Seedream 4.5

Overview

Seedream 4.5 is ByteDance Seed's image model for text-to-image and editing, and the successor to Seedream 4.0. One model covers prompt-only generation, photo edits, multi-image composition and sets of images that share a character or style.

ByteDance describes 4.5 as an all-round upgrade from scaling the model. In edits, facial features, lighting and color tone from the reference carry through the requested change. It picks out the main subjects accurately when editing across several inputs and renders small text more clearly. On its internal MagicBench evaluation, ByteDance reports gains over 4.0 in prompt adherence, alignment and aesthetics.

What you can build

• Virtual try-on. The default example dresses the person in image 1 in the outfit from image 2.
• Brand kits from one logo. A sample turns a logo into a lanyard, cap, tote bag and card in a single request.
• Storyboards and comic panels with the same person in every scene, like the four-scene astronaut sample.
• Posters, e-commerce product images with a headline, and diagrams or educational illustrations.
• Targeted edits: add, remove or replace an element, recolor one word, or change a material while keeping the pose. Arrows, boxes or doodles drawn on the input mark the area to change.

Inputs

• Prompt (required). Describe the image or the edit in plain sentences. For several images, ask for them in the prompt ("a set of 4") and set the output limit to match.
• Input images (optional). Up to 14 references as uploads or URLs, in JPG, PNG, WebP, GIF, HEIC or HEIF. Leave it empty for text-to-image; refer to several inputs by position: "image 1", "image 2".
• Advanced: resolution (auto, 2K or 4K), aspect ratio (auto, 1:1, or landscape and portrait pairs for 3:2, 4:3, 5:4, 16:9 and 21:9), an output limit from 1 to 15 (default 1), and a watermark switch that is off by default.

For text-to-image, set both a resolution and an aspect ratio. With input images, auto follows the first image's ratio and scales it into the supported size range.

Outputs

One image or a set per run. At 2K a square image is 1920x1920 and 16:9 is 2560x1440; at 4K they are 2880x2880 and 3840x2160.

Inputs and outputs share a cap of 15 per request, so 14 references leave room for one result. The output limit is a ceiling: the model decides from the prompt how many images to return, so a set can come back smaller.

Recommended settings

• Describe subject, action and setting in full sentences, then add style, color or lighting if they matter. ByteDance's guide favors short, precise prompts over keyword lists.
• Put text that must appear in the image in double quotes, for example a poster titled "Summer Sale".
• For edits, name the target plainly, say what should stay unchanged ("keep the pose and background"), and leave resolution and aspect ratio on auto.

Limitations

• Prompts are supported in English and Chinese.
• ByteDance recommends at most 600 English words or 300 Chinese characters. Longer prompts may scatter the information, so the model can overlook details and keep only the main elements.
• Text you do not spell out can come out as invented words, like the contact line and web address on the brand-kit card.

Safety & compliance

Turning on the watermark stamps an "AI-generated" label in the lower-right corner of each image.