Seedream v4.7 Sequential API by ByteDance

1. Introduction

Seedream 4.7 Sequential is the group-generation variant of ByteDance Seed's Seedream 4.7 image model, exposed as the API model identifier bytedance/seedream-v4.7/sequential. Instead of returning a single image, it produces a set of images that belong together — a season cycle, a storyboard, a product line, a brand system — from one prompt in one request.

The variant shares the base model's generation pipeline and adds sequential decoding, in which each image in the set is produced with awareness of the others. Style, palette, lighting, and viewpoint stay consistent across the group without the prompt engineering normally required to force consistency across independent calls. Results stream back as each image completes, so a partially finished set is usable before the whole request returns.

2. Key Features & Innovations

• Coherent Image Sets: Generates up to 14 related images per request, holding style, color treatment, and composition logic constant across the set.

• Model-Determined Count: The model decides how many images the prompt actually calls for, up to the requested ceiling — a prompt describing four seasons yields four panels without the count being enforced manually.

• Progressive Delivery: Each image is returned as soon as it is generated rather than after the full set completes, keeping long multi-image requests responsive.

• Per-Image Fault Isolation: If one image in a set fails content moderation, the remaining images still return; only an internal service error halts the batch.

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

• Shared Quality Baseline: Inherits Seedream 4.7's improvements in image quality, aesthetics, and instruction following, applied uniformly across the group.

3. Model Architecture & Technical Details

Sequential generation runs on the same Diffusion Transformer and high-compression VAE as the base Seedream 4.7 model; the difference is in decoding rather than in weights. The model conditions each image on the accumulated context of the set, which is what produces cross-image consistency in style and subject rendering rather than a collection of independent interpretations of the same prompt.

The upper bound on a request is 15 images total, counting both reference images supplied as input and images generated as output. For this text-to-image variant no references are consumed, so the practical ceiling is the max_images parameter. When a prompt implies fewer images than the ceiling, the model returns fewer.

Output dimensions range from 921,600 pixels (for example 1280×720) to 16,777,216 pixels (4096×4096), selectable by resolution keyword or explicit dimensions, with aspect ratios between 1:16 and 16:1.

4. Performance Highlights

Because every image in a set is billed individually, the practical benefit of sequential generation is not cost but coherence: producing the same set through separate calls requires reference-image chaining and still drifts in style, whereas a single sequential request holds the set together by construction.

5. Intended Use & Applications

• Seasonal and Thematic Series: One subject rendered across time, weather, or mood, with a single visual language.

• Storyboards and Visual Narratives: Sequential panels for pitch decks, animatics, and comic layouts.

• Brand and Design Systems: Coordinated collateral — packaging, apparel, signage — sharing one palette and style.

• Product Line Visualization: Variant renders of a product family under consistent lighting and staging.

• Campaign Asset Batches: Multiple on-brand executions of a single creative concept, generated in one pass.