Ideogram 4 GGUF Workflow for ComfyUI

A simple workflow for running Ideogram 4 in GGUF quantized format on ComfyUI, optimized for systems with limited VRAM.

Tested on RTX 3060 12GB with 16GB RAM (and 32GB Virtual RAM)

⚠️ Don't forget to update your ComfyUI first!

Available Versions:

- v1.0 - Standard (2 UNET GGUF models) - Default traditional workflow

- v2.0 - TurboTime (Single UNET GGUF) - ⚡ Faster, better quality, super low VRAM

---

🆕 What's New in v2.0

- ⚡ Optimized workflow: Removed redundant nodes (Dual Model CFG Guider, CFG Override)

- 🚀 TurboTime LoRA support: 2-step generation with CFG=0.0

- 💾 Low VRAM mode: Works on 8GB+ GPUs with Q4_0 quantization

- 🎯 Cleaner structure: Added BasicGuider for proper CFG handling

- 📝 Faster generation: ~1-2 seconds per image on modern GPUs

---

📦 Required Models

1. UNET Models → ComfyUI/models/unet/

Download unet files from HuggingFace (Recommended):

- Repository: https://huggingface.co/molbal/ideogram-4-gguf

- Main Model (choose desired quantized version):

- ideogram4-transformer-q4_0.gguf (5.64 GB) - ⭐ Best for low VRAM

- ideogram4-transformer-q4_1.gguf (6.21 GB)

- ideogram4-transformer-q5_0.gguf (6.77 GB)

- ideogram4-transformer-q5_1.gguf (7.33 GB)

- ideogram4-transformer-q8_0.gguf (10.1 GB) - Best quality

- Unconditional Model (not needed in workflow v2.0):

- ideogram4-unconditional_transformer-q4_0.gguf (5.64 GB)

- ideogram4-unconditional_transformer-q4_1.gguf (6.21 GB)

- ideogram4-unconditional_transformer-q5_0.gguf (6.77 GB)

- ideogram4-unconditional_transformer-q5_1.gguf (7.33 GB)

- ideogram4-unconditional_transformer-q8_0.gguf (10.1 GB)

💡 Note: The repository includes inference speed and memory usage charts to help you choose the best quantization for your system.

Alternative link for unet files (Civitai - Q4_0 only):

- Download from: https://civarchive.com/models/2681714/ideogram-4-gguf

- Note: File names may differ slightly from HuggingFace

2. Text Encoder → ComfyUI/models/clip/

- Qwen3-8B-Q4_K_M.gguf

- Source: https://huggingface.co/Qwen/Qwen3-8B-GGUF

3. Prompt Enhancer (optional) → ComfyUI/models/clip/

- gemma4_e4b_it_fp8_scaled.safetensors

- Source: https://huggingface.co/Comfy-Org/gemma-4

- Automatically converts natural language to JSON format

4. VAE → ComfyUI/models/vae/

- flux2-vae.safetensors

- Download: https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/vae/flux2-vae.safetensors

---

⚙️ Custom Node Required

Single GGUF custom node by molbal (for both UNET and CLIP)

- Repository: [molbal/ComfyUI-GGUF](https://github.com/molbal/ComfyUI-GGUF)

- Install: git clone https://github.com/molbal/ComfyUI-GGUF.git

- Nodes: UnetLoaderGGUF + GGUFCLIPLoader

⚠️ Important: This is a fork of city96/ComfyUI-GGUF with Ideogram 4 support. If you have city96 version installed, remove it first (same folder name causes conflicts).

Installation steps:

1. Stop ComfyUI

2. Delete the entire ComfyUI-GGUF folder in custom_nodes (not just the files inside it)

3. Clone: git clone https://github.com/molbal/ComfyUI-GGUF.git

4. Restart ComfyUI

💡 Optional: If you encounter any import errors, install requirements:

pip install -r ComfyUI-GGUF/requirements.txt

✅ That's it! This single extension handles both UNET and CLIP loading.

---

📊 Choosing the Right Quantization

The HuggingFace repository includes helpful charts for:

- Inference Speed: How fast each quant generates images

- Memory Usage: How much VRAM each quant requires

General Guidelines:

- Q4_0: Lowest VRAM (~5.5 GB), fastest, good quality

- Q5_0/Q5_1: Balanced VRAM (~6.8-7.3 GB), better quality

- Q8_0: Highest VRAM (~10 GB), best quality, slower

💡 Check the charts in the repository for detailed comparisons!

---

🗒️ JSON Prompting (Optional)

Ideogram 4 works best with structured JSON prompts. This workflow uses Gemma 4 to automatically convert your natural language prompts into JSON format for:

- ✅ Better text rendering and typography

- ✅ More accurate composition control

- ✅ Consistent results across generations

Just write a simple prompt like "A poster for a coffee shop" and let Gemma 4 handle the rest!

---

💡 Quick Tips

- For TurboTime LoRA (v2.0): Use only ideogram4-transformer-q4_0.gguf (no unconditional model needed)
- TurboTime LoRA: Download from: https://huggingface.co/ostris/ideogram_4_turbotime_lora

- For standard mode (v1.0): Use both main and unconditional models

- CFG Settings: 0.0 for TurboTime, 7.0 for standard mode

- Scheduler: mu=0.5, std=1.75 for TurboTime | mu=0.0, std=1.5 for standard

- Quantization: Q4_0 for lowest VRAM, Q8_0 for best quality

---

💻 System Requirements

Minimum:

- GPU: 8GB+ VRAM (12GB recommended)

- RAM: 16GB+ (32GB recommended)

- Virtual Memory (Pagefile): 32GB+ ⚠️

Important: If you experience crashes during VAE decode, make sure your Windows Virtual Memory (Pagefile) is set to at least 32GB:

- Settings → System → About → Advanced system settings → Performance → Settings → Advanced → Virtual memory

- Set to 32GB or more

This is especially important for systems with 16GB RAM running GGUF models!

---

💡 TurboTime LoRA Tips

Steps Guide (with TurboTime Lora):

- 2 steps: Fastest (~15s), good for quick tests and drafts

- 4-6 steps: Good balance (~30s), decent quality for iterations

- 12-20 steps: ⭐ Recommended! Best quality/speed ratio (~1.5 min)

- 30-50 steps: Maximum quality (~3-4 min), all fine details rendered

- 50+ steps: Marginal improvement over 30, not worth the extra time

Unlike other Turbo LoRAs (Flux/SDXL), increasing steps does NOT overcook the image with TurboTime Lora! You can safely use 20+ steps for better anatomy and details.

Pro Tip: If your prompt includes fine details (flowers, jewelry, text), use at least 20 steps for them to appear correctly.

⚠️ Important: Custom Node Conflicts

If you have multiple GGUF-related custom nodes installed, they can conflict and cause "Unknown model architecture" errors.

Common conflicts:

- ComfyUI-GGUF_KREA-2 - Causes GGUF loader to use wrong code

- ComfyUI-GGUF (city96 version) - Conflicts with molbal version

- Any other GGUF loader custom nodes

Solution:

1. Remove ALL conflicting custom nodes:

- Delete ComfyUI-GGUF_KREA-2 folder

- Delete ComfyUI-GGUF folder (if city96 version)

- Remove any other GGUF loaders

2. Install only molbal/ComfyUI-GGUF:

cd custom_nodes

git clone https://github.com/molbal/ComfyUI-GGUF.git

⚠️ Custom Node Conflicts - Visual Guide

The image below shows when the GGUF loader works vs fails with Ideogram 4:

✅ Only molbal/ComfyUI-GGUF installed → Works perfectly

❌ city96 version installed → "Unknown model architecture" error

❌ ComfyUI-GGUF_KREA-2 installed → Conflicts with loader

❌ Multiple GGUF nodes installed → Wrong loader selected

Solution: Remove all other GGUF custom nodes and keep only molbal/ComfyUI-GGUF

🔧 Advanced: Managing Multiple GGUF Custom Nodes

If you work with multiple GGUF models (Ideogram 4, KREA-2, etc.) and need different custom node versions:

Quick Switch Method:

1. Install each custom node version

2. Zip each folder with a descriptive name:

- ComfyUI-GGUF-molbal.zip (for Ideogram 4)

- ComfyUI-GGUF-city96.zip (for other models)

- ComfyUI-GGUF_KREA-2.zip (for KREA-2)

3. Delete the extracted folders

4. When you need a specific version, just extract the corresponding zip

5. Restart ComfyUI

This avoids conflicts and makes switching instant without reinstalling

🙏 Credits

- Ideogram 4 Model: [Ideogram AI] https://ideogram.ai/

- GGUF Conversion & Custom Node: [molbal] https://github.com/molbal/ComfyUI-GGUF

- TurboTime LoRA: [ostris] https://huggingface.co/ostris/ideogram_4_turbotime_lora

- Workflow: [dvdufo] https://civarchive.com/user/dvdufo