# Ideogram 4.0 Technical Details: Open model at the forefront of design

Ideogram publishes a public search API over its indexed pages; the guide is at https://ideogram.ai/agents/agent-access.

Source: https://ideogram.ai/blog/ideogram-4.0/
Source observed: 2026-11-03T18:37:02.271Z
Last checked: 2026-11-08T15:45:24.084+00:00
Indexed: 2026-11-03T18:42:09.635+00:00

---

# Ideogram 4.0 Technical Details: Open model at the forefront of design

**Blog — June 3, 2026**

We introduce Ideogram 4.0, our first open-weight foundation model: a 9.3B-parameter, single-stream Diffusion Transformer trained from scratch with a vision-language text encoder and structured JSON prompts.

- **Weights:** [Hugging Face](https://huggingface.co/collections/ideogram-ai/ideogram-4)
- **Code:** [GitHub](https://github.com/ideogram-oss/ideogram-4)
- **API:** [Ideogram API](https://developer.ideogram.ai/)

## Overview

Ideogram 4.0 is an open-weight text-to-image model. It uses one self-attention sequence containing text and image tokens, with shared projections through all 34 DiT layers. Its two distinguishing design choices are:

1. We use [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) as a vision-language text encoder in text-only mode. The DiT consumes hidden states from 13 intermediate encoder layers concatenated along the feature dimension.
2. We train exclusively on structured JSON captions describing image elements, styling, and optionally bounding boxes and color palettes. The reference inference pipeline parses and validates every prompt against the JSON schema before generation.

## Architecture at a glance

Our pipeline has four components:

1. A frozen vision-language text encoder.
2. A trainable single-stream Diffusion Transformer.
3. A flow-matching Euler sampler with asymmetric classifier-free guidance.
4. A frozen KL autoencoder that decodes latents to pixels.

![Ideogram 4.0 inference pipeline](https://ideogram.ai/blog/ideogram-4.0/)

*Figure 1. The 9.3B-parameter DiT is the only trained component; the text encoder and VAE are reused frozen.*

### Model spec

| Field | Value |
|---|---|
| Parameters | 9.3B |
| Embedding dimension | 4608 |
| Transformer layers | 34 |
| Attention heads | 18 |
| Intermediate dimension | 12 288 (SwiGLU) |
| AdaLN dimension | 512 |
| RoPE theta | 5 000 000 |
| MRoPE section | (24, 20, 20) |
| Latent channels | 32 × 2² = 128 |
| Maximum text tokens | 2048 |
| Resolution range | 256–2048 px per side |
| Aspect ratios | Flexible |
| Quantization | nf4 and fp8 |
| Sampler | Euler flow-matching, asymmetric CFG |

## Model

The text encoder is [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) in text-only mode. Our DiT consumes hidden states from 13 intermediate encoder layers rather than one hidden-state slice.

The DiT has 34 transformer blocks operating on one sequence of text and image latent tokens. Every block uses self-attention with QK-RMSNorm, 3D Multimodal RoPE that places text and image tokens in one positional frame, and a SwiGLU MLP. AdaLN derived from the flow-matching timestep modulates the attention and MLP residual paths in every block.

Our training objective is flow matching: the DiT predicts a velocity field `v(z_t, t)` defining an ODE from pure noise to a clean latent. During inference, an Euler sampler integrates the ODE over a logit-normal noise schedule whose mean automatically adjusts with resolution.

### Asymmetric classifier-free guidance

Classifier-free guidance combines a conditional pass containing text and image latents with an unconditional pass. In Ideogram 4.0, the unconditional pass is asymmetric: we remove text tokens entirely instead of replacing them with padding, so it runs only over image tokens. The two branches can be tuned independently, allowing prompt adherence and image quality to be scheduled separately during sampling.

The `V4_QUALITY_48` preset runs 45 steps at `gw=7`, followed by 3 polish steps at `gw=3` near `t=0`. `V4_DEFAULT_20` and `V4_TURBO_12` use the same pattern with two and one polish steps respectively. The polish tail tightens fine detail without over-saturating the global composition.

## Structured JSON prompting

We train the model exclusively on structured JSON captions. Each caption exhaustively describes image elements, includes a style block, and may include bounding boxes and color palettes. Training and inference use the same prompt format.

The reference pipeline validates every prompt against the JSON schema before generation and rejects inputs that do not parse.

### Prompt example

```json
{
  "high_level_description": "A cinematic 35mm film photograph of a lone wooden sailboat on a glassy lake at sunset, the boat on a right-third vertical with the horizon at the lower third, in a cool muted blue palette.",
  "style_description": {
    "aesthetics": "Cinematic, minimal, serene, quiet stillness.",
    "lighting": "Cool overcast dusk light with a small warm sun low at the horizon; muted and low-contrast.",
    "photo": "35mm motion-picture film still, 16:9 framing, subtle grain, slightly desaturated.",
    "medium": "Photograph.",
    "color_palette": ["#1B3A5C", "#5B8FB9"]
  },
  "compositional_deconstruction": {
    "background": "Windless evening on a wide lake; horizon at the lower third. Dusty blue-violet sky with a small amber sun at the horizon and a thin gold streak across the glassy teal water. Subtle 35mm grain.",
    "elements": [
      {
        "type": "obj",
        "bbox": [380, 590, 660, 720],
        "desc": "Lone wooden sailboat on the right-third vertical in the midground, dark varnished hull and a single tall mast with a slack white sail hanging limp in the still air. No visible crew."
      },
      {
        "type": "obj",
        "desc": "Mirror-perfect reflection of the sailboat and its mast doubling straight down into the glassy teal water directly beneath the hull."
      },
      {
        "type": "obj",
        "desc": "Out-of-focus reed tips slicing into the lower-left foreground, soft dark blades against the teal water, blurred from proximity to the camera."
      }
    ]
  }
}
```

The structured surface provides:

- **Color palette conditioning:** Up to 16 hex colors per image and 5 per element steer dominant colors directly.
- **Bounding-box layout:** Place any element with `[y_min, x_min, y_max, x_max]` in 0–1000 normalized coordinates, with the image origin at the top-left. Shared MRoPE positional space supports the requested placement.
- **Typed text elements:** A `text` element carries the literal string to render and a separate visual description for its styling, supporting multi-line and multi-font in-image text.

## Benchmarks

We evaluate Ideogram 4.0 across layout control, text rendering, spatial reasoning, and prompt alignment against GPT Image 2, Nano Banana 2, and leading open-weight releases.

| Capability | Score | Metric and meaning |
|---|---:|---|
| Layout control | 0.69 | [7Bench](https://arxiv.org/abs/2508.12919) mIoU: how tightly generated objects sit inside requested bounding boxes. |
| Text rendering | 0.97 | [X-Omni](https://arxiv.org/abs/2507.22058) English OCR accuracy on in-image text. |
| Spatial reasoning | 0.76 | [SpatialGenEval](https://arxiv.org/abs/2601.20354) accuracy, combining spatial-relationship and basic-object question splits. |
| Prompt alignment | 0.89 | [Prism-Bench](https://arxiv.org/abs/2509.09680) alignment track for long, compositional inputs. |

We judge SpatialGenEval with `gemini-2.5-flash` rather than the leaderboard’s `Qwen2.5-VL-72B`, applied uniformly across all models.

![Five-axis capability radar comparing Ideogram 4.0 with other image-generation models](https://ideogram.ai/blog/ideogram-4.0/)

*Figure 2. Ideogram 4.0 versus leading closed-source models on layout control, spatial reasoning, object fidelity, prompt alignment, and text rendering. With Magic Prompt ON.*

![Parameter efficiency on text rendering](https://ideogram.ai/blog/ideogram-4.0/)

*Figure 3. Ideogram 4.0 is positioned in the small-but-strong area of the X-Omni English OCR comparison. Hunyuan Image v3 reports 80B total parameters with approximately 13B active per token.*

![Designer-preference ELO leaderboard](https://ideogram.ai/blog/ideogram-4.0/)

*Figure 4. In an internal arena, graphic designers chose the better of two generations without being told the producing model. Ideogram 4.0 placed second overall and first among open-weight models. Bars show ELO scores and cap lines mark 95% confidence intervals.*

## Examples

We generated the qualitative examples with `seed = 0` using the `fp8` quantized checkpoint. The examples are demonstrations of model behavior, not Ideogram business results.

### Detail and realism

Examples include a grandfather cradling an infant, a dragonfly on a fiddlehead fern, and a West Village street scene.

### Typography and text control

Examples include:

- A “Mango Sago Social” team tasting poster with title, subtitle, date, tagline, credit, and RSVP text placed using six text bounding boxes.
- A “Sound & Color Festival 2026” poster with stacked headline, lineup, date strip, sponsor row, ticket information, and small print.
- A six-word typographic composition using distinct lettering styles for “WILD,” “SLOW,” “softly,” “PIXEL,” “LOUD,” and “STUDIO NINE,” with independent scale, position, and rotation.

### Layout with bounding boxes

Examples include:

- A living room with a woman reading on a mustard sofa, a window, a monstera, and a cat; four object bounding boxes place the principal elements.
- A hand-drawn junk-drawer inventory with 28 distinct everyday objects and an inventory line; 28 object boxes and one text box are used.
- A toddler’s bedroom after play with an unmade bed and scattered toys; eight object boxes place room contents and two text boxes place the train label and book spine.

### Color palette control

We include three palette variants of a fantasy floating city, using these prompt palettes:

1. `#1B3A5C`, `#5B8FB9`
2. `#5C1B1B`, `#C0392B`
3. `#7A5C12`, `#E6B422`

The scenes are sample generations of a fantasy floating city with airborne islands, a central baroque cathedral spire, suspension bridges, airships, hanging gardens, and a cloud sea.

## Resolution flexibility

One set of weights covers square thumbnails through ultrawide banners. The noise schedule adjusts automatically per resolution.

| Use case | Resolution |
|---|---:|
| Square | 1024 × 1024 |
| Landscape | 1536 × 1024 |
| Portrait | 1024 × 1536 |
| Widescreen | 1920 × 1088 |
| Ultrawide | 2048 × 768 |
| Phone wallpaper | 1024 × 1792 |
| Social banner | 1584 × 396 |

## Safety

Our open-weight release uses three mitigation layers. We filter pre-training data with NSFW classifiers across multiple categories. Post-training procedures reduce unsafe generations, including from prompts that explicitly request them. At inference time, the reference pipeline screens every prompt and output through Hive moderation. We require equivalent or stronger filtering for redistributed deployments.

## Open weight

We publish the weights, inference code, full prompting guide, and sampler presets. The repository includes fp8 and nf4 checkpoints; the nf4 variant fits on one 24 GB GPU.

### Get the model

- [Hugging Face weights](https://huggingface.co/collections/ideogram-ai/ideogram-4)
- [GitHub inference code, prompting guide, and license](https://github.com/ideogram-oss/ideogram-4)
- [API](https://developer.ideogram.ai/)

## References

1. Tencent Hunyuan, [HunyuanImage 3.0 Technical Report](https://arxiv.org/abs/2509.23951), 2025.
2. Z-Image Team, [Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer](https://arxiv.org/abs/2511.22699), 2025.
3. Cai et al., [HiDream-O1-Image: A Natively Unified Image Generative Foundation Model with Pixel-level Unified Transformer](https://arxiv.org/abs/2605.11061), 2026.
4. Black Forest Labs, [FLUX.2: Frontier Visual Intelligence](https://bfl.ai/blog/flux-2), 2025.
5. Wu et al., [Qwen-Image Technical Report](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-Image/Qwen_Image.pdf), 2025.
6. Peebles and Xie, [Scalable Diffusion Models with Transformers](https://arxiv.org/abs/2212.09748), ICCV 2023.
7. Rombach et al., [High-Resolution Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752), CVPR 2022.
8. Qwen Team, [Qwen3-VL Technical Report](https://arxiv.org/abs/2511.21631), 2025.
9. Dehghani et al., [Scaling Vision Transformers to 22 Billion Parameters](https://arxiv.org/abs/2302.05442), ICML 2023.
10. Su et al., [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/abs/2104.09864), 2021.
11. Wang et al., [Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution](https://arxiv.org/abs/2409.12191), 2024.
12. Shazeer, [GLU Variants Improve Transformer](https://arxiv.org/abs/2002.05202), 2020.
13. Lipman et al., [Flow Matching for Generative Modeling](https://arxiv.org/abs/2210.02747), ICLR 2023.
14. Liu, Gong, and Liu, [Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow](https://arxiv.org/abs/2209.03003), ICLR 2023.
15. Esser et al., [Scaling Rectified Flow Transformers for High-Resolution Image Synthesis](https://arxiv.org/abs/2403.03206), ICML 2024.
16. Ho and Salimans, [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598), 2022.
17. Izzo et al., [7Bench: a Comprehensive Benchmark for Layout-guided Text-to-Image Models](https://arxiv.org/abs/2508.12919), ICIAP 2025.
18. Wang et al., [Everything in Its Place: Benchmarking Spatial Intelligence of Text-to-Image Models](https://arxiv.org/abs/2601.20354), ICLR 2026.
19. Geng et al., [X-Omni: Reinforcement Learning Makes Discrete Autoregressive Image Generative Models Great Again](https://arxiv.org/abs/2507.22058), 2025.
20. Fang et al., [FLUX-Reason-6M & PRISM-Bench: A Million-Scale Text-to-Image Reasoning Dataset and Comprehensive Benchmark](https://arxiv.org/abs/2509.09680), ICLR 2026.
21. Gutflaish et al., [Generating an Image From 1,000 Words: Enhancing Text-to-Image With Structured Captions](https://arxiv.org/abs/2511.06876), 2025.

## Cite this work

```bibtex
@misc{ideogram-4-2026,
  author       = {Ideogram AI},
  title        = {{Ideogram 4.0}},
  year         = {2026},
  howpublished = {\\url{https://ideogram.ai/blog/ideogram-4.0/}}
}
```

## We’re hiring

We’re looking for Research Scientists and Research Engineers to work on next-generation generative models and the products built on top of them. [Apply at Ideogram Careers](https://ideogram.ai/careers).

---

[Read this page’s summary](https://ideogram.ai/agents/api/public/v1/page-tldr?url=https%3A%2F%2Fideogram.ai%2Fblog%2Fideogram-4.0%2F&trace=t_380edc68d53d)

## Ask a follow-up about this page

Need a more specific answer? Ask the exact question you have. It returns a fresh, cited answer from Ideogram’s knowledge base. No account or API key required.

Open this URL with your own question, URL-encoded. Include your question in q. All research inputs are optional; no account or setup.

`GET https://ideogram.ai/agents/agent/faq?q={question}&trace=t_380edc68d53d`

[Tool guide and examples](https://ideogram.ai/agents/agent-access) · [OpenAPI](https://ideogram.ai/agents/api/public/v1/openapi.json) · [Page TLDR form](https://ideogram.ai/agents/agent-tools/page-tldr?url=https%3A%2F%2Fideogram.ai%2Fblog%2Fideogram-4.0%2F). Opening a form does not execute a tool.