Ideogram has released model 4.5, which is focused on precise editing of existing images rather than generation from scratch. The Precise Edit mode edits only the specified part of the frame and returns the result in the original resolution. The model is already available in the service itself, via API, and through partners; open weights are promised for later, and for now, all supporting evidence comes from the vendor itself.

image
image
image

What happened

Ideogram introduced model 4.5, aimed at editing existing images. The key Precise Edit mode operates according to three rules: pixels not affected by the edit remain unchanged; the result is returned in the original resolution of the frame; and the area of change can be limited with a mask. According to Ideogram, in chains of sequential edits under such a scheme, fewer artifacts, detail shifts, and color changes accumulate. The vendor shows applied scenarios: changing the color of objects and the lighting of a scene with consistent shadows and reflections, editing and translating text directly in a layout, and editing an enlarged fragment of the source with subsequent seamless assembly. A file of 4016 × 6016 pixels was used in the demonstration. The model is already available in the Ideogram service itself, via API, and through partners: Picsart, fal, Krea, Runway, Pika, Gamma, Luma, and ComfyUI.

Context

The problem that Precise Edit tackles is known to anyone who has run chains of edits in generative editors: the model drifts between iterations — pixels shift, colors and textures change, and after several steps, the image imperceptibly moves away from the original. For professional processes — product photography, restoration, interiors, working with text in a layout — such instability is unacceptable: there, editing must be surgical, and the untouched part of the frame is valued higher than a new version of the image. It is precisely in these Photoshop workflows that generative models have so far penetrated weakly, because the conservatism of the result is more important there than the ability to invent an image from scratch. Against this backdrop, Ideogram's bet looks like a change of niche rather than another round in the generator competition.

Why this matters for the industry

For the industry, 4.5 is not "just another generation model" but a building block for products on top of existing images. Preservation of untouched areas, original resolution, masks, and multi-step chains of edits make the AI editor a predictable component of automated pipelines: batch retouching of catalogs, light correction in product photography, restoration, localization of creatives through text editing in a layout. Access via API and the partner network immediately provides a distribution channel, and an MVP on top of point editing — for example, batch replacement of product color — can be assembled in days. Strategically, Ideogram is moving away from the oversaturated race of generation from scratch into a niche where the conservatism of editing is valued higher than novelty, and if practice confirms the claims of low drift, multi-step editing may become a working standard for catalogs and marketing layouts, and the metric of preserving untouched pixels may become the main one for comparing edit models.

Why this matters for users

You can try it right away. Open Ideogram, select model 4.5, upload a finished image, and describe one specific change — for example, "make the sneakers blue, leave the rest unchanged." Set the next edit for the obtained result, not for the original; in the API, the Precise Edit mode is provided for this scenario, and partners provide access to the same model from their platforms. The most honest assessment will be given by your own test: take a frame characteristic of your work, perform 5–10 sequential edits, and see what changed in the areas that should not have been changed, and at which step the differences began. If the frame remains stable after such a chain, the model covers a class of tasks that previously required manual retouching.

What is still unknown / limitations

The entire evidence base is primary sources from the vendor itself: the model page, a post on X, and a demo video. There is no technical report, article, benchmarks, or pixel drift metrics; all comparative claims are self-declared, and the demonstration cases were selected by Ideogram itself. The formulation "untouched pixels are unchanged" more likely describes an engineering guarantee — compositing the editing area over the original using a mask, a standard technique — than a property of a new architecture: the implementation is not disclosed, and it is impossible to draw conclusions about the training method from the statement. For a migration decision, pricing, latency, and limits are missing, so unit economics cannot be calculated. Open weights are promised but not published, and the date is unknown; scenarios for further development of the topic are interpretations, not facts.

Sources

Author

Look at AI, editorial team