Image reference limits: GPT Image 2.5 takes 16, Ideogram 4.5 takes 5

On Sume, openai/gpt-image-2.5 and openai/gpt-image-2.5-sunburst accept up to 16 input_references, while ideogram/ideogram-v4.5 takes 5 in total: with references it edits the first image and uses up to 4 more as references. If your job is combining many assets, such as a model plus several garments, 16 gives room. If it is editing one picture with a few style cues, 5 is plenty.

What the docs say

The Image API page states both limits: Flare and Sunburst support text-to-image, up to 16 image references and an optional mask_url; Ideogram 4.5 generates from text without references, and with them edits the first image. An edit without aspect_ratio keeps the source image's shape. Other families list their own input_references range, and a model with {"min": 0, "max": 0} is text-to-image only and rejects references.

OpenAI's image generation guide describes the same capability in its own words: one or more images as references for new creations, plus mask-based editing.

Order matters

The prompt should name references by position: image 1 is the person, images 2 to 4 are garments. On Ideogram the first position is special because it is the image being edited, so put the picture you want changed there and say what the other four are for. Keep the list short: send only the references the prompt names.

{
  "model": "ideogram/ideogram-v4.5",
  "prompt": "Edit image 1: swap the mug for the teapot in image 2. Keep the table and light.",
  "input_references": [
    {"type": "image_url", "image_url": {"url": "https://example.com/table.jpg"}},
    {"type": "image_url", "image_url": {"url": "https://example.com/teapot.jpg"}}
  ],
  "quality": "medium"
}

Rules for every reference

Reference URLs must be public HTTPS. Localhost, private-network and non-HTTPS URLs are rejected before submission, as the Image API docs say, and the Media inputs page covers URL rules for uploads. Ask a text-only model for a reference and you get a rejection, not a quiet ignore.

Picking between them

• Many assets into one scene (outfits, kits, collages): GPT Image 2.5, up to 16.
• One photo to retouch with a couple of style cues: Ideogram 4.5 keeps the source shape by default.
• Local fix to one region: GPT Image 2.5 with mask_url.
• Unsure: send the same two-image edit to both at n: 1 and compare the results and usage.cost.