Ideogram 4.5 targets image-edit drift, with mixed benchmark results

Ideogram has released Ideogram 4.5 in its web app, API, and partner platforms. The model focuses on precise, repeated edits that preserve composition, color, and unaffected pixels across multiple turns. Open weights are also planned.

Public benchmark results support that specialist positioning. Ideogram 4.5 ranks No. 23 on Artificial Analysis’s editing leaderboard and No. 34 for text-to-image generation. Its strongest results appear in typography, composition, and workflows that modify an existing image without rebuilding the full frame.

No. 23 conceals narrower strengths

On the AA-Image-Editing v2.0 leaderboard, Ideogram 4.5 sits behind HiDream-O1-Edit-1.5 and HunyuanImage 3.0 Instruct, and ahead of ByteDance’s Seedream 5.0 Lite. It is the first Ideogram model listed on the editing board.

Artificial Analysis divides editing performance into seven actions and 10 use cases. Ideogram 4.5 performs closest to the category leaders in three areas:

• Text or Symbol Edits: Modifying typography, logos, signage, and other symbols within an image.
• Scene and Style Edits: Changing lighting, visual style, backgrounds, or settings.
• Composition and Framing: Cropping, extending, reframing, and changing the apparent camera view.

Those capabilities translate most directly to social media graphics, creator assets, architectural visualization, real-estate staging, and UI or UX mockups. They also match Ideogram’s established strength in rendering legible text inside generated images.

Generation improves selectively

Ideogram 4.5 ranks No. 34 for text-to-image generation, up from No. 40 for Ideogram 4.0 Quality. It narrows the gap to the category leaders in five of nine measured capabilities, with its largest gains in text rendering and complex compositions.

Its strongest generation categories are Knowledge, Physics, and Complex Compositions. These tests cover recognizable landmarks and species, physical relationships such as gravity and support, and prompt details involving counts, positions, and correctly assigned attributes.

At the use-case level, the largest improvement over version 4.0 appears in Animation and Gaming. Marketing, advertising, book covers, stock imagery, and editorial illustration also align with the model’s stronger categories.

Repeated edits are the core bet

Ideogram describes version 4.5 as resistant to edit drift, the gradual accumulation of artifacts, color changes, softened details, and shifted pixels after successive revisions. The company says the model can preserve an image through multiple edits more reliably than GPT Image and Nano Banana, though teams should test that claim against their own assets and prompts.

The API supports two main workflows:

• Precise edit: Produces an edited result at the input image’s dimensions.
• Generate and edit: Combines reference-driven generation with editing instructions.

Requests can include up to four reference images, an optional mask, and a high-precision mode designed to preserve unchanged pixels. Developers can also edit a crop from a high-resolution image and merge the result back into the source, avoiding a full-image downscale during localized changes.

The model offers four quality modes, prompts of up to 10,000 characters, native 2K output, and supported dimensions reaching roughly 3K. Its text tools can alter stylized lettering in place or translate copy while retaining the surrounding layout and visual treatment.

Editing prices vary by 27 times

Artificial Analysis evaluated the High quality tier, where editing costs more than generation. Lower modes substantially reduce the per-image price for batch workloads.

The 27.5-fold spread between the lowest and High edit tiers gives developers room to trade output quality for throughput. Production evaluations should compare quality modes using representative masks, reference images, and multi-turn sequences because single-edit tests will not expose cumulative drift.

Where it fits in production

Ideogram 4.5’s benchmark profile favors pipelines that begin with an approved image and require controlled revisions. Suitable workloads include:

1. In-image copy changes: Replace text in posters, advertisements, interfaces, or signs without regenerating the entire frame.
2. Product campaign variants: Change backgrounds, props, or copy across several turns while preserving the product’s appearance.
3. Architectural visualization: Swap furnishings, materials, finishes, and color palettes while retaining the underlying structure.
4. Localization: Translate headlines and labels while preserving typography, spacing, and layout.
5. Targeted high-resolution edits: Modify a crop and return it to the original image without processing every pixel.

Teams using GPT Image or Nano Banana for initial generation can evaluate Ideogram 4.5 as a dedicated editing stage, particularly when later revisions introduce artifacts or unwanted composition changes. The planned open-weight release could also support self-hosted editing pipelines, subject to its eventual license, hardware requirements, and model size.