First look at Ideogram4.0 ! A new industry leader or just a marketing gimmick? (Patreon)

Ideogram 4.0 is Ideogram's first open weight model. Its main features

Precise text rendering and typesetting: Xuancheng boasts the industry's strongest in-image text generation capabilities. It not only supports multiple languages but also can accurately and completely embed long text in images, perfectly adapting to posters, covers, and social media creatives.

Structured JSON Prompt: The world's first model trained on structured JSON prompt data. Users can precisely control the position, layout, and color of text and objects by accurately specifying bounding boxes, achieving pixel-level composition.

Powerful open-source base model: With 9.3 billion parameters (9.3B), it uses the single-stream diffusion Transformer (DiT) architecture, is trained from scratch, and is fully open-source, allowing developers to download the weights and run it locally.

Expert-level design sense: The model is specifically optimized for poster, marketing graphic, logo, and advertising design, ranking among the top in human blind tests (DesignArena), and has a deep understanding of color schemes, fonts, and visual hierarchy.

ComfyUI has implemented Day0 adaptation for it, indicating that it still has a certain level of prestige. ComfyUI version v0.24.0 and above can support it. There are also templates in the template library that support Ideogram v4.

In the ComfyUI template, there are two main sections: the left side is the regular text-to-image section, and the right side is a section for generating JSON structured prompts.

The template already includes a default JSON prompt. Try to first get the text-to-image process up and running. The model download address is

https://www.modelscope.cn/models/Comfy-Org/Ideogram-4

This model requires downloading two UNets, a main UNet and an unconditional UNet. The combined size of the two UNets is still relatively large, even the nvfp4 quantized version is 11GB. Text Encoder uses Qwen3-VL-8B. ComfyUI only provides the fp8 quantized version. In fact, the smaller quantized gguf version can be used, and the download address is

https://www.modelscope.cn/models/Qwen/Qwen3-VL-8B-Instruct-GGUF

Remember to also change ClipLoader to the GGUF version. It should be noted that the template provided in v0.24.0 has incorrect sampler parameters.The latest advice from ComfyUI developers is that the start percent in CFG Override should be set to 0.7, and the K sampler should use euler

Image drawn using the default prompt of the template

The picture has a strong grainy texture, but this is required by the prompt, which asks for "heavy film grain". To use this model, you also need to write a good JSON prompt. It's impossible to write it yourself, as that's too complicated. Besides using the prompt descriptor in the ComfyUI template, which is generated by gemma 4 4b, you can also use external larger models to generate it, such as Doubao and Gemini. JSON generator for web AI

The following is a prompt for generating JSON using Gemini Flash: "Create a Chinese ink painting for 'Night Mooring at Maple Bridge'".

"A four-panel cartoon introducing how to make scrambled eggs with tomatoes. Step prompts are written in Chinese"

The rendering of Chinese doesn't seem to be as strong as claimed, and it's obvious that the second and third steps are confused. The JSON prompt requires frying the eggs in the second step and stir-frying the tomatoes in the third step. Moreover, this image has been changed to Default mode, yet it's just as messy as when rendered in Turbo mode. "Under the night sky, a beautiful woman is soaking in a hot spring."

"In the snowfield forest, there is a magical weather bottle."

It is worth mentioning that it has built-in support for partitioned prompts

Divide the canvas into separate sections and fill each section with distinct prompts. Logographic models can identify the content within these segmented areas, generate corresponding artwork for every part, and stitch all generated pieces into a single complete image. To a certain extent, this approach fully replaces conventional inpainting techniques and even delivers partial functionality of dedicated image-editing models. This raises a question: do you think it could evolve into a new paradigm for AI image generation?

This feature is very practical in advertising photography. For example, the perfume in a woman's hand can be replaced with any product.

Judging from the final rendering effect of ideogram4, the skin rendering is still relatively delicate. It's just that a beautiful woman aged 20-30 has been paired with a pair of 60-year-old hands. Moreover, text rendering is less likely to go wrong only in the Default crotch. After testing, the rendering of this model in other languages is not as good as advertised. Instead, the generated images have good details and a strong sense of realism. However, the drawbacks are also obvious. First of all, image generation is slow.Turbo mode also takes 12 steps to generate an image, and two UNets plus Clip cannot fit in 16GB of video memory. Each time an image is generated, Clip has to be reloaded and then kicked out of video memory after use. Secondly, modifying the prompt in JSON is too cumbersome. Although it offers stronger controllability, for general use, traditional plain text prompts are more convenient and easier to modify.

OK! That wraps up all the content for the initial test. Hope you enjoy this test format. What do you think of Ideogram 4? What score would you give it on a scale of 1 to 10? Leave your thoughts down below.

🫡 May the AI-power be with you, see you soon !🫡