Ideogram 4: Specs, License, Benchmarks, and Typography Control (2026)
Last updated August 7, 2026
Ideogram 4 (Jun 3, 2026) is the top open-weight design model: a 9.3B single-stream DiT with a Qwen3-VL-8B encoder, 256–2048px any-size output, native transparency, and structured JSON prompting — bounding boxes on a 0–1000 grid plus 16-color hex palettes. Inference code is Apache 2.0 but weights are gated non-commercial. Runs inside invideo as Ideogram 4, 4 Remix, 3, and Remove Object.
Updated August 2026
Ideogram 4 is the strongest open-weight image model for typography and graphic design as of August 2026: a 9.3-billion-parameter diffusion transformer released June 3, 2026 — the first open-weight model in Ideogram's history — that renders legible multi-line, multi-font text and ranks as the top open-weight model on DesignArena (Ideogram announcement). Its signature capability is structured JSON prompting: every element can be pinned to a bounding box on a 0–1000 grid with the palette locked to up to 16 hex colors — deterministic layout control no other major family offers. In invideo's model picker the family appears as Ideogram 4, Ideogram 4 Remix, Ideogram 3, and Ideogram Remove Object. One caveat belongs up front: "open-weight" here does not mean commercially usable — the license is the most misreported fact about this release.
Who makes Ideogram, and how did it get to 4.0?
Ideogram was founded in 2022 in Toronto by Mohammad Norouzi, William Chan, Chitwan Saharia, and Jonathan Ho — the ex-Google Brain researchers behind the Imagen text-to-image model; Ho co-authored the DDPM paper that underpins modern diffusion models (Wikipedia). Legible in-image text has been the family's identity from day one:
| Version | Date | What it added |
|---|---|---|
| 0.1 | Aug 2023 | Launch with a16z/Index seed backing; text rendering as the core pitch |
| 2.0 | Aug 2024 | Style system: realistic, design, 3D, anime |
| 3.0 | Mar 2025 | Better realism, complex text layout, up to 3 style reference images |
| 4.0 | Jun 3, 2026 | First open-weight release; new architecture trained from scratch |
The through-line is the day-one pitch — text you can actually read — with 4.0 converting it into an open-weight, layout-controllable system.
What's inside Ideogram 4?
| Spec | Ideogram 4.0 |
|---|---|
| Architecture | 9.3B-parameter single-stream Diffusion Transformer (DiT), 34 layers; text and image tokens in one unified sequence; trained from scratch |
| Text encoder | Qwen3-VL-8B-Instruct, with hidden states drawn from 13 intermediate layers |
| Resolution | Any size from 256 to 2048 px (multiples of 16); aspect ratios up to 6:1; native 2K |
| Transparency | Native alpha/transparent output layers |
| Layout control | Structured JSON prompts: bounding boxes on a 0–1000 grid, per-element text strings, up to 16 hex palette colors |
| Speed presets | Turbo (12 steps), Default (20), Quality (48) |
| Self-hosting | nf4 quantization runs on a single 24GB GPU; fp8 build for non-CUDA hardware |
| Weights | Open but gated non-commercial; inference code Apache 2.0 |
Sources: the Hugging Face model card and GitHub repository, as of August 2026. The row that decides most deployments is the last one — open code, gated weights — so start there.
Is Ideogram 4 open source?
Partly, and the distinction is widely misreported. The inference code is Apache 2.0 — genuinely open source. The weights are gated under a separate non-commercial license on Hugging Face (model card): you can download, self-host, research, and experiment, but not use self-hosted output commercially. There is no revenue threshold — a $200 freelance logo needs a commercial path just like an enterprise deployment, via Ideogram's paid subscriptions or the API, both of which include commercial rights. Articles calling Ideogram 4 "Apache 2.0 open source" without qualification are describing the code, not the weights.
What is structured JSON prompting?
Ideogram 4 was trained on structured JSON captions, so instead of describing a poster and hoping, you can specify it: each element gets a typed text string and a bounding box on a normalized 0–1000 grid, and the whole image can be conditioned on a palette of up to 16 hex codes (Hugging Face model card). Overlapping boxes respect z-order, so a logo can sit over a photo region without masks. For programmatic brand assets — catalogs, localized ad sets, templated thumbnails — this is the difference between art direction and dice rolls, and as of August 2026 no competing family documents an equivalent.
Which Ideogram variant does what?
The family is one model line with several task-specific endpoints. In invideo's Image tab the roster is:
| Picker entry | What it does | Notes |
|---|---|---|
| Ideogram 4 | Text-to-image on the 4.0 model | The default for typography and layout work |
| Ideogram 4 Remix | Image + prompt reinterpretation | An image_weight dial trades structure preservation against prompt influence; accepts JPEG/WebP/PNG up to 10MB; 40+ preset resolutions up to 3328×1248 (API reference) |
| Ideogram 3 | The March 2025 model | Still strong for styled text work; three style reference images |
| Ideogram Remove Object | Object removal | Mask-based, not prompt-based — see below |
| P-Image-Ideogram | Speed-optimized build by Pruna | Faster inference for volume drafting |
The Remove Object detail trips people up: the endpoint takes a mask image (white pixels mark the removal region), not a text description of what to remove (API reference); prompt-driven removal goes through Ideogram's separate edit-with-prompt endpoint. The broader v4 API also includes background removal, reframing, upscaling, and a quietly important layerize endpoint that decomposes a flat graphic into editable text layers (developer.ideogram.ai).
How good is Ideogram 4 on benchmarks?
Three results carry the weight, all as of the June 2026 launch:
- Top open-weight model on DesignArena, ranking 9th overall in text-to-image and 1st in quality mode among open models — behind only closed frontier models (Ideogram announcement).
- 47.9% first-place win rate in a blind typography evaluation with professional designers (ContraLabs eval, reported on the Hugging Face model card) — designers picked Ideogram 4's typography over every competitor's nearly half the time.
- At 9.3B parameters it beats far larger open models — including 32B and 20B competitors — on typography, per independent testing in AI Video Bootcamp's guide (June 2026).
One caution: exact Elo numbers for Ideogram 4 differ across trackers; the rankings — top open-weight, below the closed frontier — are consistent everywhere.
Where Ideogram 4 struggles
Per community testing (AI Video Bootcamp, June 2026) and the model's own documentation:
- Diacritics — Polish, Turkish, and Vietnamese marks get dropped or duplicated.
- Non-Latin scripts — Arabic, Cyrillic, and CJK text remain unreliable despite the multilingual encoder.
- Hands in photorealistic scenes — the classic failure persists.
- Body copy — quality degrades past roughly 25–30 words; headlines are the sweet spot, paragraphs are not.
- Photorealism — dedicated photoreal models still win on naturalism; Ideogram's edge is design, not photography.
How to prompt Ideogram 4
- Double-quote literal text.
a poster that says "SUMMER SALE"renders the exact string; unquoted text gets paraphrased. - Name real typefaces. "Cooper Black" or "Futura" beats "bold retro font."
- Put hex codes in the palette field, not the prose prompt; use plain color words in the description.
- Use bounding boxes for anything positional — specify the box instead of writing "at the top."
- Keep rendered text under ~25 words per image.
What should you make with Ideogram 4?
The typography edge maps to specific jobs. Logos and wordmarks with consistent kerning are the flagship case — why a logo generator workflow leans on Ideogram over photoreal models. Event posters exploit the layout grid directly, which is what an AI poster maker needs when date, venue, and headline all have to land legibly. Bold, readable overlay text is where an AI thumbnail maker gets the most from the 47.9% typography win rate. And palette-locked layouts across dozens of sizes make display ads in a banner maker consistent instead of approximately-on-brand.
Ideogram 4, asked and answered
What is Ideogram 4?
A 9.3B-parameter open-weight diffusion transformer released June 3, 2026, specialized in in-image typography and layout control — the first Ideogram model with downloadable weights and the top open-weight model on DesignArena at launch.
Can I use Ideogram 4 commercially if I self-host it?
No. The self-hosted weights carry a non-commercial license with no revenue threshold. Commercial use requires Ideogram's paid subscription tiers or the API, which include commercial rights (Hugging Face model gate, August 2026).
How much does Ideogram cost?
API generation runs roughly $0.03–$0.10 per image by rendering speed (Turbo/Default/Quality); the v3 endpoints are confirmed at $0.03/$0.06/$0.09. Subscription figures conflict across sources — entry tiers reported at $7–9/month, pro tiers $42–45 — so verify on ideogram.ai's pricing page before budgeting. All paid tiers include commercial rights.
What GPU do you need to run Ideogram 4 locally?
The nf4 quantization runs on a single 24GB GPU (e.g., an RTX 4090); an fp8 build covers non-CUDA hardware (GitHub). Expect roughly a minute per 2K image on consumer hardware.
Is Ideogram Remove Object prompt-based?
No — it takes a mask image, with white pixels marking the region to remove (API reference). Prompt-driven editing goes through the separate edit-with-prompt endpoint.
How is Ideogram 4 different from Ideogram 3?
New 9.3B single-stream architecture trained from scratch, open weights, native transparency, structured JSON prompting, and any-size output from 256 to 2048 px. Ideogram 3 remains a capable closed-API model with style references but none of the layout-control surface.
Where does Ideogram fit in an invideo workflow?
All four production variants — Ideogram 4, Ideogram 4 Remix, Ideogram 3, and Ideogram Remove Object — run inside invideo's agent, listed in the Image tab of the model picker. That placement matters for design work: the agent can route a typography-heavy job (a titled thumbnail, a poster with a locked palette) to Ideogram 4 while sending photoreal or volume jobs to models that win those. The family's dedicated hub with examples and workflow notes is at invideo.io/ai-models/ideogram.
Version history: Ideogram 0.1 launch Aug 2023 → 2.0 style system Aug 2024 → 3.0 realism + complex text Mar 2025 → 4.0 first open-weight release, structured JSON prompting, Jun 3, 2026.