Ideogram V4 Is a 9.3B Diffusion Transformer With Open Weights and $0.03 Turbo Image Generation

Ideogram V4 is a 9.3-billion-parameter single-stream Diffusion Transformer for text-to-image generation, developed by Ideogram and released on June 3, 2026, with a Qwen3-VL-8B-Instruct text encoder driving its industry-leading typography and layout control. The model ranks #1 among open-weight image models on Design Arena, reports 0.97 on X-Omni English OCR, and is offered through Ideogram's own API at $0.03 per Turbo image, $0.06 Default, and $0.10 Quality. The Instant and Fast variants were open-sourced on Hugging Face on July 13, 2026 under a non-commercial license.

Key Takeaways

• Ideogram V4 is a 9.3B single-stream Diffusion Transformer with native 256–2048 px generation and a Qwen3-VL-8B-Instruct text encoder.
• It accepts structured JSON prompts with bounding boxes, hex palettes, and per-region CFG weights for fine-grained layout control.
• Official API pricing: Turbo $0.03, Default $0.06, Quality $0.10 per image.
• Instant (8-step, ~0.4 s / 1K) and Fast (~1 s / 1K) weights are open under the Ideogram Non-Commercial Model Agreement; inference code is Apache 2.0.
• On Design Arena it ranks #1 open-weight and #2 overall in designer preferences, with X-Omni English OCR of 0.97 and 7Bench mIoU of 0.69.
• Community integrations already live: ComfyUI workflows, LoRA training via AI Toolkit, and hosted access on fal.

What Is Ideogram V4?

Ideogram V4 is the fourth-generation image model from Ideogram, positioned as a designer-first text-to-image system. It targets the workloads where prior diffusion models tended to fail: legible in-image text, brand-consistent typography, poster and UI layouts, and photorealistic portraits with clean skin rendering.

Architecturally it is a single-stream Diffusion Transformer with 9.3B parameters, trained from scratch with flow matching. Text conditioning comes from a Qwen3-VL-8B-Instruct vision-language model, and the model exposes multi-layer hidden states so prompts can carry both semantic instructions and structured layout metadata. The native resolution range spans 256 to 2048 pixels.

Ideogram V4 is officially released. It was announced on Ideogram's own tech blog on June 3, 2026, and the Instant and Fast open-weight variants followed on July 13, 2026 via fal on Hugging Face.

Ideogram V4 at a Glance

Field	Detail
Developer	Ideogram
Type	Single-stream Diffusion Transformer (flow matching)
Parameters	9.3B (base)
Text encoder	Qwen3-VL-8B-Instruct (multi-layer hidden states)
Modality	Text-to-image, structured JSON prompts
Native resolution	256–2048 px
Prompt features	Bounding boxes, hex palettes, per-region CFG
Hosted pricing	Turbo $0.03 / Default $0.06 / Quality $0.10 per image
Open variants	Instant (8-step) and Fast weights
License (weights)	Ideogram Non-Commercial Model Agreement
License (inference code)	Apache 2.0
Availability	Ideogram web + API, Hugging Face, fal
Release date	June 3, 2026 (base); July 13, 2026 (open Instant/Fast weights)

How Ideogram V4 Works and What Makes It Different

The core mechanism is a Single-Stream Diffusion Transformer rather than the dual-DiT designs used by several competitors. All modalities flow through one transformer stack, which simplifies the guidance path and made later CFG distillation possible.

The text pipeline is the differentiator. Instead of a CLIP or T5 encoder, Ideogram V4 uses the full Qwen3-VL-8B-Instruct model and taps its hidden states across multiple layers. That capacity is what the model spends on typography: rendering long strings, honoring font weight cues, and holding character shapes together at small pixel counts. On the X-Omni English OCR benchmark it scores 0.97, and on 7Bench mIoU (a layout adherence measure) it lands at 0.69.

Prompts are not just strings. Ideogram V4 accepts Structured JSON Prompts carrying bounding boxes, hex color palettes, and per-region CFG weights. That is why designer benchmarks weight it so heavily: a user can pin "brand-red H1 in the top-third band" and the model will respect it. The API exposes this directly through the json_prompt field on the /v1/ideogram-v4/generate endpoint, and there is a matching describe endpoint that returns a structured V4JsonPrompt from any image so a layout can be round-tripped.

Two speed-optimized variants shipped alongside the base model. Instant runs in 8 steps with no runtime CFG (achieved via CFG distillation) and pre-QAD BF16 weights around 9.3B. Fast applies timestep distillation on top. On fal's hosted infrastructure, Instant reaches ~0.4 s per 1K image and Fast ~1 s, versus roughly 2.75 s for the base BF16 model — a 6–8× speedup that fal attributes to NVFP4 quantization, epilogue fusion of RMSNorm and SwiGLU, and quantization-aware distillation for color recovery, per fal's July 9 technical blog.

Community engineers are extending the model within weeks of the open-weight release — a fair proxy for real open-weight momentum, with ComfyUI ports of the Fast and Instant variants and LoRA training pipelines both landing in mid-July.

What You Can Do With Ideogram V4

• Poster and marketing artwork with legible headline typography and precise color spec.
• UI mockups and app screens using structured JSON prompts to place components in bounding boxes.
• Consistent characters via LoRA training, now documented in ComfyUI + AI Toolkit workflows by @mickmumpitz's Consistent Character Creator 4.0 tutorial.
• Photorealistic portraits where community reads praise skin texture and lighting realism.
• High-volume iteration on Instant / Fast — sub-second generation enables live prompt-tuning rather than batch waits.
• Multi-model pipelines, chaining Ideogram V4 for text-heavy layers with other models for upscaling and refinement.

One consistent visual tell has been flagged by users: @el_mejnun observed a persistent tendency toward smoke, haze, volumetric fog, and vignette lighting even on lightly prompted scenes. Treat it as a style bias worth countering with explicit negative prompts.

How Ideogram V4 Compares

Model	Params	Text rendering	Open weights	Hosted price/image
Ideogram V4 (base)	9.3B	X-Omni OCR 0.97	Instant + Fast, non-commercial	$0.03–$0.10
Ideogram V3	Not disclosed	Strong but lower	No	Ideogram API only
Typical open DiT peers	2–12B	Varies, generally lower on OCR	Yes (varied licenses)	Varies

On Design Arena's designer-preference vote, Ideogram V4 sits at #1 among open-weight models and #2 overall, per Ideogram's June 3, 2026 launch post. Independent blind comparisons at 2K resolution and complex multi-region JSON layouts are still limited.

Availability: How to Access Ideogram V4

The official access path is the Ideogram web app and Ideogram API, with the three-tier pricing above. Commercial self-hosting requires a paid Ideogram tier — the open weights alone do not license commercial deployment.

For research and personal use, @fal open-sourced the Instant and Fast weights on July 13, 2026, inheriting Ideogram's non-commercial license. Quantized nf4 and fp8 checkpoints are on Hugging Face, and ComfyUI has community workflows for local inference. fal also offers free sandbox generations of Instant and Fast on its hosted platform.

If you are building image workflows and want to test comparable hosted models with API-first access, Ideogram V3 remains a solid production choice for text-heavy generation while you evaluate V4 locally.

What We Don't Know Yet

• Exact commercial self-hosting license pricing, revenue thresholds, and IP terms for LoRAs / derivatives are not fully public — the non-commercial free tier is clear, paid tier details are less so.
• Independent large-scale blind studies comparing Instant (8-step) and Fast against base Quality at 2K and on complex JSON layouts are not published. Community reads range from "no visible loss" to minor color shifts.
• Whether Ideogram will release an official distilled checkpoint under more permissive commercial terms is open.
• Performance characteristics on very complex multi-element JSON layouts at maximum 2048 px are lightly covered.

Frequently Asked Questions

What is Ideogram V4?

Ideogram V4 is a 9.3-billion-parameter single-stream Diffusion Transformer for text-to-image generation, released by Ideogram on June 3, 2026. It uses a Qwen3-VL-8B-Instruct text encoder and accepts structured JSON prompts with bounding boxes and hex palettes.

Is Ideogram V4 open source?

The Instant and Fast weights are open on Hugging Face, but they ship under the Ideogram Non-Commercial Model Agreement, which restricts self-hosted commercial use to paid tiers. Inference code is Apache 2.0.

How much does Ideogram V4 cost?

On the official Ideogram API, Turbo generations cost $0.03 per image, Default $0.06, and Quality $0.10. The Instant and Fast open-weight variants can be run locally for free under the non-commercial license.

How fast is Ideogram V4 Instant?

Ideogram V4 Instant generates a 1K image in about 0.4 seconds on fal's optimized hosted infrastructure, compared with roughly 2.75 seconds for the base BF16 model. The Fast variant lands around 1 second per 1K image.

What makes Ideogram V4 different from other image models?

Ideogram V4 leans on a large VLM text encoder with multi-layer hidden states and accepts structured JSON prompts, which drives its strong typography and layout scores. It ranks #1 among open-weight models on Design Arena and #2 overall in designer preferences.

How does Ideogram V4 compare to Ideogram V3?

Ideogram V4 is a larger single-stream Diffusion Transformer with a newer Qwen3-VL text encoder, higher benchmark scores in text rendering, and open-weight Instant and Fast variants that Ideogram V3 did not offer. V3 remains available on the Ideogram API for existing pipelines.

Where can I try Ideogram V4?

Ideogram V4 is live on the official Ideogram web app and API, with paid tiers for commercial use. The Instant and Fast weights are downloadable from Hugging Face for research and personal projects, and are also available on fal's hosted platform.

What to Watch Next

Three signals will shape the next chapter for Ideogram V4. First, whether Ideogram publishes clearer commercial self-hosting terms — the ambiguity is the main friction point creators keep flagging. Second, whether an independent blind study confirms or challenges the "no visible loss" claim for Instant at 2K resolutions. Third, whether Ideogram itself releases an official distilled or commercial-friendly open variant, which would reset the field again.

Building similar text-to-image or design-first workflows? On kie.ai you can try Ideogram V3, Seedream 4.0, and Imagen 4.