Ideogram 4 Release: Open-Weight 9.3B Text-to-Image Model

Ideogram 4 introduces structured JSON prompting for high-precision design

Ideogram 4 is a state-of-the-art 9.3B parameter text-to-image foundation model trained from scratch. It distinguishes itself from other open-weight models by utilizing a structured JSON prompting interface, which allows for explicit control over bounding-box layouts, color palettes, and multilingual text rendering. Unlike many contemporary releases, Ideogram 4 is not a fine-tune of an existing model but a ground-up build designed specifically for the frontier of graphic design.

Model Architecture and Technical Specifications

Ideogram 4 is built on a fully single-stream Diffusion Transformer (DiT) architecture, enabling deep cross-modal interaction by processing text and image tokens in a unified sequence through a 34-layer transformer.

Core Architectural Components

  • Single-Stream DiT: Text and image tokens are concatenated into one sequence, eliminating separate branches for different modalities.
  • Vision-Language Text Encoder: The model uses Qwen3-VL-8B-Instruct as its text encoder. By extracting hidden states from 13 intermediate layers, the model gains multi-scale semantic features ranging from basic token information to complex compositional understanding.
  • Dual-Branch Classifier-Free Guidance: This allows independent refinement of conditional (positive) and unconditional (negative) branches to separately control image quality and prompt adherence.
  • Flexible Resolution Support: Ideogram 4 natively supports resolutions from 256 to 2048 (in multiples of 16) with aspect ratios up to 6:1, with an auto-adjusting noise schedule based on the resolution.

Performance and Benchmarks

Ideogram 4 is positioned as a leading open-weight model for design-oriented tasks, outperforming several proprietary and open models in typography and layout control.

Design and Typography Leadership

  • Design Arena: Ideogram 4 is the top-ranked open-weight model on the Design Arena Elo leaderboard, trailing only proprietary models from GPT and Gemini.
  • ContraLabs Typography Evaluation: In a blind test by professional designers, Ideogram 4 achieved a 47.9% first-place win rate, significantly beating Gemini 3.1 Flash Image Preview (30.0%), FLUX.2 [max] (15.5%), and Grok Imagine 1.0 (15.0%).
  • Practical Usability: Professional designers rated Ideogram 4's suitability for real client work at 3.55/5, the highest among the compared models.

General Capability Benchmarks

  • LMArena: Ideogram 4 is the top-ranked open-weight lab and a top-5 image generation lab overall.
  • Open-Source Benchmarks: The model demonstrates superior layout control on the 7Bench benchmark compared to all closed-source models. It also leads other open-weight models in text rendering efficiency, outperforming larger models such as Qwen-Image (20B), FLUX.2 [dev] (32B), and HunyuanImage 3.0 (80B MoE) despite its smaller 9.3B parameter size.

Prompting Strategy: The JSON Interface

Ideogram 4 was trained exclusively on structured JSON captions to maximize training efficiency and grounded supervision. While it supports plain-text prompts, the JSON format is the native language of the model and is required for maximum controllability.

Advanced Control Mechanisms

  • Bounding-Box Layout: Users can specify bbox coordinates in the JSON prompt to explicitly place subjects, text, and background elements.
  • Color Palette Conditioning: The colour_palette array allows users to steer the dominant color scheme using hex codes.
  • Compositional Deconstruction: Precise spatial layout is achieved through per-element descriptions combined with bounding boxes.

Magic Prompt

To bridge the gap between casual users and structured JSON, Ideogram provides a "magic prompt" LLM. This tool expands plain-text prompts into the full structured JSON captions the model expects, allowing users to achieve high-quality results without writing JSON manually.

Deployment and Access

Model Zoo and Licensing

Ideogram 4 is available in two quantizations:

  • nf4 (9.3B): CUDA supported, Diffusers compatible.
  • fp8 (9.3B): Supported across all hardware.

Both versions are released under the Ideogram 4 Non-Commercial License.

Quick Start Implementation

Users must accept the license gate on Hugging Face before downloading weights. Inference can be run via the CLI using run_inference.py. For high-quality output, the recommended settings are --height 2048 --width 2048 and --sampler-preset V4_QUALITY_48.

Community Insights and Counterpoints

While the technical capabilities are praised, the community has raised points regarding licensing and safety:

"Non-commercial license, you should not call that 'open-weight'. Words have meaning. And people are having a laugh at how censored the model is."

Additionally, some users have noted that third-party API routers may not emphasize the importance of the JSON-structured prompt format, which could lead to suboptimal results if users rely solely on natural language prompts.

Sources