I think ideogram-ai 4 is 18.6B + 8B mode not 9.3B model.
Unconditional model double size.
The ideogram4_unconditional_fp8_scaled.safetensors model is not a separate image generator. It is a required auxiliary component that the main generator relies on. It plays a key role in controlling the final image quality and prompt adherence.
🧠 Why is an "unconditional" model needed?
To understand this, we need to look at the generation process. Most modern models, including Ideogram 4.0, use a technique called Classifier-Free Guidance (CFG).
The essence of CFG is comparing two generation paths:
Conditional generation: The model creates an image trying to follow your text prompt as closely as possible.
Unconditional generation: The model creates an image “on its own”, without any text guidance.
The final result is produced by “pulling” the unconditional output toward the conditional one. The strength of this pull is controlled by the CFG scale parameter.
💡 Why Ideogram 4.0 does this differently
In many models, the unconditional branch is handled by the same neural network, just with an empty prompt. But Ideogram 4.0 uses an asymmetric architecture – the conditional and unconditional paths are separate and can be optimised independently.
That’s why the diffusion_models folder contains two files:
ideogram4_fp8_scaled.safetensors – the main (conditional) model. It generates images according to your prompt. It understands your JSON, places objects, and builds the composition.
ideogram4_unconditional_fp8_scaled.safetensors – the unconditional model. Its job is to generate a “raw” image without following the prompt. It acts as an anchor that the main model pulls away from to strengthen alignment with your request.
🤝 How they work together
The system runs both models simultaneously: the conditional one (with your prompt) and the unconditional one (with no prompt).
At each step, it computes the difference between their outputs. The larger the difference, the more strongly the main model follows your prompt.
You control this effect with the CFG scale (also called guidance_scale).
🎯 Practical benefits
Separating the two models gives developers and advanced users more control:
Independent tuning – you can adjust each branch to find the ideal balance between prompt adherence and image quality.
CFG scheduling – you can change the guidance strength during generation. For example, follow the prompt strictly in early steps to set composition, then reduce guidance later to refine details without oversaturation.
Quality presets – this flexibility allows the Ideogram 4.0 authors to provide presets like V4_QUALITY_48, V4_DEFAULT_20, and V4_TURBO_12, which differ in step count and CFG dynamics.
⚠️ Can you run without it?
Technically, the model might run without the unconditional component (some sources mention this), but it is strongly discouraged. Without ideogram4_unconditional_fp8_scaled, you lose the CFG mechanism. This will significantly degrade image quality – results will likely be blurry, poorly structured, and weakly aligned with your prompt.
For best results, keep both files in ComfyUI/models/diffusion_models/ and use them together.