Following the usual style of the Basic Workflow series, this workflow focuses on simplicity and practicality. No unnecessary custom nodes are required.
The interface also follows the familiar layout used throughout the series, so existing users should feel right at home.
Before using this workflow, make sure your ComfyUI installation is updated to the latest version. Older versions may not support Ideogram 4. All testing was done on ComfyUI 0.25.0.
Here are a few quick notes:
## I. Models
1. Model Versions
I tested both the official FP8 and NVFP4 models released by ComfyUI. Other variants based on these models should work as well.
The NVFP4 model uses slightly less VRAM, but is actually slower than the FP8 version. It also shows a noticeable drop in image quality. Unless you're limited by available VRAM, I would recommend sticking with FP8.
2. Single Main Model Loading
By default, the workflow follows the official setup and loads both the ideogram4 and ideogram4_unconditional models.
If you'd like to load only the main ideogram4 model, open the Models Subgraph and Bypass the second Load Diffusion Model node to disable loading the unconditional model.
This noticeably reduces VRAM usage, speeds up generation, and still produces reasonably good results.
II. Options
I've simplified the official workflow as much as possible, removing nearly every non-essential node and consolidating the important settings into the Options panel.
Parameters such as steps, width, height, sampler, and seed follow the same layout and usage found in other Basic Workflow releases.
Additional Ideogram 4 specific settings, including mu and std, have also been added.
III. LoRA Support
This workflow continues to use the convenient and user-friendly Power Lora Loader (rgthree) node.
When paired with Turbo-related LoRAs available online, it's possible to generate images in as few as 4 or even 2 steps while maintaining surprisingly good results.
IV. Prompt
The Prompt section currently remains intentionally simple.
Traditional prompts already produce solid results, but for better precision and control, I recommend following Ideogram's official JSON prompt format.
Popular LLMs such as ChatGPT and Gemini can generate JSON prompts very well.
If needed, I may further improve this section in future updates, possibly by adding dedicated JSON processing nodes.
Download link
