Seedream v4.7 Text-to-Image API by ByteDance

1. Introduction

Seedream 4.7 is ByteDance Seed's latest image generation and editing model, positioned as the balanced member of the Seedream family — tuned so that visual quality, editing fidelity, throughput, and cost land in the same place rather than trading against each other. This README covers the following API model identifiers:

• bytedance/seedream-v4.7/text-to-image — text-to-image, single output
• bytedance/seedream-v4.7/sequential — text-to-image, coherent image group
• bytedance/seedream-v4.7/edit — reference-image editing, single output
• bytedance/seedream-v4.7/edit-sequential — reference-image editing, coherent image group

Relative to Seedream 4.5 and Seedream 5.0 Lite, version 4.7 delivers a more even improvement across image quality, aesthetic judgment, instruction response, and overall image-to-image experience. Its distinguishing characteristic is efficiency: ByteDance positions its performance and cost profile as on par with the 4.0/4.2 generation while measurably ahead of 4.5, which makes it the family member intended for large-scale production workloads rather than one-off showpiece renders.

2. Key Features & Innovations

• Higher Image Quality and Aesthetics: In both text-to-image and image-to-image scenarios, outputs are cleaner and more natural, with stronger overall aesthetic composition than the preceding generation.

• Edit-Local Instruction Following: The model executes editing instructions more precisely while holding the source image's key information, subject consistency, and local structure stable — the combination that determines whether an edit is usable without manual repair.

• Multi-Reference Image Input: Accepts multiple reference images per request, as URLs or Base64, and reasons across them jointly for subject transfer, style grafting, and cross-image composition.

• Sequential Group Generation: The /sequential and /edit-sequential variants produce up to 15 thematically coherent images in a single call, holding style, palette, and viewpoint consistent across the set — suited to seasonal series, storyboards, and brand design systems.

• Native 4K Output: Generates up to 4096×4096 natively, alongside 1K and 2K tiers, with fine detail and texture preserved rather than upscaled after the fact.

• Flexible Sizing: Output size is specified either by resolution keyword (1K, 2K, 4K) or by explicit WIDTH*HEIGHT dimensions across a wide range of aspect ratios. In keyword mode the model selects an aspect ratio appropriate to the prompt, so a cinematic prompt returns a wide frame and a scene prompt returns a balanced one.

• Production-Oriented Efficiency: Performance and cost characteristics comparable to the 4.0/4.2 generation and clearly better than 4.5, supporting stable, large-scale deployment.

3. Model Architecture & Technical Details

Seedream 4.7 continues the Seedream 4 family's unified framework, in which text-to-image synthesis, single-image editing, and multi-image composition are served by one model rather than by separate specialized checkpoints. The family is built on an efficient Diffusion Transformer (DiT) paired with a high-compression Variational Autoencoder, an arrangement that keeps the number of latent image tokens low and is the structural reason the generation's inference cost stays flat as quality rises.

Version 4.7's gains concentrate in post-training rather than in a change of architecture family. Refinement against human preference feedback targets two properties in particular: aesthetic quality in unconstrained generation, and edit locality in reference-driven generation — changing what the instruction asks for while leaving identity, lighting, and surrounding structure untouched.

Output dimensions run from roughly 1M pixels (1K tier) to 16.7M pixels (4096×4096, 4K tier), selectable by keyword or explicit dimensions. Results are returned as hosted URLs or Base64, and sequential variants stream partial results as each image in the group completes.

4. Performance Highlights

ByteDance positions 4.7 within the family as follows: image quality, aesthetics, and instruction response improve over both Seedream 4.5 and Seedream 5.0 Lite, while performance and cost remain at the 4.0/4.2 level — a combination that no earlier family member offered simultaneously.

The practical consequence for high-volume pipelines is that 4.7 raises the acceptance rate of first-pass output — fewer regenerations and less manual retouching per delivered asset — without the per-request cost increase that previously accompanied a quality upgrade.

5. Intended Use & Applications

• Portrait Retouching: Targeted corrections to skin, lighting, wardrobe, and background with facial identity and tone held stable — the model's strongest editing scenario.

• Portrait Photography: Studio- and lifestyle-grade portrait generation with natural skin rendering and believable lighting.

• Stylized Generation: Converting subjects and scenes into a defined artistic style while preserving recognizable structure.

• Poster and Key Visual Design: Layout-aware compositions integrating typography, product, and background for campaign and print work.

• E-commerce and Product Visualization: Product recontextualization, background replacement, and packaging revisions at catalog scale.

• Visual Storytelling: Sequential variants generate coherent multi-image sets — seasonal series, storyboards, and brand design systems — in a single request.