Ideogram 4.0: Open-Weight Image Generation — How to Run, API & JSON Prompts (2026)
Ideogram 4.0 launched June 3, 2026 with open weights, native 2K output, and JSON-first prompting. Step-by-step for the API, local inference on a 24GB GPU, magic-prompt, bounding-box layout, and how it compares to FLUX and GPT Image.
The headline is not just "another open diffusion model." Ideogram 4.0 closes the quality gap between proprietary frontier image models and the open ecosystem on the axes that matter for production design work: typography in scene, deterministic layout, and 2K photoreal output. CEO Mohammad Norouzi put it directly: "The hardest problems at the forefront of design generation — headline-grade typography, deterministic layout, branded layered output — need a foundation engineered for them."
Update — July 10, 2026:Reve 2.1 targets the same layout-and-typography lane with 4K output and stronger foreign-text rendering — closed hosted weights vs Ideogram's open JSON/bounding-box stack.
This guide covers what shipped, how the architecture differs from unified multimodal stacks, and how to run Ideogram 4.0 — via API, CLI, and self-hosted inference.
Ideogram has led on in-scene typography since its 2023 launch. Version 4.0 extends that with multilingual support, denser type at smaller scales, and reliable rendering of headlines, packaging copy, and signage. In a ContraLabs blind evaluation judged by ten professional designers, Ideogram 4.0 was picked as best 47.9% of the time — ahead of Gemini 3.1 Flash Image Preview (30.0%), FLUX.2 [max] (15.5%), and Grok Imagine 1.0 (15.0%).
2. Bounding-box layout control
You specify where a logo, headline, callout, or subject belongs on the canvas using normalized [y_min, x_min, y_max, x_max] coordinates on a 0–1000 grid. Layout is directed by the brief, not sampled and corrected afterward.
3. Photoreal output at 2K
Native support for resolutions from 256 to 2048 (multiples of 16), with aspect ratios up to 6:1. For highest quality locally, the README recommends --height 2048 --width 2048 --sampler-preset V4_QUALITY_48.
Layer-based roadmap
Most professional design work is not a single pixel layer. Ideogram 4.0 is the start of a generation stack:
Capability
Status
Transparent background cutouts
Available via Background Remover API
Editable text + movable image layers
Follow-up 4.0 release
Branded assets (typography, palette, logo fidelity)
Scheduled
Architecture: a specialized foundation, not a unified multimodal model
Ideogram 4.0 is a foundation model trained entirely from scratch — not a fine-tune or distillation of any existing checkpoint. Key architectural choices from the GitHub README:
Component
Detail
Backbone
34-layer single-stream Diffusion Transformer (DiT) — text and image tokens in one unified sequence
The bet is explicit: unified multimodal models (GPT Image, Gemini) are strong generalists, but headline-grade typography, deterministic layout, and brand fidelity require a foundation engineered for design specifically. At 9.3B parameters, Ideogram 4.0 delivers the best text rendering of any open-weight release Ideogram benchmarked — ahead of Qwen-Image (20B), FLUX.2 [dev] (32B), and HunyuanImage 3.0 (80B MoE).
For a general primer on how diffusion image models work under the hood, see our diffusion explainer.
We train exclusively on JSON so that training and inference share a single, common prompt format. The training captions themselves are deliberately extremely descriptive: each JSON exhaustively describes everything in the image.
Plain-text prompts create train/eval mismatch. JSON mirrors the training distribution and unlocks full model quality.
The caption schema (three top-level fields)
json
{"high_level_description":"A clean business card layout for a tech startup.","style_description":{"aesthetics":"minimal, professional, geometric","lighting":"even, diffuse studio lighting","medium":"graphic_design","art_style":"flat vector design, generous whitespace, sans-serif typography","color_palette":["#FFFFFF","#F0F0F0","#333333","#0066FF","#00CC88"]},"compositional_deconstruction":{"background":"A solid off-white card surface with subtle paper texture.","elements":[{"type":"text","text":"ACME TECH","desc":"Bold dark grey sans-serif company name across the upper third."},{"type":"text","text":"[email protected]","desc":"Small blue sans-serif contact email near the bottom."}]}}
Field
Required
Purpose
high_level_description
Strongly recommended
One- or two-sentence summary
style_description
Optional
Aesthetics, lighting, medium, color palette
compositional_deconstruction
Required
Background + spatial elements
Element types: "obj" for objects/subjects, "text" for in-image text (include a text field with the literal string to render).
Magic-prompt: JSON without writing JSON
Don't want to hand-write captions? Magic-prompt expands plain text into full structured JSON before generation.
Three backends ship in the repo:
Config
Registry key
Backend
Ideogram4MagicPromptV1
ideogram-4-v1
Ideogram hosted API (free)
ClaudeOpusMagicPromptV1
claude-opus-v1
OpenRouter
ClaudeSonnetMagicPromptV1
claude-sonnet-v1
OpenRouter
The hosted ideogram-4-v1 backend is the default in run_inference.py and only needs IDEOGRAM_API_KEY. The magic-prompt system prompts are open source in src/ideogram4/magic_prompt_system_prompts/.
Via the API, two endpoints scaffold the JSON workflow:
Practical workflow: Start with text_prompt for fast ideation. Migrate to json_prompt once layout precision, brand hex colors, or multi-line typography matter.
Bounding-box layout and color palettes
Spatial control with bbox
Each element can include a bounding box in normalized 0–1000 coordinates (origin top-left):
json
{"type":"text","bbox":[100,200,300,800],"text":"SUMMER SALE","desc":"Large bold red headline across the upper center of the poster."}
Format: [y_min, x_min, y_max, x_max]. This is native to the model — no ControlNet pipeline required.
Color palette conditioning
Steer dominant colors with hex codes in style_description.color_palette:
Up to 16 colors in style_description.color_palette
Up to 5 colors per element
Uppercase hex only — #RRGGBB form (not #fff or lowercase)
Include both highlight and shadow colors for controlled lighting
On 7Bench (layout control), Ideogram 4.0 scored significantly better than all closed-source models tested — the bbox + palette system is the differentiator.
API endpoints beyond generate
The Ideogram API is not just text-to-image. Full capability list from ideogram.ai/api-learn:
Capability
Endpoint family
Notes
Generate
/v1/ideogram-v4/generate
Text or JSON prompt → image
Transparent backgrounds
v4 endpoints
Native alpha cutouts
Edit with prompt
v3 endpoints
Describe changes in plain language
Remix
v3 endpoints
Reimagine with image_weight control
Reframe
v3 endpoints
Extend to new aspect ratio
Remove background
v4 endpoints
Clean cutout in one call
Layerized text
v3 endpoints
Pull editable text layers
Custom models
Training + generate
Fine-tune on brand assets
Upscale
Upscale endpoint
Raise resolution for delivery
Magic-prompt
/v1/ideogram-v4/magic-prompt
Plain text → JSON caption
Describe
/v1/ideogram-v4/describe
Image → JSON caption
Ideogram 4.0 also supports MCP for agent workflows — useful if you're wiring image generation into coding agents or design automation pipelines. For agent harness concepts, see our Agent Harness guide.
Requires ComfyUI 0.24.0+ and image_ideogram4_t2i.json template
For most developers building image generation into a product, start with the API (Turbo at $0.03/image for prototyping). Move to local inference when you need custom fine-tunes, synthetic data pipelines, or on-premise deployment.
The NF4/FP8 Hugging Face checkpoints use a non-commercial license for the open release. Commercial use through the API or enterprise licensing is the production path.
Summary
Ideogram 4.0 is the most significant open-weight image release of 2026 for anyone who ships visual assets — not hobbyists generating cats, but teams that need readable type, controlled layout, and 2K fidelity.
Three things to remember:
JSON is the native prompt format. Use magic-prompt for casual input; write JSON when layout and typography matter.
Three ways in: API for products, CLI for research/self-hosting, app for hands-on design.
It closes the open-vs-closed gap on design benchmarks while staying at 9.3B parameters — a fraction of FLUX.2 [dev]'s 32B.
Model specs, API pricing, and benchmark numbers in this post reflect publicly available information as of June 20, 2026. Verify current pricing and license terms at ideogram.ai before production deployment.
