Seedream v4.7 Edit Sequential API by ByteDance

1. Introduction

Seedream 4.7 Edit Sequential combines ByteDance Seed's Seedream 4.7 reference-based editing with group generation, exposed as the API model identifier bytedance/seedream-v4.7/edit-sequential. Given reference images and a natural-language instruction, it returns a coherent set of edited results rather than one — the same subject across several variations, all consistent with the references and with each other.

This is the variant for work where one source asset must yield a family of outputs: a model photographed in several outfits, a product staged in several environments, a character carried through several panels. Consistency is enforced in two directions at once — against the reference inputs, and across the generated set.

2. Key Features & Innovations

• Reference-Anchored Sets: Generates multiple edited images that stay faithful to the supplied references while varying along the axis the prompt describes.

• Dual Consistency: Holds identity and structure stable against the source images, and holds style, palette, and lighting stable across the generated group.

• Multi-Reference Input: Accepts up to 10 reference images per request as URLs or Base64, reasoning across them jointly.

• Combined Request Budget: Reference images and generated images share a single ceiling of 15 per request — a request with one reference can return up to 14 images. Requests exceeding the budget silently return fewer images rather than failing, so keep the sum at or below 15.

• Progressive Delivery with Fault Isolation: Images stream back as they complete, and a single moderation rejection does not discard the rest of the set.

• Full Resolution Range: Every image in the set may be generated at 1K, 2K, or 4K, or at explicit WIDTH*HEIGHT dimensions.

3. Model Architecture & Technical Details

This variant runs the Seedream 4.7 unified model with both reference conditioning and sequential decoding active. Reference images are embedded into a shared representation that anchors subject identity, while each generated image is additionally conditioned on the accumulated context of the set — the mechanism that keeps the group internally consistent rather than producing several independent edits of the same source.

The binding constraint is the combined budget: reference images plus generated images must not exceed 15. With the maximum of 10 references, at most 5 images are returned; with a single reference, up to 14. Exceeding the budget does not raise an error — the model reduces the number of generated images silently, so max_images should be set with the reference count in mind.

Output dimensions range from 921,600 pixels to 16,777,216 pixels (4096×4096), with aspect ratios between 1:16 and 16:1.

4. Performance Highlights

Producing the same set through repeated single-edit calls requires re-supplying references on every call and still drifts in style between results; a single sequential-edit request holds both anchors at once.

5. Intended Use & Applications

• Lookbooks and Outfit Series: One subject rendered in several coordinated wardrobe variations.

• Product Staging Sets: A single product placed in multiple environments with consistent lighting and branding.

• Character Sequences: A reference character carried across panels or scenes with stable identity.

• Campaign Variant Batches: Multiple on-brand executions derived from one source asset.

• Catalog Production: Bulk generation of consistent presentation variants from existing product photography.