Ideogram 4 in ComfyUI: local image generation with layout control

1. What Ideogram 4 adds

In the video, AI Search argues that Ideogram 4 is especially strong at three things that often trip up local image models: world knowledge, text rendering, and prompt adherence. The important difference is that the recommended ComfyUI workflow does not rely only on one big text prompt. It uses a prompt-builder canvas where you draw boxes for the elements you want and describe each object or text region.

• World knowledge: the presenter shows examples involving recognizable fictional/game/anime-style characters and mixed group scenes. As always, verify rights before publishing or selling outputs involving protected characters or logos.
• Text rendering: the workflow can place labels, titles, speech bubbles, and typographic instructions more reliably than many local models.
• Prompt adherence: when objects are placed with bounding boxes, it can follow complex layouts such as posters, multi-panel manga pages, and scenes with many separate objects.
• Fine placement control: bounding boxes can define title blocks, background areas, people, hands, props, windows, posters, panels, and small details.

2. Requirements and downloads

Software

• ComfyUI installed and working. The video uses the Windows portable build, but the model files/folders are the same conceptually on other systems.
• ComfyUI Manager installed and enabled, so missing nodes can be detected and installed from the Manager UI.
• Git for manually cloning KJ Nodes if Manager does not install it automatically.

Model files confirmed from Hugging Face

The official Comfy-Org/Ideogram-4 repository currently lists these relevant files:

• diffusion_models/ideogram4_fp8_scaled.safetensors
• diffusion_models/ideogram4_nvfp4_mixed.safetensors
• diffusion_models/ideogram4_unconditional_fp8_scaled.safetensors
• diffusion_models/ideogram4_unconditional_nvfp4_mixed.safetensors
• text_encoders/qwen3vl_8b_fp8_scaled.safetensors
• text_encoders/qwen3vl_8b_nvfp4.safetensors
• vae/flux2-vae.safetensors

Choose one main Ideogram model, one matching unconditional model, one Qwen3-VL text encoder, and the Flux 2 VAE. The video uses FP8 versions; NVFP4 is smaller/more compressed if your GPU/runtime supports it.

3. Installation steps

Start with a known-good ComfyUI install

Launch ComfyUI once and confirm the browser interface loads. If you are new to ComfyUI, do not start with Ideogram 4 as your first test; first verify that ComfyUI can run a smaller default workflow.

Download the video’s workflow

The source description links a workflow file on Google Drive. Download it from the Workflow link and save it somewhere convenient, such as your main ComfyUI folder or a dedicated workflows folder.

Then drag the workflow file onto the ComfyUI browser window. Expect red/missing nodes or missing model warnings at first; the next steps resolve those.

Install or enable ComfyUI Manager

If Manager is already visible in the ComfyUI interface, open it and skip to the missing-node install. If not, follow the official ComfyUI Manager installation docs. For the Windows portable build, the docs/video emphasize launching ComfyUI with --enable-manager.

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --enable-manager

If your run.bat does not include --enable-manager, add it to the ComfyUI startup command, save the file, and relaunch ComfyUI.

Install missing nodes from Manager

1. Open Manager in ComfyUI.
2. Go to Missing Nodes.
3. Install the nodes listed for the imported workflow.
4. Apply changes and restart ComfyUI when prompted.

Manually install KJ Nodes if the prompt builder is still missing

The key node in the video is the KJ prompt-builder/canvas node. If it remains missing after Manager install, clone KJ Nodes directly into ComfyUI/custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-KJNodes.git

If it is already installed but stale, update it:

cd ComfyUI/custom_nodes/ComfyUI-KJNodes
git pull

Restart ComfyUI. The missing KJ node warning should disappear, and the workflow should show a canvas where you can drag boxes.

Place the model files in the right folders

From Comfy-Org/Ideogram-4, download:

• Main diffusion model: FP8 or NVFP4 Ideogram 4 main model → ComfyUI/models/diffusion_models/
• Unconditional diffusion model: FP8 or NVFP4 unconditional model → ComfyUI/models/diffusion_models/
• Text encoder: Qwen3VL 8B FP8 or NVFP4 → ComfyUI/models/text_encoders/
• VAE: Flux 2 VAE → ComfyUI/models/vae/

Back in ComfyUI, press R or use the refresh option so model dropdowns update, then select the matching files in the workflow. When everything is selected correctly, red model warnings should clear.

4. How to use the workflow

The workflow has three conceptual areas:

1. Canvas/image size node: choose aspect ratio and megapixels. The video shows ratios ranging from very vertical to very wide.
2. Ideogram prompt-builder node: enter the high-level description, background, style, aesthetics, lighting, medium, and per-object/per-text boxes.
3. Seed/batch controls: choose fixed seed, randomized seed, and number of outputs per run.

Suggested first test

Start with a simple poster because it exercises the text and layout strengths without requiring too many objects:

• Aspect ratio: 3:4 or 4:5.
• High-level description: “minimalist product poster for three matcha drinks.”
• Background: “minimalist cafe setting.”
• Style/medium: poster, product photography, smooth professional lighting.
• Boxes: title at top, three glasses in the middle, three labels below, optional small logo/brand placeholder if you have rights to use it.

5. Bounding-box prompting pattern

Use the canvas as a layout editor. Each box should answer: what is here, what type is it, what style/color/typography should it use, and how should it relate to nearby boxes?

Objects

• Use object boxes for people, props, animals, furniture, products, backgrounds, windows, signs, and small details.
• Large background boxes can sit behind everything and describe the scene environment.
• Micro boxes can control hands, props, feet, facial expression, or a small object that a larger subject should hold.

Text

• Switch the element type to text for titles, labels, speech bubbles, onomatopoeia, buttons, and poster copy.
• Put the exact text in the text field.
• Put typography instructions in the description: “large bold blocky sans-serif,” “small italic subtitle,” “white text on black strip,” “deep green color,” etc.
• Use the color picker/reference colors when you need stronger color adherence.

Comics and manga

For a manga page, make each panel a box and describe its camera angle, action, mood, and any dialogue/sound effects. The video demonstrates wide establishing panels, close-ups, speech bubbles, and bottom action panels. Expect to iterate; text-heavy comic pages are still more fragile than simple posters.

6. Settings, seeds, variations, and batch size

• Aspect ratio and megapixels: set these before carefully arranging boxes, because changing the canvas shape can alter the layout.
• Fixed seed: keep the seed fixed when you want roughly the same style while moving or resizing boxes.
• New fixed random: choose a new seed to get a different variation while keeping the same prompt and boxes.
• Randomize each time: use when exploring broadly rather than refining one composition.
• Batch size: increase to 2 or 4 only if your GPU/RAM can handle it. The video estimates roughly one minute per image on the presenter’s system, slower than Z-Image or Flux-style faster workflows.
• Grab background: this sets the generated output as a visual reference behind your boxes so you can reposition elements. It is not image-to-image conditioning; Ideogram 4 in this workflow is still text-to-image.
• Alt-click: when boxes overlap, hold Alt and click to cycle/select elements behind the front item.
• Ctrl+D: duplicate a selected box, useful for product labels, repeated drinks, panels, or similar objects.

7. Troubleshooting

“Missing node” errors

• Install missing nodes through ComfyUI Manager, apply changes, and restart.
• If the KJ prompt builder is still missing, manually clone kijai/ComfyUI-KJNodes into ComfyUI/custom_nodes and restart.
• If KJ Nodes already exists, run git pull inside its folder to update.

Model dropdowns are empty or red

• Confirm files are in the exact folders: models/diffusion_models, models/text_encoders, and models/vae.
• Press R or restart ComfyUI to refresh model lists.
• Make sure you selected both the main Ideogram model and the unconditional model. The video stresses that both are required.

Generation appears blocked or returns a safety-filter image

• Do not treat this as proof the model cannot generate. In the tutorial, the fix was to draw bounding boxes for the main objects instead of using only the high-level prompt.
• Break the prompt into specific boxes: subject 1, subject 2, prop, background, text, etc.
• Keep the prompt lawful and consent-based; avoid prohibited or exploitative content.

Output layout is close but not exact

• Use Grab background as a visual guide, then resize or move boxes.
• Keep the seed fixed for small layout edits.
• For hidden/overlapping elements, use Alt-click to select lower layers.
• Use shorter, clearer per-box descriptions instead of one long paragraph.

Out of memory or very slow generation

• Try the more compressed model variants if supported by your GPU/runtime.
• Reduce megapixels or batch size.
• Close other GPU-heavy apps.
• ComfyUI can offload to system RAM, but large model files still require patience and enough total RAM/storage.

8. Sources and links

• Primary video: New BEST local AI image generator is here! — AI Search
• Workflow from video description: Google Drive workflow file
• ComfyUI Manager installation docs: docs.comfy.org/manager/install
• KJ Nodes GitHub repository: github.com/kijai/ComfyUI-KJNodes
• Ideogram 4 model repository: huggingface.co/Comfy-Org/Ideogram-4
• Related tutorials from the source description: ComfyUI tutorial, Z-Image tutorial, Ernie image tutorial