Seedream 5.0 Pro is live!

ByteDance Seed launches Seedream 5.0 Pro, a multimodal AI model improving infographic layouts, precision editing, layer separation, and multilingual text.

New iteration for ByteDance Seed's multimodal image generation model, dubbed Seedream 5.0 Pro. The publisher reports progress on text-image alignment, structural consistency, text rendering, and aesthetics, then details several new capabilities.

The first concerns infographics: the model converts data, concepts, and dense text into hierarchical layouts, with, according to Seed, the ability to combine a timeline, graphs (bar, pie, line), and realistic photos in a single image, with text rendered without spelling errors. The same engine produces interface mock-ups, such as an e-commerce page with a navigation bar and floating cards.

Next comes precision editing, supported by grounding, which is the spatial localization of elements within the image. Seedream 5.0 Pro reasons about these positions (solving and then writing exercise answers in the designated boxes, translating a menu while preserving its layout) and accepts selections by point, lasso, frame, or sketch for local retouching: recoloring by Hex code or reference swatch, material replacement, adding or removing objects, isolating areas delimited by colored frames. Its layer separation function decomposes a poster into more than ten editable and transparent layers, reconstructs masked parts by inpainting, and allows for the replacement of the main subject. Multi-image merging completes the set.

Photorealism also progresses: finer light, materials, and skin textures, reflections and refractions based on real physics, support for panning, and recomposition of multiple faces into a group photo. Regarding languages, both input and generation are native in more than ten idioms (including French, German, Russian, Japanese, Korean, Spanish, Arabic), with adaptation of typographic rules: Arabic read from right to left, Spanish accents preserved.

Seed acknowledges room for improvement in the finesse of text rendering and the consistency of pixel-perfect editing.