Seedream v4.7 Edit API by ByteDance

1. Introduction

Seedream 4.7 Edit is the reference-driven editing variant of ByteDance Seed's Seedream 4.7 image model, exposed as the API model identifier bytedance/seedream-v4.7/edit. Supply one or more reference images together with a natural-language instruction, and the model returns a single edited result that applies the change while leaving everything else intact.

Editing is where version 4.7's post-training gains are most visible. The model's stated advantage over Seedream 4.5 and 5.0 Lite is that it executes an edit instruction more faithfully while holding the source image's key information, subject consistency, and local structure stable — the property that determines whether an edited image ships as-is or goes back for manual repair.

2. Key Features & Innovations

• Edit Locality: Applies the requested modification while preserving identity, lighting direction, color grading, and surrounding structure — the difference between an edit and a regeneration.

• Multi-Reference Composition: Accepts up to 10 reference images and reasons across them jointly, resolving instructions that span inputs ("replace the garment in image 1 with the one from image 2").

• Conversational Instructions: Edits are described in plain language, with numbered references mapped to inputs automatically — no masks or region annotations required.

• Broad Input Format Support: References may be JPEG, PNG, WEBP, BMP, TIFF, GIF, HEIC, or HEIF, supplied as URLs or Base64, up to 30 MB and 36M pixels each, with aspect ratios from 1:16 to 16:1.

• Full Resolution Range: Output at 1K, 2K, or 4K, or at explicit WIDTH*HEIGHT dimensions, independent of the reference images' own dimensions.

• Efficiency at Editing Scale: Retains the 4.0/4.2 generation's cost and performance profile, which matters most in editing workflows where every asset passes through several revision rounds.

3. Model Architecture & Technical Details

Seedream 4.7 Edit is the same unified model as the text-to-image variant rather than a separately trained editing checkpoint — the Seedream 4 family serves generation, single-image editing, and multi-image composition from one set of weights. Reference images are embedded into a shared representation so that cross-image instructions can be resolved before synthesis begins.

The improvement in 4.7 comes from post-training against human preference feedback that rewards edit locality specifically: changing what the instruction names and nothing else. This is what produces the model's stability in facial features, lighting, and local structure under edits.

Output dimensions range from 921,600 pixels (for example 1280×720) to 16,777,216 pixels (4096×4096), with aspect ratios between 1:16 and 16:1. Results are returned as hosted URLs or Base64.

4. Performance Highlights

ByteDance positions 4.7's image-to-image experience as improved over both 4.5 and 5.0 Lite while holding the earlier generation's cost profile — the combination that makes it suited to high-volume retouching pipelines rather than one-off hero images.

5. Intended Use & Applications

• Portrait Retouching: Wardrobe, background, skin, and lighting adjustments with facial identity and tone held constant.

• Virtual Try-On and Outfit Transfer: Move garments and accessories between reference photos while preserving the subject's pose and scene lighting.

• Product Recontextualization: New backgrounds, staged environments, and packaging revisions with product geometry and branding faithful.

• Stylization of Existing Imagery: Convert a photograph into a defined artistic style while keeping recognizable structure.

• Poster and Layout Revisions: Headline swaps, localization passes, and layout adjustments on text-bearing designs.