Seedream 4.5 is ByteDance's image generation and editing model, suited for transforming text concepts, product photos, and visual references into advertising assets, image series, and scene images. It balances text understanding, visual aesthetics, and multi-image creation, supporting 2K and 4K output as well as sequential image generation. doubao-seedream-4-5-251128 is suited for workflows requiring a specific version, reference image editing, and related image sets.
Submit requests to the public API at api.acedata.cloud using the documented parameters, then use the results in your application.
Input parameters and result formats vary by service. Use the public API for this model and follow its guide for generation, task retrieval and editing operations.
The above parameters apply to this platform's doubao-seedream-4-5-251128 endpoint. Size presets are 2K or 4K; explicit width and height must meet this model's pixel and aspect ratio limits. The reference image array supports up to 14 items, and max_images for image sets is the upper limit for the number generated, without guaranteeing the actual number of images delivered.
Seedream 4.5 is suited to visual creation that includes titles, slogans, and layout intent. Prompts can specify the subject, lighting, materials, text content, and whitespace placement at the same time, allowing advertising concepts to enter the image more completely. Clearly specify short copy and its position, and check it character by character before delivery, rather than treating the generated image as the final layout.
Submit product photos or visual references, then describe which elements to retain and change; this can be used for background replacement, style adjustments, and scene reconstruction. Multi-image input is suitable for combining subject appearance, environment, and style cues; explain the purpose of each image to avoid composition, color, or subject features from different references interfering with one another.
After enabling linked image groups, you can generate content-related images around the same theme, suitable for storyboards, product series, and event visuals. When creating, clearly define the shared subject, unified color palette, and the variation points of each image, which is easier to review than simply asking for “consistent style.” Streaming mode can also deliver image results progressively, making it convenient for applications to display them as they are received.
Enter product photos, the target scene, and composition requirements, such as retaining the bottle's appearance, changing to a warm indoor background, and reserving a title area, to generate e-commerce scene images or advertising candidates. Before delivery, focus on checking packaging text, product proportions, and details, so the creative result serves the presentation rather than replacing real product information.
Provide the theme, short title, color scheme, and landscape or portrait requirements to create event key visuals, covers, or post graphics. 2K and 4K output suit different display needs; determine the composition first, then refine text and decorative elements. For dates, brand names, and long copy that must be presented accurately, retain manual layout and proofreading steps.
Organize a script into frame-by-frame descriptions, attach character or scene references, and use linked image groups to explore continuous visuals. The results can serve as visual drafts for storyboards, product stories, or content series. Review each frame for character appearance, actions, and shot relationships, and edit drifting images individually rather than publishing the entire group directly.
When you need 2K or 4K images, text understanding, and reference-based editing, 4.5 is a previous-generation flagship option worth trying; if your workflow must use the 1K preset, 4.0 is more suitable. Both support image sets and streaming output. Existing workflows with 4.0 assets can test 4.5 using the same prompts and reference images, then decide whether to switch after comparing text, subject retention, and visual style.
If you need high-resolution related image sets but do not rely on web search, you can choose 4.5. If the task requires web search or a 3K preset, consider 5.0 Lite; if layer separation is needed, consider 5.0 Pro. Pro's transparent-background control applies only to single-image editing: you must input a PNG with an alpha channel and set output_format to png, and it cannot be used together with layer separation. Pro does not support image sets or streaming, so choose based on input conditions and delivery format rather than judging suitability solely by version number.
Clearly specify the purpose of each image, subject retention requirements, and target copy; for editing, use image to submit an image URL or Base64, and avoid uploading assets without modification instructions.
For /seedream/images, specify model=doubao-seedream-4-5-251128, prompt, and size=2K. For related image sets, set sequential_image_generation=auto and use max_images to limit the maximum; choose streaming and asynchronous/callback modes separately.
Read the images actually returned rather than inferring the number of images from the request limit; synchronous, streaming, and asynchronous modes each use their own response formats, while asynchronous tasks are associated through /seedream/tasks.
Using the same ceramic vase as a reference, create three visually cohesive display images: a light-colored desk, a warmly lit living room, and greenery by a window; preserve the vase's proportions and glaze color, and avoid adding text.
Compare against the reference image to check the vase opening, patterns, size relationships, and image-set style, then inspect edges and textures in the high-resolution images.
Submit the full doubao-seedream-4-5-251128 to POST /seedream/images and provide a prompt. Do not use the Seedream 4.5 display name as the invocation value. When editing an existing image, add image and describe what content to retain and what content needs to be modified.
Presets support 2K and 4K. You can specify landscape, portrait, or square composition in the prompt; you can also specify width x height pixel values through size, but they must comply with the size range. Do not directly apply 1K or 3K presets from other versions to 4.5, and do not mix the two size expression methods at the same time.
Multi-image references use multiple images as creative inputs to help express subjects, scenes, or styles; related image groups generate a set of related images. The former is provided through image, while the latter is enabled by setting sequential_image_generation to auto. Neither guarantees one-to-one correspondence between inputs and outputs.
You can write short titles, slogans, and placement requirements into the prompt, and explore poster visuals with sizes and reference images. It is recommended to clearly specify the text that needs to appear to reduce ambiguity. After generation, spelling errors, repeated characters, and layout should still be checked; formal brand materials can additionally undergo manual text overlay and proofreading.
For background tasks, set async to true, obtain the task_id and query it through /seedream/tasks, or use callback_url to receive completion notifications. When images need to be displayed progressively, use stream and receive application/x-ndjson; streaming cannot be combined with async or callbacks.
Choose a development approach or application suited to this model and follow its setup guide.
Python SDK Integration Guide
SDK installation, authentication and examplesTypeScript SDK Integration Guide
SDK installation, authentication and examples
