Ideogram Drops 9.3B Open-Weight Model That Rivals Closed Frontier Image AI
Ideogram's first open-weight model is a 9.3B Diffusion Transformer with native 2K output and JSON-structured prompts for layout control.
Takeaways
- Ideogram 4.0 launches as a 9.3B open-weight Diffusion Transformer with weights on Hugging Face
- Native 2048x2048 output with no separate upscaler, supporting any aspect ratio up to 6:1
- Trained on structured JSON captions with bounding boxes, hex color palettes, and per-element text styling
- Uses frozen Qwen3-VL-8B-Instruct text encoder and asymmetric dual-branch classifier-free guidance
- API runs $0.03 to $0.09 per image, web plans start at $8/month, available in Krea and Picsart Flow
- Designers rated it 3.55/5 for real client work, ahead of Nano Banana 2, Grok Imagine, and FLUX.2 max
Ideogram just dropped its first open-weight image model, and it is aiming straight at the production design workflow. Ideogram 4.0 is a 9.3 billion parameter Diffusion Transformer trained from scratch, with native 2K output, a structured JSON prompting interface, and weights you can download and run locally. It is already live in Krea, Picsart Flow, and the Ideogram app, with API access alongside.
The headline numbers are aggressive. On LMArena, Ideogram is the top-ranked open-weight lab and a top-5 image generation lab overall, and on an internal blind benchmark of graphic designers Bradley-Terry scores rank Ideogram 4 second overall behind only GPT Image 2 medium. The team also claims it outperforms open-weight models at 20B, 32B, and 80B MoE scale, which is the kind of efficiency claim that is worth testing yourself.
JSON in, pixels out
The most distinctive thing about 4.0 is how you prompt it. The model was trained exclusively on structured JSON captions rather than plain text prompts, with every training image described by per-element styling and optional bounding boxes and color specifications. Plain prose still works, but the model was built to consume something more like a layout spec.
What that schema actually exposes:
- Bounding boxes in
[y_min, x_min, y_max, x_max]normalised to 0 to 1000 coordinates with the origin at the top left - Colour palettes of up to 16 hex colours per image, with 5 per element
- Typed text elements that carry both the literal string and a styling description for multi-font posters
- Composable
objandtextentries under acompositional_deconstructionfield
The samples in the technical report use six text bounding boxes to anchor a poster's title, subtitle, date, tagline, credit and RSVP, twenty-eight boxes to mark inventory items, and four boxes to place a woman, window, monstera and cat in one scene. The intent is clear: stop rolling the dice on layout and just declare it.
The reference pipeline validates every input and rejects non-conforming JSON, while plain strings are expanded via an optional magic prompt, either through the hosted API or a local LLM, into the structured format. That sidesteps the usual problem with structured-input models, which is that nobody wants to hand-write JSON for casual generations.
Inside the 9.3B transformer
Ideogram 4.0 is a 9.3B parameter open-weight text-to-image model that follows the pattern of recent open-weight releases: text and image tokens share the same projections at every layer of a 34-layer DiT, with two design choices that distinguish it from peer releases. The encoder is interesting. The model uses Qwen3-VL-8B-Instruct as its text encoder, extracting features from 13 intermediate layers rather than just the final output, and implements dual-branch classifier-free guidance.
The second twist is how guidance is computed at sampling time. Classifier-free guidance combines a conditional pass over text plus image latents with an unconditional pass, and Ideogram 4.0 makes the unconditional pass asymmetric: it drops the text tokens entirely instead of replacing them with padding, so the unconditional pass runs only over image tokens, and the two branches can be tuned independently to schedule prompt adherence and image quality separately across the sampling trajectory.
That gets used in the sampler presets shipped with the repo. The V4_QUALITY_48 preset runs 45 steps at gw=7 followed by 3 polish steps at gw=3 near t=0, while V4_DEFAULT_20 and V4_TURBO_12 follow the same shape with two and one polish steps, with the polish tail tightening fine detail without over-saturating the global composition.
Running it locally
The weights are real and downloadable. They are gated on Hugging Face as ideogram-ai/ideogram-4-nf4 and ideogram-4-fp8 behind the Ideogram 4 Non-Commercial Model Agreement, the inference code is Apache 2.0 on github.com/ideogram-oss/ideogram4, nf4 fits a single 24 GB GPU, and fp8 is offered for broader hardware. A minimal local run looks like this:
python run_inference.py \
--prompt "campaign poster with clean type" \
--output out.png \
--quantization nf4 \
--magic-prompt-key "$IDEOGRAM_API_KEY"
NF4 quantisation makes this meaningfully more accessible than full bfloat16 for a 9.3B model, but 2K inference on a 9.3B DiT still needs a proper GPU, and VRAM requirements at 2048px specifically are not clearly documented yet.
Resolution is not a separate step anymore
The other big practical claim is that 2K is the default. Ideogram 4.0 generates at native 2K resolution without a separate upscaling step, whereas most open-weight image models generate at lower resolutions and depend on external upscaling pipelines to produce print-ready output. A single model handles any resolution from 256 to 2048 in multiples of 16, with aspect ratios up to 6:1, and the noise schedule auto-adjusts per resolution. So square thumbnails, phone wallpapers, social headers and ultrawide banners all come out of the same weights.
Pricing and where to use it
Costs are straightforward. API pricing runs from $0.03 per image at Turbo quality to $0.09 per image at the Quality tier, and the web platform starts at $8 per month on the annual plan. Krea rolled it into its image tab, Picsart Flow added it as a model option, and the Ideogram web app is still the simplest place to try the JSON interface without writing any code.
What it is actually good at
The benchmark story is more interesting than the usual win-rate parade. Professional designers rated Ideogram 4.0 at 3.55 out of 5 on real client-work usability, ahead of Nano Banana 2 at 2.84, Grok Imagine 1.0 at 2.61, and FLUX.2 max at 2.49. The strongest areas:
- Frontier multilingual text rendering, including dense small text and natural handwritten text
- Bounding-box layout control for directing where objects and text land
- Transparent backgrounds ready to drop straight into a design
- Photorealistic images with accurate lighting, texture depth and surface detail at native 2K, holding up for product photography, fashion editorial and food imagery without post-production correction
Where to be careful: typography benchmark prompts often have particular characteristics, specific scripts and font contexts, that may not match the full diversity of production use. And on negative prompting, instead of using a negative prompt to exclude unwanted elements, you describe the negative space you want as positive instruction; saying "clean white background, no other objects in frame" is more reliable than listing what you do not want, because the model was trained on descriptive captions, not exclusion lists.
Why this matters
Two things stand out. First, this is the first time an open-weight model has credibly competed with closed frontier image models on the design-grade tasks that text-to-image has historically been worst at: legible typography, multi-element layout, and 2K output without an upscaler bolted on. Picsart's writeup positions it in the same performance band as Google DeepMind's Gemini 3.0 Pro Image Gen and Gemini 3.1 Flash Image Gen, while being fully open weight.
Second, the JSON-native training data is a real architectural bet. If models keep being trained on structured captions instead of prose, prompting becomes a layout language rather than a creative writing exercise, and a strong result becomes reproducible rather than a lucky seed. Structured JSON prompts give explicit control over description, composition, and individual elements, so a strong result is reproducible rather than a lucky roll. For anyone building agent pipelines that generate marketing assets, packaging mocks, or any output where placement matters, that reproducibility is the actual unlock.