On June 3, 2026, Ideogram released 4.0 — the company’s first open-weight model and, per DesignArena’s leaderboard, the top-ranked open-weight image generation model on that benchmark (1285 Elo, ~114 points ahead of the next open-weight model, confirmed independently). For builders, the headline is not the benchmark ranking. It is the JSON prompting interface: a structured way to describe layouts, palettes, and typography that turns one-off image generation into a templated design pipeline.
At a glance: Ideogram 4 is a 9.3 billion-parameter single-stream Diffusion Transformer trained from scratch. Weights are available in nf4 format and fp8 format on Hugging Face. Inference code is Apache 2.0; weights are under Ideogram’s Non-Commercial Model Agreement — free for research, evaluation, and local experimentation; commercial production deployment requires a separate paid license. ComfyUI added native support the same day. The JSON interface supports per-element text bounding boxes (no fixed cap in the schema), up to sixteen hex color palette slots, and explicit object positioning. Part of our Builder’s Log.
Why This One Is Different
Ideogram has shipped multiple model generations. What changes with 4.0 is the architecture of how you talk to it.
Every prior text-to-image model uses flat prompts: you describe an image in prose, the model makes a best-effort interpretation, and the result is not reproducible at scale. Each call is a negotiation between your description and the model’s learned priors.
Ideogram 4 is trained on structured JSON captions. You can still use flat prose — but the JSON surface gives builders explicit control over six properties that flat prompts leave to chance:
- Bounding-box placement: Any element in the image — subjects, text, background regions — can be assigned a bounding box in normalized coordinates
[y_min, x_min, y_max, x_max]from 0 to 1000. You specify where the object lives in the frame, and the model respects it. - Color palette: Up to 16 hex colors total, up to 5 per element, specified as an array. The model constrains the output palette to those values.
- Typography control: You specify text content, placement, and a bounding box per text element; style notes like weight or font family go in the element’s free-text description rather than a dedicated field. Ideogram’s own JSON prompting docs don’t state a fixed cap on the number of text elements, though hands-on testing suggests keeping dense typographic layouts to a handful of elements for reliable placement.
- Object positioning: Objects and subjects can be positioned relative to the frame, not just described.
- Style attributes: Lighting, rendering style, and texture can be specified declaratively.
- Multi-element layouts: Different elements can carry different constraints simultaneously.
The practical implication: a flat prompt gets you a single image. The JSON interface gets you a template. Run it fifty times with the same layout schema and a different subject line, and you get fifty images with the same compositional rules — the kind of consistency that product, marketing, and brand pipelines require.
Model Specifications
| Property | Value |
|---|---|
| Parameters | 9.3 billion |
| Architecture | Single-stream Diffusion Transformer |
| Training | From scratch (no fine-tune of existing model) |
| Text encoder | Qwen3-VL-8B-Instruct, a vision-language model used in text-only mode |
| Native resolution | 2K (any multiple-of-16 from 256 to 2048 per side) |
| Text rendering | 0.97 accuracy on the X-Omni English OCR benchmark for in-image text, per Ideogram’s own release notes |
| DesignArena rank | #1 open-weight, 1285 Elo — a ~114-point lead over the next open-weight model (HunyuanImage 3.0 at 1171) |
| Inference code license | Apache 2.0 |
| Weights license | Ideogram Non-Commercial Model Agreement |
| Release date | June 3, 2026 |
Ideogram 4 was trained from scratch, not derived from or fine-tuned on any other model — but that applies to the diffusion transformer itself, not the text encoder. The text encoder is Qwen3-VL-8B-Instruct, an existing open vision-language model run in text-only mode, not a proprietary Ideogram component. The from-scratch training still matters for licensing clarity — the diffusion transformer has no inherited IP obligations from FLUX, Stable Diffusion, or any other image-model base.
The License Split and What It Means for Your Project
The weight license is the critical decision gate for any builder evaluating Ideogram 4.
Non-commercial use is free. Research, personal projects, local experimentation, fine-tuning for research purposes, academic work — all of this falls under the Non-Commercial Model Agreement without cost. You can download the weights from Hugging Face today and start running inference.
Commercial use requires a paid license, and it’s tiered. Ideogram’s self-serve commercial license covers self-hosting the public nf4/fp8 quantized weights for 10K–100K images/month at $300/month (billed annually) — but it does not include full-precision weights. Full-precision weights, higher volume (100K+ images/month), customer-facing products, or API-like access for third parties require Ideogram’s Enterprise track, which adds support, onboarding, security review, and custom legal terms.
The Ideogram API is a separate track. If you want production deployment without managing the licensing negotiation and GPU infrastructure, Ideogram operates three per-image API pricing tiers — Turbo ($0.03/image), Default ($0.06/image), and Quality ($0.10/image). This is the faster path for most product builders — you pay per generation rather than per weight access, and the API already carries commercial rights. The tradeoff is API dependency versus model ownership.
For builders evaluating where to sit: if you’re prototyping, exploring the JSON interface, or building a non-commercial product, the free weights are the right starting point. If you’re building a commercial pipeline and need to ship in the next few months, the API path removes friction. If you need on-premises deployment, data isolation, or volume pricing, contact Ideogram directly about commercial weight licensing.
Hosting and Deployment Options
| Option | Who It’s For | Notes |
|---|---|---|
Hugging Face (ideogram-ai/ideogram-4-nf4) |
Research, local experimentation | Quantized (nf4); free non-commercial |
Hugging Face (ideogram-ai/ideogram-4-fp8) |
Higher-fidelity local runs | Quantized (fp8); free non-commercial |
GitHub (ideogram-oss/ideogram4) |
Developers who want the inference code | Apache 2.0 code; full control |
| ComfyUI | Node-graph pipelines, JSON control | Native day-0 support (ComfyUI 0.24.0+); standard image nodes + JSON prompt input |
| fal.ai | Hosted inference, no GPU management | Serverless per-image pricing |
| Ideogram API | Product builders, commercial use | Three per-image pricing tiers; commercial rights included |
ComfyUI’s native support is worth noting: recent ComfyUI releases added native nodes for Ideogram 4.0 on launch day, so you can pass structured JSON prompts directly through standard image nodes and get bounding-box layout and palette control in a visual workflow. For builders already using ComfyUI for image pipelines, the upgrade path is the same workflow with a richer prompt format.
What Ideogram 4 Is Good At
In-image text rendering is the flagship capability. Ideogram’s founders — former Google Brain researchers who worked on Imagen — started the company specifically to solve legible in-image text, for use cases like packaging, posters, signage, UI mockups, and marketing assets with embedded type. Ideogram 4 continues that trajectory, scoring 0.97 on English OCR accuracy for in-image text per Ideogram’s own benchmark. Designers and product teams generating brand assets programmatically are the primary audience.
Templated design pipelines are the new territory that the JSON interface opens. If your application generates personalized images at scale — marketing email headers, product cards, localized ad creatives, greeting cards, report covers — the JSON prompting surface lets you build a layout schema once and vary only the parameters that need to change across outputs. This is the use case that flat-prompt models cannot reliably serve.
Typography-heavy compositions — posters, book covers, announcement graphics, event graphics, social card templates — benefit from explicit text bounding-box control. You stop fighting the model’s default text placement and start directing it.
Brand consistency at scale becomes tractable when you can constrain color palettes to a brand’s hex codes and specify compositional rules per asset type.
Where Ideogram 4 Is Not the Best Choice
General photorealism is not where Ideogram 4 leads. FLUX.2 is specifically optimized for photorealism, and Ideogram’s photorealistic output still trails Midjourney for purely artistic, non-text images. If your use case is photorealistic product photography or portrait generation with no text element, competing models remain stronger.
Unstructured creative generation — the “surprise me” use case — does not benefit from Ideogram 4’s architecture in the same way. The JSON interface adds value when you know what you want; it does not add value when exploration is the goal.
Teams that need managed compliance for commercial deployment right now face more friction than with the API track. The weight licensing negotiation adds a step. For fast-moving commercial builds, the Ideogram API or fal.ai hosted inference removes that step.
Practical Builder Decision Map
You want to prototype or evaluate the model: Download the nf4 or fp8 checkpoint from Hugging Face. Run locally. Use the JSON interface to explore layout control. No cost, no licensing friction for non-commercial work.
You’re building a non-commercial product (open source tool, research project, educational product): Weights are free under the Non-Commercial Model Agreement. Set up via GitHub inference code or ComfyUI.
You’re building a commercial product and need to ship in weeks: Use the Ideogram API. Three tiers available, commercial rights included, no GPU management required.
You’re building a commercial pipeline that requires on-premises deployment or data isolation: Contact Ideogram about commercial weight licensing. Expect a negotiation step before weights can be used in production.
You already have a ComfyUI pipeline for image generation: Add Ideogram 4 as a new model node. Pass structured JSON prompts using the bounding-box and palette parameters. Compare output consistency against your existing model on the specific use cases you care about.
You’re building marketing automation or content generation at scale with brand guidelines to maintain: The JSON palette constraint and bounding-box layout control are specifically designed for this. Ideogram 4 plus a JSON template schema for your brand’s layout rules is closer to a designed output pipeline than to a probabilistic image generator.
Connection to the Broader Image Generation Landscape
Ideogram 4 arrives at a moment when the open-weight image model tier is consolidating. FLUX.2, Qwen’s visual models, and now Ideogram 4 are all competing for the “best open-weight image model” position. On DesignArena’s leaderboard, Ideogram 4 sits roughly 114 Elo points above the next open-weight model (HunyuanImage 3.0) in design-oriented tasks.
The closed-model providers remain ahead on photorealism, but as of this writing none of them publish a comparable structured bounding-box + palette + typography JSON interface — Ideogram 4’s JSON prompting schema is documented specifically as part of this release.
The existing Ideogram review in our Reviews section covers the platform history, the founding team (ex-Google Brain, Imagen), the Ideogram 3.0 model era, and the Canvas editor. Ideogram 4 builds on that review’s context — it is the same company’s text-rendering DNA extended into a fully open-weight, JSON-native architecture.
Summary
Ideogram 4.0 is the first open-weight image generation model that offers a structured programmatic interface for layout control. The JSON prompting surface — bounding boxes, palette constraints, multi-element typography — is what separates it from every other open-weight image model available today.
The licensing split is the practical decision gate: non-commercial use is free; commercial pipelines need either a paid weight license or the API track. For builders who work in design automation, brand asset generation, or any pipeline where compositional consistency across many images matters, Ideogram 4 is worth evaluating now. The nf4 checkpoint on Hugging Face and the ComfyUI native node provide the lowest-friction entry point.