Ideogram 4 Is Here: The Ultimate JSON Prompting Masterclass

Learn how to run Ideogram 4 locally with SwarmUI and ComfyUI, download the required model bundle, use the ready Turbo, Balanced, and Highest Quality presets, and create accurate structured JSON prompts with Ultimate Image Captioner Pro.

The workflow covers image recreation, reliable text rendering, bounding-box editing, batch captioning, and training-dataset preparation.

Full Tutorial

Watch the complete step-by-step tutorial on YouTube:

Ideogram 4: The Ultimate JSON Prompting Masterclass

Tutorial Resources

Resource	Link
Ultimate Image Captioner Pro download and installer	Patreon post
SwarmUI installer, model downloader, and presets	Patreon post
ComfyUI installer	Patreon post
Windows requirements tutorial	YouTube
Requirements guide with links and screenshots	Patreon post
Community support	SECourses Discord
More tutorials, installers, and resources	SECourses GitHub

Supported Models

Ultimate Image Captioner Pro supports the following models with robust torch.compile integration.

Qwen Vision Models

• Qwen3-VL 8B Instruct (default)
• Huihui Qwen3-VL 8B Instruct Abliterated
• Qwen3-VL 4B Instruct
• Qwen3-VL 2B Instruct
• Qwen3-VL 30B-A3B Instruct
• Qwen3.6 27B
• Huihui Qwen3.6 27B Abliterated

Joy Caption Models

• Joy Caption Beta 1
• Joy Caption Alpha 2
• Joy Caption Alpha 1
• Joy Caption Pre Alpha

Torch 2.13 Runtime

The application uses Torch 2.13 with the latest project-tested, precompiled supporting libraries.

Application Preview

Click the image to open the full-size screenshot.

Torch Compile Performance

The fully compiled captioning path delivers an 84% speed improvement in the demonstrated benchmark.

Installers

Installer workflows are available for Windows, RunPod, SimplePod, Massed Compute, and local Linux systems.

Video Chapters

Show all tutorial chapters

• 00:00:00 - Ideogram 4 overview: JSON prompting, SwarmUI presets, ComfyUI workflows, and model bundle
• 00:00:53 - Ultimate Image Captioner Pro for turning reference images into Ideogram JSON prompts
• 00:01:10 - Editing JSON elements, bounding boxes, wanted text fields, captions, and prompt layout
• 00:02:02 - Regeneration examples showing structure, objects, scene layout, and image text matching
• 00:03:18 - Captioner Pro feature tour: Qwen, JoyCaption, saved outputs, and JSON builder
• 00:04:30 - Dataset workflow: prompt presets, batch folder captioning, and automatic VRAM presets
• 00:05:13 - Tutorial roadmap: ComfyUI update, SwarmUI update, model download, app installation, and usage
• 00:05:40 - Updating ComfyUI by extracting the latest installer ZIP and overwriting old files
• 00:05:56 - Optional fresh ComfyUI virtual environment rebuild for outdated or broken installations
• 00:06:15 - Running the ComfyUI update script, Python choice, UV speed, and quantization support
• 00:07:05 - Installing recommended custom nodes bundle 100 for ComfyUI and SwarmUI compatibility
• 00:07:50 - Launching fresh ComfyUI and testing the Ideogram Turbo preset workflow
• 00:08:47 - Setting width, height, resolution, and matching the prompt aspect ratio
• 00:09:07 - Updating SwarmUI with the latest ZIP, overwrite method, and safe folder paths
• 00:09:48 - Automatic .NET SDK 10 installation and why SwarmUI needs the correct SDK version
• 00:10:51 - SwarmUI backend setup: ComfyUI backend, Triton, Sage Attention cautions, and extra arguments
• 00:11:44 - Downloading the Ideogram 4 core bundle with hash verification
• 00:12:28 - 16-connection parallel downloads, target folders, ComfyUI mode, and URL downloader
• 00:13:20 - Merging model parts and sharing SwarmUI models through extra_model_paths.yaml
• 00:13:51 - Setting the SwarmUI model root to reuse another model folder and avoid duplicates
• 00:14:12 - Updating SwarmUI presets with delete import, normal import, overwrite, and backup
• 00:14:58 - Refreshing presets and confirming Ideogram Turbo, Balanced, and Highest Quality
• 00:15:14 - First simple Ideogram prompt, false safety-filter block, and weak plain prompting
• 00:15:34 - Using Realism Engine Ideogram 5 LoRA to fix the blocked car prompt
• 00:15:57 - Why detailed JSON prompts are needed and downloading Captioner Pro
• 00:16:23 - Installing Captioner Pro with the Windows install/update app, virtual environment, and model downloads
• 00:16:34 - Windows requirements: Python, CUDA, cuDNN, C++ tools, FFmpeg, Git, and setup guide
• 00:17:03 - Cloud and Linux notes plus the Massed Compute interface, creator image, GPU, and coupon
• 00:17:34 - Captioner installer downloader: 16 connections, hash checks, and accurate setup
• 00:17:57 - Starting Ultimate Image Captioner Pro and saving custom user presets
• 00:18:14 - Loading the Bugatti reference image and generating official Ideogram JSON
• 00:18:39 - Prompt generation speed, copying the prompt, and understanding VRAM usage
• 00:19:09 - Subprocess mode to release all VRAM and RAM after each captioning run
• 00:19:54 - Reviewing generated JSON: high-level description, visible text, boxes, and details
• 00:20:21 - Pasting JSON into SwarmUI and matching the custom 5:3 aspect ratio
• 00:20:43 - Aspect-ratio calculator, side-length control, and high-resolution generation
• 00:21:36 - Comparing results with and without aspect-ratio metadata and avoiding false safety blocks
• 00:21:58 - Realism Engine LoRA strength, when to use it, and output comparison
• 00:22:34 - Choosing Turbo, Balanced, or Highest Quality and testing Turbo speed
• 00:22:54 - Ideogram 4 image-to-image, inpainting, image creativity, and image prompts
• 00:23:19 - Captioner Pro batch-folder processing: subfolders, overwrite, and append modes
• 00:23:35 - Post-processing captions with prefixes, suffixes, replacements, and sensitivity
• 00:24:07 - Final options, automatic quantization by GPU VRAM, support channels, and closing

Covered in the Tutorial

• Local Ideogram 4 installation
• SwarmUI and ComfyUI preset usage
• Automatic model downloads and hash verification
• Structured JSON prompt creation
• Bounding-box and visible-text editing
• Reference-image recreation
• Safety-filter troubleshooting and LoRA realism settings
• Folder-based batch captioning
• VRAM-friendly caption generation