Ideogram 4.0 in ComfyUI: Complete Install and Setup Guide (2026)

What to Expect from Ideogram 4.0

Before you spend time on setup, here are four real examples generated with the workflow you're about to install. Each one uses a structured JSON prompt — the format this model was trained on. Click View Prompt on any card to see the exact prompt, and hit Copy Prompt to grab it and try it yourself once the workflow is running.

🔍 Click to zoomFantasy Movie Poster

Whimsical adventure movie poster combining storybook fantasy, mixed-media collage design, and large-scale cinematic typography.

▶View PromptCopy Prompt

🔍 Click to zoomFashion Editorial Portrait

High-end fashion and travel editorial photography with natural beauty, cinematic lighting, and realistic environmental details.

▶View PromptCopy Prompt

🔍 Click to zoomAction Photography

Cinematic high-speed motorcycle photography featuring a perfectly sharp rider against dynamic urban motion blur and light trails.

▶View PromptCopy Prompt

🔍 Click to zoomSci-Fi Movie Poster

Hollywood-scale sci-fi poster featuring planetary collision, cosmic energy, monumental architecture, and cinematic worldbuilding.

▶View PromptCopy Prompt

What Is Ideogram 4.0 and Why Does It Work Differently?

Ideogram 4.0 is a 9.3B parameter open-weight text-to-image model released by Ideogram AI on June 3, 2026. "Open-weight" means the model files are publicly downloadable — you run them on your own machine, not through a paid service. It was built entirely from scratch, not fine-tuned from any existing model like FLUX or Stable Diffusion.

Two things make it structurally different from every other model you may have used. First, the text encoder. Most image models use CLIP or T5 — text-only encoders with limited language understanding. Ideogram 4.0 uses Qwen3-VL-8B-Instruct (a full vision-language model) plus Gemma4, and extracts hidden states from 13 of Qwen's intermediate layers. This gives the model far deeper understanding of what you're describing — which is why it renders in-image text, precise layouts, and complex compositions better than models three times its size.

Second, the prompt format. Ideogram 4.0 was trained exclusively on structured JSON captions, not plain text. Plain-text prompts still work, but produce noticeably weaker results because the model was never trained on them. The JSON format is covered in detail in the Ideogram 4.0 Prompt Guide.

How Does Ideogram 4.0 Compare to FLUX and Stable Diffusion?

At 9.3B parameters it outperforms FLUX.2 [dev] (32B) and Qwen-Image (20B) on text rendering benchmarks. In a blind evaluation by professional graphic designers, it ranked #2 overall — behind only GPT Image 2 — and #1 among all open-weight models.

What Is the Ideogram 4.0 License? Can You Use It Commercially?

There are two license tiers. Get this straight before you download anything.

One important distinction: the inference code on GitHub is Apache 2.0 — free for any use including commercial. The restriction is on the model weights, not the code that runs them. Images you generate are yours — Ideogram does not claim ownership of outputs. You can use generated images commercially as long as you're not using the weights themselves in a commercial deployment.

What You Need Before You Start

The 5 Model Files You Need

The table below covers the FP8 versions — recommended for 16 GB+ VRAM cards. If you have less than 16 GB, see the NVFP4 note in the section below.

⚠ Yellow rows = both text encoders are required. Missing either one causes generation to fail.

FP8 vs NVFP4 — Which Should You Download?

The Comfy-Org HuggingFace repo contains two quantization options for the diffusion models. Here's what each one actually is and when to use it:

Where to Put the Files in ComfyUI

This is where most beginners run into problems. Ideogram 4.0 does not use the models/checkpoints/ folder that FLUX and SD models go into. It uses three separate subfolders. The ComfyUI root folder is the main folder where ComfyUI is installed — it contains main.py, a models/ folder, a custom_nodes/ folder, and others.

📂 ComfyUI/
└── 📂 models/
    ├── 📂 vae/
    │   └── flux2-vae.safetensors
    ├── 📂 diffusion_models/
    │   ├── ideogram4_fp8_scaled.safetensors
    │   └── ideogram4_unconditional_fp8_scaled.safetensors
    │   (or _nvfp4_mixed.safetensors variants for lower VRAM)
    └── 📂 text_encoders/
        ├── qwen3vl_8b_fp8_scaled.safetensors
        └── gemma4_e4b_it_fp8_scaled.safetensors

How to Load and Run the Ideogram 4.0 Workflow in ComfyUI

Step 1 — Update ComfyUI First

Ideogram 4.0 requires a recent version of ComfyUI. If you're running an older version, some of the workflow nodes won't exist and the canvas loads with red errors.

1. Open ComfyUI Manager — the Manager button in the top menu bar.
2. Click Update ComfyUI.
3. Wait for the update to complete, then restart ComfyUI completely — close the terminal window and relaunch it.

Step 2 — Load the Workflow JSON

1. Save the downloaded .json file anywhere — your Downloads folder is fine.
2. In ComfyUI, drag the JSON file directly onto the canvas. Or click the Load button in the top menu and navigate to the file.
3. The canvas populates with the workflow nodes connected by wires. You'll see a Load Diffusion Model node, a DualCLIPLoader node, a VAELoader node, and a sampler chain leading to an output node.

Step 3 — Check the Model Nodes Are Pointing to the Right Files

Before running anything, confirm each model node is pointing to the correct file. ComfyUI reads from the folder you set up in the previous step — if a file is missing or misplaced, the node dropdown shows nothing.

1. Click the Load Diffusion Model node. The filename dropdown should show ideogram4_fp8_scaled.safetensors. If it shows "None", the file isn't in models/diffusion_models/ — go back and move it.
2. Click the DualCLIPLoader node. Both text encoder fields should show the Qwen3 and Gemma4 filenames. These two encoders convert your prompt into something the model understands — without them, generation fails or produces noise.
3. Click the VAELoader node. It should show flux2-vae.safetensors. The VAE (Variational Autoencoder — the component that converts the model's internal latent image into an actual pixel image you can see and save) is required for output.

Step 4 — Write Your First Prompt

Find the prompt text box on the canvas — it's connected to the node chain that feeds into the sampler. For your first run, write a short plain-text description:

a red apple on a white marble table, photorealistic, soft natural light

Plain text works fine here — the model will produce a good result. After you've confirmed the workflow runs cleanly, read the Ideogram 4.0 Prompt Guide to learn the JSON format. That's where you'll see a real jump in quality.

Step 5 — Run and What to Expect

Click the orange Queue Prompt button in the top-right corner of the ComfyUI interface. A progress bar appears below it and the active node highlights as each step runs.

The first run is always slower regardless of preset — model files load into VRAM for the first time. From the second generation onward, speed improves because the models stay cached in VRAM. The finished image appears in the output preview node on the right side of the canvas.

Why Does the Prompt Look Like JSON? (The Prompting Difference Explained)

Every other image model you've used accepts plain text: "a woman standing in a forest at sunset." Ideogram 4.0 was trained on something structurally different — structured JSON captions. JSON (JavaScript Object Notation) is a structured text format that uses { } brackets to organize information into named fields. You don't need to know any code to use it — you're filling in a template, not writing a program.

The JSON format gives Ideogram 4.0 three capabilities that plain text can't match: exact hex color palette control (up to 16 colors per image), bounding-box layout for placing subjects at exact coordinates, and typed text elements that let the model know precisely what characters to render and where. Because the model was trained exclusively on captions structured this way, plain-text prompts sample outside the training distribution and leave quality on the table — especially for composition, color, and typography.

You can also skip writing JSON entirely using Magic Prompt — a free tool built into the reference pipeline that auto-converts a plain-text prompt into a full JSON caption. It runs server-side using a free Ideogram API key with no local model required. The Ideogram 4.0 Prompt Guide covers every JSON field with real before/after examples, and includes a downloadable skill file that converts your normal prompts to proper Ideogram 4.0 JSON automatically.

Troubleshooting Ideogram 4.0 in ComfyUI

"NODE NOT FOUND" or Red Nodes After Loading the Workflow

What it looks like: One or more nodes on the canvas turn red with a "NODE NOT FOUND" label after you load the workflow JSON.

What causes it: ComfyUI is out of date and doesn't yet support the Ideogram 4.0 node types.

How to fix it:

1. Open ComfyUI Manager from the top menu.
2. Click Update ComfyUI.
3. Restart ComfyUI completely.
4. Reload the workflow JSON.

If nodes are still red after updating, follow the step-by-step guide to fixing missing nodes in ComfyUI.

Model Files Not Appearing in the Node Dropdowns

What it looks like: The Load Diffusion Model, DualCLIPLoader, or VAELoader node dropdowns show "None" or an empty list.

What causes it: Files are in the wrong subfolder (most common: placed in models/checkpoints/ instead of models/diffusion_models/), or the filenames were modified during download.

How to fix it: Check each file is in the exact subfolder shown in the folder structure diagram. Confirm the filenames match exactly. Restart ComfyUI — it only scans model folders on launch.

"Image Blocked by Safety Filter" — Gray Screen Output

What it looks like: Generation completes but instead of an image you see a plain gray screen with the text "Image blocked by safety filter."

What causes it: The safety filter flagged your prompt. The official Ideogram 4.0 documentation confirms false positive rates are higher for plain-text prompts than for structured JSON prompts.

How to fix it: Rewrite the prompt to be more specific and descriptive. Switch from plain text to a JSON-structured prompt — this alone resolves most false positives. The Ideogram 4.0 Prompt Guide shows you how.

CUDA Out of Memory — Generation Fails Mid-Run

What it looks like: ComfyUI shows a red error in the terminal or on the canvas: "CUDA out of memory" or "RuntimeError: CUDA error."

What causes it: The FP8 model files require at least 16 GB VRAM. If your GPU has less, the model can't fully load.

How to fix it:

1. Close any other applications using your GPU (games, other AI tools).
2. Add --lowvram or --medvram to your ComfyUI launch command.
3. If you have less than 16 GB VRAM, switch to the NVFP4 setup — use FP8 for the main model and NVFP4 for the unconditional model to reduce VRAM usage while preserving most of the quality.

Frequently Asked Questions

Published: 2026-06-03 · Last updated: 2026-06-11 · Tested on RTX 4090 (24 GB VRAM) · ComfyUI version: [ADD VERSION]