Ideogram 4 Settings

Ideogram 4 is the interesting one of the 2026 open-weights wave: the text-in-image benchmark that finally shipped weights, but non-commercial, quantized-only, with a safety refusal image trained straight into the model. Locally it's also structurally unusual - it runs dual-model guidance, where a separate unconditional diffusion model drives the CFG pass. AIOIdeogram4Settings is the pack's control surface for that. It outputs one AIO_MODEL_SETTINGS dict, which you connect to the model_settings socket of AIO Image Generate with model_type set to ideogram4.

The inputs that matter

• preset - Default, Quality, Turbo, or Workflow Compatible. The first three use the official Ideogram 4 sigma schedules; Workflow Compatible uses the saved workflow's simple-scheduler path, which matters if you're porting an existing Ideogram workflow that isn't on the official presets. The profile default is 20 steps with the official scheduler.
• unconditional_model - path to the separate unconditional diffusion model (diffusion_models/ideogram4/ideogram4_unconditional_fp8_scaled.safetensors). This is the second model in the dual-model guider, and yes, you need both checkpoints on disk.
• dual_cfg - the dual-model CFG passed to the guider (default 7.0). This is the "real" guidance knob for Ideogram 4; the main node's cfg is separate and defaults to 0 (profile default).
• cfg_override_enabled / cfg_override / cfg_override_start_percent / cfg_override_end_percent - a final CFG override window applied late in sampling (default: override at CFG 3.0 from 70% to 100%). This is the "back off guidance near the end so details don't overcook" pattern. Leave it on unless you have a reason not to.
• sampling_shift - the AuraFlow model sampling shift applied to the conditional model (default 5.0).
• run_unconditional_model - runs (default true) the separate unconditional model for dual-model guidance. Turn this off for turbo-LoRA workflows that should skip the second model and run the guider with the conditional model only - that's also how you avoid loading a second 8GB checkpoint when you don't need it.
• The shared performance block: precision_policy, attention_mode, torch_compile_mode/torch_compile_backend, performance_apply_timing. Same semantics as the rest of the pack.

There's also one optional socket worth knowing: prompt_builder accepts an AIO_IDEOGRAM4_PROMPT from the Ideogram 4 Prompt Builder node. When connected, the structured JSON prompt replaces the main node's positive_prompt, and the builder's resolved dimensions override the main node's size controls. That's not optional in spirit - Ideogram 4 is a JSON-prompting model, and the structured format is the documented way to get text and layout right (the community's finding, per the KB's ideogram-4 panel, is that structured prompting even clears the refusal filter).

Wiring and install

The node lives in the AIO pack, installed once:

cd ComfyUI/custom_nodes
git clone https://github.com/helto4real/comfyui-all-on-one-image-generation-node
cd comfyui-all-on-one-image-generation-node
python -m pip install -r requirements.txt

Or ComfyUI Manager → search "AIO Image Generate" → install → restart. Everything appears under AIO/Image. Then in AIO Image Generate: model_type=ideogram4, pick the conditional diffusion model, the Qwen3-VL text encoder, and the FLUX.2 VAE, and make sure the unconditional model path here resolves.

Gotchas

• It needs four files, not two. Conditional model + unconditional model + Qwen3-VL encoder + FLUX.2 VAE, per the Comfy-Org Ideogram 4 packaging. Miss the unconditional one and dual-model guidance breaks.
• Output dimensions are strict. Multiples of 16, 256–2048 per side, aspect ratio no wider than 6:1. Set multiple value to 16 on the main node.
• No references, no GGUF. Ideogram 4 in this adapter is text-to-image and AIO Inpaint only - reference images, legacy masks and GGUF files aren't implemented for it. Don't fight it.
• Licence first. Non-commercial agreement, termination-on-notice. If that matters to your project, that's a decision to make before you spend an afternoon on it.

One dial, honestly labeled: dual_cfg is the Ideogram way, cfg_override is the finisher. Learn those two and the rest of the node is set-and-forget.