Ideogram 4.0 is the vendor's first open-weight foundation model, a 9.3B-parameter Diffusion Transformer built for flawless in-scene typography, deterministic layout, and native 2K photoreal design output.
Renders headlines, logos, and paragraph copy legibly inside the scene, scoring 0.97 on English OCR accuracy and leading every other open-weight model on text benchmarks.
Go beyond plain prompts with a structured JSON schema: bounding-box coordinates place each element, and hex palette conditioning locks exact brand colors.
Generates up to 2048px per side and ships public fp8 and nf4 checkpoints on Hugging Face and GitHub for local inference.
Choose Turbo for fast drafts, Balanced for everyday work, or Quality for maximum-fidelity final renders.
Ideogram 4.0, released on June 3, 2026, is the Toronto-based lab's most capable text-to-image release and its first open-weight foundation model. At its core is a 9.3 billion-parameter single-stream Diffusion Transformer trained from scratch, where text and image tokens share the same projections rather than running through separate branches. It pairs this with Qwen3-VL-8B-Instruct as its text encoder, concatenating hidden states from 13 intermediate layers so the model reads prompts with the semantic depth of a full vision-language model.
The release targets the parts of image generation that matter most for real design work: typography that stays legible inside the scene, layouts you can control precisely, and clean 2K photoreal output. A structured JSON prompting format lets you define bounding boxes for element placement, assign hex color palettes, and separate the text string from its visual styling for each element in the frame. Ideogram ships public fp8 and nf4 checkpoints for local inference, while the hosted API and partner endpoints cover production-scale and commercial use.
Ideogram 4.0's headline strength is text rendering. It reliably places legible words, headlines, and even paragraph-length copy inside a generated image, scoring 0.97 on English OCR accuracy and topping open-weight text-rendering benchmarks ahead of much larger models.
Instead of relying on a single natural-language string, you can pass a JSON schema that defines regions, assigns content to them, and controls styling per element. This gives layout-level determinism that plain-text prompts on most models cannot match.
Specify bounding-box coordinates to pin logos, headings, and product elements exactly where you want them, turning composition from a guessing game into a spec you define up front.
Feed exact hex colors into the prompt to constrain the output palette, keeping generated designs on-brand without color drift across a batch.
The model generates natively up to 2048px per side across a range of aspect ratios, producing print- and screen-ready design assets without a separate upscaling pass.
As Ideogram's first open-weight model, 4.0 publishes its weights and inference code on Hugging Face and GitHub, with quantized fp8 and nf4 checkpoints that make local and self-hosted generation practical.
Turbo trades steps for speed on quick iterations, Balanced covers everyday production, and Quality maximizes fidelity for final renders, letting you dial cost and detail per job.
| Parameters | 9.3 billion |
|---|---|
| Architecture | Single-stream Diffusion Transformer (trained from scratch) |
| Text encoder | Qwen3-VL-8B-Instruct (13 intermediate layers concatenated) |
| Modality | Text-to-image |
| Max resolution | Native 2K (up to 2048px per side) |
|---|---|
| Tiers | Turbo, Balanced, Quality |
| Formats | PNG, WEBP (with transparency support) |
| Aspect ratios | Multiple supported |
| Prompt format | Natural language + structured JSON schema |
|---|---|
| Layout control | Bounding-box coordinates per element |
| Color control | Hex palette conditioning |
| Per-element control | Separate text string and visual styling |
| Weights | Open (Hugging Face + GitHub) |
|---|---|
| Checkpoints | fp8 and nf4 quantized variants |
| License | Ideogram 4 Non-Commercial Model Agreement (weights); commercial via API/subscription |
| Released | June 3, 2026 |
Generate finished posters, ads, and social graphics with headlines and body copy rendered legibly in-scene, no manual text overlay required.
Explore wordmarks, badges, and typographic treatments where accurate letterforms and spelling are non-negotiable.
Use hex palette conditioning and bounding-box layout to produce large sets of visuals that stay consistent with a brand's colors and composition rules.
Place labels, product names, and callouts precisely with JSON layout control to prototype packaging and merchandise designs.
Run the open fp8 or nf4 checkpoints on your own hardware for private, offline, or high-volume design pipelines.
The published weights are released under a non-commercial agreement. Commercial use runs through Ideogram's subscription, developer API, or hosted partner endpoints rather than the open checkpoints.
Ideogram 4.0 is a text-to-image foundation model. It does not generate video or audio, and image-editing or inpainting capabilities are separate from base generation.
The structured JSON schema that unlocks precise layout and color control is more involved than plain-text prompting and takes experimentation to use well.
Figures such as the 0.97 OCR score, layout mIoU, and DesignArena rankings are largely reported by Ideogram; independent results may vary by prompt, tier, and settings.
Design-ready images with text that actually reads are on the way. Ideogram 4.0 support is coming soon to SharkFoto. Explore our available AI image tools and effects today, and be ready to generate poster-perfect typography the moment it lands.
Try Ideogram 4.0 now