gemini models, according to the project's README. On LMArena's text-to-image leaderboard, the picture is more mixed: in the snapshot from November 8, 2026, Ideogram 4.0 (Quality) ranks 22nd overall with a score of 1199, based on 77,744 votes, while the leader, GPT-Image-2.5 Sunburst, sits at 1428. That gap is a useful reality check. Ideogram 4 leads the open-weight field on design-specific measures, but the best closed models still win on general-purpose image generation. (GitHub, LMArena)
How to Run It Yourself
Running Ideogram 4 locally is straightforward, with one gate to clear first. The weights are gated on Hugging Face, so you accept the license on the model page and authenticate with a Hugging Face token before the download works. There are two variants: nf4, which runs on CUDA and supports Diffusers, and fp8, which runs on all hardware but has no Diffusers support yet. The nf4 build fits on a single 24 GB GPU, which puts it within reach of a good consumer card. (Hugging Face)
From there, install the repository with pip, get a free Ideogram API key for the magic prompt, and run the inference script with a plain-text prompt. The script expands your prompt into the structured JSON caption, generates the image, and saves it. For the highest quality, the README recommends 2048 by 2048 output with the V4 quality sampler preset. Prompt and output safety screening runs through Hive, which needs its own text and visual moderation keys. (GitHub)
The release also introduced a layer-based generation stack, starting with transparent layers, plus editable text and movable image layers in a follow-up release. That points at where Ideogram is heading: not just generating flat images, but generating editable design files. (Ideogram)
The License Catch
Here is the part to read carefully before you build a business on it. The GitHub repository is Apache-2.0, and that license covers the inference code. The model weights are not Apache-2.0. Both the nf4 and fp8 weights are listed under the Ideogram 4 Non-Commercial license in the official model zoo, and the press release describes a commercial license with paid tiers, plus enterprise fine-tuning on customer infrastructure. So the correct way to describe Ideogram 4 is: open-weight code under Apache-2.0, weights free for non-commercial use, commercial use through a separate Ideogram license. (GitHub, Ideogram)
That split is increasingly common for frontier open-weight models, and it is not a reason to dismiss the project. It does mean you should check the license page before using Ideogram 4 output in a client project or a product. The model is genuinely open for research, learning, and personal projects, and that is where most readers will use it.
Who Should Try It
Ideogram 4 fits anyone who generates images with text in them: poster designers, social media managers, indie developers making game art with labels, and small teams that need marketing graphics without a designer on staff. It also fits developers who care about layout control, because bounding-box prompting is a genuinely different way to direct an image model, and the JSON interface is easy to generate programmatically.
It is less suited to anyone who needs the absolute best general image quality regardless of openness, since the closed models still lead on LMArena's overall board. And if your use case is commercial, sort out the license first. The honest shape of Ideogram 4 in November 2026 is this: the best open-weight model for design work, with typography that finally holds up in real client work, running on hardware you might already own.
Try this first: accept the license gate on Hugging Face, run the nf4 model with a prompt for a poster, and compare the text against anything you have generated before. Then look at what open source means in practice, and at PhotoCraft, the open source Photoshop alternative rebuilt in Rust, for more open tools in the same creative workflow.
FAQ
Is Ideogram 4 really open source?
Partly. The inference code is Apache-2.0, and the weights are openly downloadable, but the weights carry a non-commercial license. Commercial use needs a separate license from Ideogram.
Can Ideogram 4 run on a home GPU?
Yes, the nf4 build fits on a single 24 GB GPU, according to AMD's write-up of the model. The fp8 build runs on all hardware but has no Diffusers support yet.
Do I have to write JSON prompts by hand?
No. The magic prompt feature expands a plain-text prompt into the structured JSON caption automatically, using Ideogram's free hosted API or your own language model provider.
How good is Ideogram 4 at text, really?
In a blind test by 10 professional designers, it won first place on typography 47.9% of the time, ahead of Gemini at 30.0%, and it scored 0.97 on the X-Omni English OCR benchmark.
Is Ideogram 4 the best image model overall?
No. It is the top-ranked open-weight model on Design Arena, but on LMArena's overall text-to-image board it ranks 22nd, behind the leading closed models. Its edge is design and typography, not general image generation.
