Ideogram 4.5 Coming Soon as Open Weights, a Quick Look

Ideogram 4.5 will soon be released as open weights. Qwen Image 2.1, announced recently, also shows very strong editing and generation, and Ideogram 4.5 shows impressive editing abilities and original-content preservation too, so I got curious and took a look.

Of course, it's also a generation-capable model, and according to them it shows strong performance in "multi-turn" editing.

Specs (API)

• Resolution: Native 2K, up to about 4.19MP (e.g., 2048×2048, 2560×1440, 3328×1248, etc.)
• Aspect ratio: 1:6 to 6:1
• Editing: Edits at the original resolution. Demonstrated examples up to 4,016×6,016 (24.2MP). You can crop, edit, and then stitch the result back onto the original.
• Reference images: up to 4 (3 plus a mask when using a mask)
• Mask: Specify the editing area with a black-and-white mask
• Input file limit: 25MB
• Output: JPG / PNG / WEBP
• Other: Seed setting, Magic Prompt (automatic prompt expansion) on/off/auto
• Speed: About 24–32 seconds per request (Runware basis)

Price (per image, Runware basis)

Performance

On the benchmark, it ranked 34th out of 167 for T2I and 23rd for editing, which shows it's stronger than Google's Nano Banana 2. It's especially strong at editing: according to the results, it kept 95.4% of the pixels unchanged during editing, while GPT Image 2.5 redrew about 25%. However, for text, while all the spelling was correct, there were punctuation mistakes such as hyphens being converted to em dashes and errors in spacing around decimal points. GPT Image 2.5 got all 731 characters exactly right, so GPT Image seems stronger for text.

Estimated local running requirements?

VRAM: Estimated at 20–22GB for BF16, 12–15GB for FP8, and 8–10GB for NF4.