toobusy Ideogram4 T2I (local model)

The name is a lie, in the good way: toobusy Ideogram4 T2I doesn't call any API and needs no key. It's a full local text-to-image sampler for the Ideogram 4 weights, folded from the whole load→encode→sample→decode chain into one node. Ideogram 4 is the model whose text-in-image quality made it the benchmark for years, and whose June 2026 open-weights release made it runnable locally - with the catch that it's JSON-only prompting and it ships quantized. This node exists to make that workflow not miserable.

How it works

It runs the complete local chain: loads the Ideogram 4 diffusion model (two of them, actually - a conditional and an unconditional checkpoint, because Ideogram's CFG pass drops text tokens entirely on the uncond side), the Qwen3-VL text encoder, and the VAE; encodes your structured prompt; samples; decodes. The prompt input expects the Ideogram JSON schema - the same shape the Layout Builder produces, which is why the two nodes are a pair. quality gives you the vendor's recommended presets (Quality / Default / Turbo / Custom) that set sampler, CFG, and schedule for you, and steps = 0 means "use the preset's step count." If KJNodes is installed, use_sage_attention patches just this node's models with SageAttention for a speedup.

The inputs that matter

• model_name / unconditional_model_name - the two checkpoints, defaulting to ideogram4_fp8_scaled.safetensors for both. The fuzzy auto-detection excludes "uncond"-named files from the conditional slot so the two don't grab the same weight.
• clip_name - qwen3vl_8b_fp8_scaled.safetensors; vae_name - flux2-vae.safetensors.
• prompt - paste the Layout Builder's ideogram_json here. This is the intended flow.
• quality - Turbo is a great default for iterating; step down to Quality for finals.
• ratio_preset + megapixels - sizing, with width/height optional overrides. The tooltip tells you the intended wiring: "Connect Layout Builder's width here to use its resolution."
• cfg (7), sampler_name (res_multistep), seed - overridable; the quality preset overrides them when steps is 0.
• mu / std / cfg_override* - only used in Custom quality; ignore until you're tuning.
• The lora_N_enable/name/strength slots (up to 5) - Ideogram 4's LoRA scene is thin (Krea 2 ate its ecosystem), but the slots are there.

Outputs and where they go

• image - the render. latent - for a second pass.
• width / height - actual geometry, for downstream compositing.

Install and gotchas

Standard pack install. The model files are not bundled - you need the local Ideogram 4 checkpoints in models/diffusion_models/, the Qwen3-VL text encoder, and the Flux2 VAE. And critically, your ComfyUI build must have Ideogram4 support nodes and the ideogram4 CLIP type; on an unsupported build the node fails at runtime with a clear error - the README's FAQ is blunt about this: it's a local-model node, not the web API.

Two things worth knowing. First, the license: Ideogram 4's weights are non-commercial, quantized-only, and Ideogram trained a refusal image into the model. Second, its real strength after the dust settled turned out to be character and IP recall plus layout control - which is exactly why this pack pairs it with the Layout Builder. Don't use it as your general-purpose model; use it when you need the layout and text control it was built for.