ByteDance released Seedream 4.5, a scaled-up version of its image generation model that focuses on better typography, multi-image editing, and preserving fine details from reference images.
Key improvements over Seedream 4.0
Typography and dense text rendering - The model can now handle text-heavy design scenarios like posters, ecommerce product pages, and small-type layouts without losing legibility. Examples show it translating entire designs into different languages with specific font styles while keeping all other layout elements unchanged.
Multi-image editing accuracy - Seedream 4.5 better identifies which subject to keep or modify when working with multiple images. The demo page shows it isolating a single character from a group photo while removing others, or swapping materials (metallic dress to transparent water) while preserving pose, lighting, and composition.
Benchmark improvements - ByteDance's MagicBench evaluation shows Seedream 4.5 with significant gains over 4.0 across prompt adherence, alignment, and aesthetics in both text-to-image and single-image editing workflows.
