Introduction

Ideogram has upgraded Precise Edit on its Ideogram 4.5 image model with a context window. You pick a region of up to roughly 2K (4.19 megapixels) inside a much larger image, and only that region is edited at its original pixels. Every pixel outside the window comes back exactly as you sent it, at the image's full width and height. It's a closed, API-served model available today in the Ideogram app, the Ideogram API, and partner platforms including fal, Runware, Runway, Luma, Luminal, ComfyUI, and Scenario.

Precise Edit itself isn't new. We covered the Ideogram 4.5 precise multi-turn edits launch in September. The news here is resolution: until now, big images had to shrink before they could be edited.

The Problem It Fixes

Editing the whole image is still capped at 4,194,304 pixels, the area of a 2048×2048 square. Ideogram's API docs give a blunt example: send an 8192×8192 image with the default context_window=none and it comes back as 2048×2048. Ideogram's launch page shows the same thing with a 4K product label that returns at 2K without a context window. With a window, the result stays at 4K.

The context window takes a different approach. Instead of shrinking the whole image, the model works only on the window at native resolution and stitches it back in. Ideogram's pitch is "no seams, no artifacts, no Photoshop," with small text, logos, and fine detail rendered more sharply.

How the Context Window Works

The new field is context_window on POST /v2/image/precise-edit/ideogram-4-5, and it has three modes:

  • none (default): Edits the whole image. Anything over 2048×2048 in area is scaled down.
  • auto: Needs a mask. Ideogram fits a window around the masked area, and the output keeps the full resolution.
  • y_min,x_min,y_max,x_max: You name the window by corner pixels, rows first. A mask is optional, but any black mask area must fall inside the window.

The window has these constraints:

  • At least 256 px on each side
  • At most 4,194,304 px in area, so 2048×2048 and 1024×4096 both fit
  • Aspect ratio between 1:6 and 6:1

The mask decides which pixels may change. The window is the surrounding context the model reads so the change blends in. If a mask covers more than 4.19 MP, no window fits, and the request falls back to a scaled-down whole-image edit.

Input Limits and Extras

  • Input image: JPEG, PNG, or WEBP, up to 50 MB and 67,108,864 pixels. That's exactly 8192×8192, so "8K" here means the pixel budget rather than a strict edge length. Ideogram's own gallery includes a 14,495×4,568 (66.2 MP) panorama.
  • References: Up to 4 reference_images, or 3 when you send a mask because the mask takes a slot. References guide the edit and are never altered.
  • Other fields: quality, seed, num_images, async with polling, signed webhook_url delivery, optional copyright detection, and a dry_run flag that prices a request without running it.

Here's one of Ideogram's own demo prompts, used with a mask and an explicit window on a 4,016×6,016 photo:

Change the handbag from pale blue to burgundy. Keep the leather texture and lighting.

Pricing

Price depends only on quality. Image size and the context window don't change it:

  • Very low: $0.008
  • Low: $0.03
  • Medium: $0.06
  • High: $0.22

Is Scenario's "16K Edit" the Same Thing?

Yes. Ideogram reposted Scenario's "type the fix on a 16K image" post. Scenario's link goes to its ideogram-v4-5-precise-edit model page, and the "up to 67 MP" figure it quotes matches Ideogram's 67,108,864-pixel input cap. "16K" there describes a very wide image (like Ideogram's 14.5K-wide panorama), not a separate model or a higher limit.

Where to Run It Today

  • Ideogram app: Precise Edit
  • Ideogram API: /v2/image/precise-edit/ideogram-4-5
  • Partners named by Ideogram: Runware, Runway, IDEA, fal, Luminal, Luma, ComfyUI, and Sociaro. Scenario and MyDesigns announced support separately.

Original Source

Conclusion

This is the quietly useful kind of upgrade. Ecommerce, real-estate, and print teams no longer have to choose between AI editing and keeping their master file. You fix one label, fixture, or coat color at native pixels, and the other 60-odd megapixels never touch a model. It sits nicely beside Ideogram's recent Vector 1.0 as another production-first tool. Start with context_window=auto and a tight mask, then grab the coordinates yourself when you need full control.

—Aurelia ♡