Optimized Workflow for Ideogram 4 with Advanced Dual Guidance & CFGOverride

Another new, refined and enhanced workflow by Experimental_dAIver.

This workflow brings the impressive image quality, prompt adherence and cinematic strengths of Ideogram 4 directly into your local ComfyUI setup. It features dedicated loading of both the main and unconditional models, a sophisticated DualModelGuider, the specialized Ideogram4Scheduler, high-quality text encoding via Qwen3-VL, Flux VAE decoding and a complete integrated SeedVR2 Video Upscaler pipeline.

The highlight of this version is the CFGOverride node, which elegantly exploits Ideogram 4’s asymmetric classifier-free guidance: gentle, organic guidance in the early steps for natural composition, followed by a targeted raise in CFG during the final phase for exceptional detail, sharpness and prompt fidelity.

Key features in v1.5:

• Full native support for Ideogram 4 fp8_scaled together with its unconditional variant using DualModelGuider + ConditioningZeroOut

• CFGOverride node – gentle guidance (CFG ~1) in the first 70–85 % of steps for natural structure, then raised CFG (3–7) in the final phase for precise detail refinement

• Custom Ideogram4Scheduler with quality presets (Turbo / Default / Quality) and dynamic mu/std control via JSON-driven MATH subgraphs

• Powerful text understanding powered by Qwen3-VL 8B fp8_scaled (with optional Gemma4 support)

• High-quality decoding via Flux VAE

• ResolutionSelector with automatic multiple-of-16 calculation and dynamic width/height routing

• Modular architecture with a full-featured SeedVR2 Video Upscaler subgraph for exceptional 4K+ stills and short video sequences

• Smart VRAM optimization helpers (easy cleanGpuUsed) and offload options

• Includes a rich, highly detailed structured JSON example prompt demonstrating conceptual food photography with precise material rendering and dramatic lighting

Required Custom Nodes

• ComfyUI-easy-use – Provides easy cleanGpuUsed, easy float, easy compare and other helper nodes. Install via ComfyUI Manager.

• ComfyUI-KJNodes – Required for the LazySwitchKJ node in the upscaler section.

• ComfyUI-SeedVR2_VideoUpscaler – https://github.com/ainvfx/ComfyUI-SeedVR2_VideoUpscaler – Powers the high-quality upscaling and video features.

Important: Update your ComfyUI to the latest version first, at least v0.24. Ideogram 4 support, Ideogram4Scheduler, DualModelGuider and CFGOverride are part of recent core updates.

Models & Downloads (exact paths)

The following list explains the base models I am most frequently using with this workflow. The list as well explains where to put each file after you downloaded it.

1. VAE:

• File: flux2-vae.safetensors

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

• Target folder: ComfyUI/models/vae/

2. Main Diffusion Model (Ideogram 4 fp8 scaled):

• File: ideogram4_fp8_scaled.safetensors

• Download: https://huggingface.co/Comfy-Org/Ideogram-4/resolve/main/diffusion_models/ideogram4_fp8_scaled.safetensors

• Target folder: ComfyUI/models/diffusion_models/

3. Unconditional Diffusion Model (Ideogram 4):

• File: ideogram4_unconditional_fp8_scaled.safetensors

• Download: https://huggingface.co/Comfy-Org/Ideogram-4/resolve/main/diffusion_models/ideogram4_unconditional_fp8_scaled.safetensors

• Target folder: ComfyUI/models/diffusion_models/

4. Text Encoder (Qwen3-VL 8B):

• File: qwen3vl_8b_fp8_scaled.safetensors

• Download: https://huggingface.co/Comfy-Org/Qwen3-VL/resolve/main/text_encoders/qwen3vl_8b_fp8_scaled.safetensors

• Target folder: ComfyUI/models/text_encoders/

5. Text Encoder (Gemma4 E4B IT – optional):

• File: gemma4_e4b_it_fp8_scaled.safetensors

• Download: https://huggingface.co/Comfy-Org/gemma-4/resolve/main/text_encoders/gemma4_e4b_it_fp8_scaled.safetensors

• Target folder: ComfyUI/models/text_encoders/

Model Storage Location

📂 ComfyUI/
├── 📂 models/
│   ├── 📂 vae/
│   │   └── flux2-vae.safetensors
│   ├── 📂 diffusion_models/
│   │   ├── ideogram4_fp8_scaled.safetensors
│   │   └── ideogram4_unconditional_fp8_scaled.safetensors
│   └── 📂 text_encoders/
│       ├── qwen3vl_8b_fp8_scaled.safetensors
│       └── gemma4_e4b_it_fp8_scaled.safetensors

Recommended Settings

Sampler: Euler (or Euler ancestral / res_multistep)

Steps: Turbo (12), Default (20) or Quality (48) – selectable via the Quality Selector

Base Resolution: Freely selectable via ResolutionSelector (automatically adjusted to multiple of 16)

CFG Scale (DualModelGuider): 1.0 (base)

CFGOverride: Start at CFG 1, raise to 3–7 from ~70–85 % of steps onward (start_percent 0.7–0.9) Ideogram4Scheduler mu / std: Automatically set by the quality preset (Turbo: mu 0.5 / std 1.75, Default: mu 0.0 / std 1.75, Quality: mu 0.0 / std 1.5)

How to use the workflow

1. Update ComfyUI to the latest version that includes Ideogram 4 support.

2. Install the required custom nodes listed above via ComfyUI Manager or by cloning the repositories.

3. Download all models and place them exactly in the folders specified above.

4. Load the JSON workflow file into ComfyUI.

5. Prompting: The large “CLIP Text Encode (Positive Prompt)” node accepts structured JSON prompts (the model was trained on this format). The workflow includes a detailed internal guide explaining three ways to create prompts:

   • Raw natural language or structured JSON (recommended for maximum control)

   • LLM-assisted prompting via the integrated Gemma 4 Prompt Builder subgraph

   • Direct editing of the high_level_description, style_description (with aesthetics, lighting, color_palette) and compositional_deconstruction (with normalized bounding boxes and per-object color palettes)

   The included example prompt (hyper-realistic conceptual food photography of an ice cream cone with Earth, Jupiter and Uranus scoops) demonstrates the model’s excellent capabilities in precise material rendering, dramatic cinematic lighting and creative conceptual twists.

6. Quality & Resolution: Use the Quality Selector (Turbo / Default / Quality) to automatically set steps, mu and std. Use the ResolutionSelector to choose aspect ratio and megapixels – width and height are automatically calculated and routed correctly.

7. CFGOverride (the secret weapon): The node keeps guidance gentle in the early steps (natural composition and proportions) and raises CFG in the final phase for razor-sharp detail, texture and prompt adherence. This technique is particularly effective with Ideogram 4’s asymmetric guidance.

8. Queue the prompt. The upper section generates the base image using the powerful dual guidance + CFGOverride system.

9. Optional high-quality upscaling: The lower SeedVR2 section can be fed with the generated image to produce stunning high-resolution stills or short video sequences.

10. Review results in the PreviewImage node and save with the extended SaveImageExtended node (includes full metadata).

This is a clean, powerful and future-proof base workflow for Ideogram 4. You can bypass or mute the upscaler group entirely if you only need the base generation. It runs comfortably even on mid-range hardware.

I’ve refined this version further with the CFGOverride technique and dynamic quality presets. The combination of gentle early guidance and targeted late-stage refinement, together with the structured JSON prompting approach, really lets Ideogram 4 shine — especially in conceptual, atmospheric and highly detailed scenes.

Huge thanks to Comfy-Org for the excellent Ideogram 4 implementation and to realrebelai for the original base workflow that I built upon and optimized.

If you have questions, feedback or want to share your results – I’m always happy to hear from you in the comments. Enjoy the workflow and happy creating!