Ideogram 4 Studio Launcher

Local Pinokio launcher for ideogram-ai/ideogram-4-nf4, Ideogram's open-weights text-to-image model, wrapped in a Studio UI for structured JSON prompting.

Model Details

• Params: 9.3B — selectable nf4 (4-bit, least VRAM) or fp8 (8-bit, higher fidelity); only one is resident at a time
• Architecture: Single-stream Diffusion Transformer (DiT), flow-matching, 34 layers
• Text encoder: Qwen3-VL-8B-Instruct (vision-language model)
• Resolution: any multiple-of-16 from 256 to 2048 per side (native 2k)
• License: Ideogram 4 Non-Commercial

Features

• Structured JSON prompting (the format the model was trained on) via a visual Studio editor
• Bounding-box layout control — drag boxes on the canvas to place objects and text
• Color-palette conditioning (global + per-element)
• Prompt upsampling: Ideogram's free hosted magic-prompt (remote) or local Qwen graft
• Best-in-class in-image text rendering
• Speed/quality modes: Turbo (12 steps), Default (20 steps), Quality (48 steps)

Requirements

• NVIDIA GPU with CUDA support
• Gated model — accept the license at the model page and provide an HF token

Installation

1. Accept the gate at ideogram-ai/ideogram-4-nf4 (and ideogram-4-fp8 if you want the fp8 variant — it is gated separately)
2. Click Install in the Pinokio UI
3. Enter your HF token (required — the weights are gated)
4. Optionally enter an Ideogram API key (developer.ideogram.ai) to enable remote prompt upsampling

Usage

1. Click Start to launch the web UI (FastAPI + a static HTML/JS frontend)
2. Pick a model variant and click ⬇️ Download / Load model
3. Type a prompt and click ✨ Generate prompt to draft a structured JSON caption
4. Refine it in the visual or JSON editor (move/resize/describe elements, set colors)
5. Click 🎨 Generate image — the result lands on the canvas for further editing

API Access

Python (diffusers)

import os, torch
from diffusers import Ideogram4Pipeline

pipe = Ideogram4Pipeline.from_pretrained(
    "ideogram-ai/ideogram-4-nf4",
    torch_dtype=torch.bfloat16,
    token=os.environ["HF_TOKEN"],
).to("cuda")

image = pipe(
    "a ginger cat wearing a tiny wizard hat reading a spellbook",
    height=1024, width=1024,
    generator=torch.Generator("cuda").manual_seed(0),
).images[0]
image.save("ideogram4.png")

License

Ideogram 4 Non-Commercial — see ideogram-ai/ideogram-4-nf4 for details.