import json, os, torch, requests
from diffusers import Ideogram4Pipeline

pipe = Ideogram4Pipeline.from_pretrained(
    "ideogram-ai/ideogram-4-nf4",
    torch_dtype=torch.bfloat16,
    token=os.environ["HF_TOKEN"], # or: token="hf_xxxxxxxxx", token is needed as the repo is gated
).to("cuda")

# Expand the prompt into a structured JSON caption with Ideogram's free hosted magic-prompt API.
# Get a key at https://developer.ideogram.ai/  (set IDEOGRAM_API_KEY).
resp = requests.post(
    "https://api.ideogram.ai/v1/ideogram-v4/magic-prompt",
    headers={"Api-Key": "your_ideogram_api_key"},
    json={"text_prompt": "a ginger cat wearing a tiny wizard hat reading a spellbook", "aspect_ratio": "1x1"},
).json()
caption = json.dumps(resp["json_prompt"])  # already a structured JSON caption

# Pass the caption straight to the pipeline (no prompt_upsampling — it's already upsampled).
image = pipe(
    caption,
    height=1024,  # model supports up to 2048
    width=1024,   # model supports up to 2048
    generator=torch.Generator("cuda").manual_seed(0),
).images[0]
image.save("ideogram4.png")

import torch
from diffusers import Ideogram4Pipeline

pipe = Ideogram4Pipeline.from_pretrained(
    "ideogram-ai/ideogram-4-nf4",
    torch_dtype=torch.bfloat16,
    token=os.environ["HF_TOKEN"], # or: token="hf_xxxxxxxxx", token is needed as the repo is gated
).to("cuda")

# prompt_upsampling=True rewrites the prompt into Ideogram's structured JSON caption locally on-device.
image = pipe(
    "a ginger cat wearing a tiny wizard hat reading a spellbook",
    height=1024,  # model supports up to 2048
    width=1024,   # model supports up to 2048
    prompt_upsampling=True,
    generator=torch.Generator("cuda").manual_seed(0),
).images[0]
image.save("ideogram4.png")