ideogram-ultra

Download the conditional and unconditional FP8 diffusion models, Qwen3-VL text encoder, and Flux 2 VAE from Hugging Face, then wire UNETLoader nodes through ModelSamplingAuraFlow into a DualModelGuider. Load the text encoder with CLIPLoader type ideogram4 and sample with SamplerCustomAdvanced.

Ideogram 4 expects a structured JSON caption with high_level_description and compositional_deconstruction containing background and elements arrays. Each element uses a bbox in [top, left, bottom, right] format with values 0 to 1000, and text elements specify the exact string to render.

The node re-serializes its elements_data widget from an in-browser array every time the graph is queued, so external widget edits are overwritten. Wire a PrimitiveStringMultiline node with the full caption JSON into the import_json input and set import_mode to always.

No, this is a fully local open-weights pipeline with no API key or network call at generation time. The weights run on your GPU via standard UNETLoader, CLIPLoader, and VAELoader nodes, separate from ComfyUI's paid hosted Ideogram partner nodes.

Loading both ~13.8 GB UNETs plus the Qwen3-VL encoder is heavy on 24 GB GPUs; official docs cite 16 GB minimum for FP8 with model swapping. Clear VRAM before switching model families, and use --lowvram or lower resolution if you run out of memory.