till/managarten
1
0
Fork 0
mirror of https://github.com/Memo-2023/mana-monorepo.git synced 2026-05-15 19:19:41 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 4
managarten/apps/picture/docs/models/seedream-4.md
Wuesteon d36b321d9d style: auto-format codebase with Prettier 
Applied formatting to 1487+ files using pnpm format:write
  - TypeScript/JavaScript files
  - Svelte components
  - Astro pages
  - JSON configs
  - Markdown docs

  13 files still need manual review (Astro JSX comments)
2025-11-27 18:33:16 +01:00

3.3 KiB
Raw Blame History

ByteDance SeeDream 4

Overview

SeeDream 4 is ByteDance's latest generation image model featuring unified text-to-image generation and precise single-sentence editing capabilities. It offers significant improvements over SeeDream 3 with higher resolution support and more flexible workflows.

Model Details

  • Provider: ByteDance
  • Replicate ID: bytedance/seedream-4
  • Version: 054cd8c667f535616fd66710ce20c8949bf64ac3d9a3459e338f026424be8bec

Key Features

  • Unified Architecture: Single model for both generation and editing
  • Ultra High Resolution: Up to 4K (4096x4096) output
  • Multi-Reference Support: Use up to 10 reference images
  • Batch Generation: Generate up to 15 images in one request
  • Precise Editing: Natural language prompt-based editing
  • Consistent Characters: Maintains character consistency across multiple outputs

Default Parameters

  • Resolution: 2048x2048 (2K preset)
  • Steps: 50 (automatic based on size preset)
  • Guidance Scale: 7.5 (automatic)
  • Supports Negative Prompts: No
  • Supports Seed: No
  • Supports Image-to-Image: Yes (via image_input array)

Supported Aspect Ratios

8 fixed ratios:

  • Square: 1:1
  • Landscape: 4:3, 16:9, 3:2, 21:9
  • Portrait: 3:4, 9:16, 2:3

Additionally supports "match_input_image" when using reference images.

Size Presets

  • 1K: Best for quick previews (1024-2047px)
  • 2K: Balanced quality and speed (2048-3071px) - Default
  • 4K: Maximum quality (4096px)
  • Custom: Specify exact width/height (1024-4096px range)

Best Use Cases

  • Character consistency across multiple scenes
  • High-resolution commercial imagery
  • Image editing with natural language prompts
  • Multi-view generation from single prompt
  • Reference-based generation
  • Batch creation of variations

Example Prompts

  1. "A professional portrait of a woman in business attire, modern office background, natural lighting"
  2. "A selection of photos of this character [reference] exploring a bookshop called 'SeeDream 4'"
  3. "Multiple views of a futuristic car design, different angles and lighting"

Tips for Best Results

  • Use the 2K preset for optimal balance of quality and speed
  • Leverage multi-reference input for character consistency
  • Use natural language for precise editing instructions
  • Request multiple outputs for variations in one go
  • Specify detailed scene descriptions for better results
  • For ultra-high quality, use 4K preset

Strengths

  • Exceptional high-resolution output (up to 4K)
  • Unified generation and editing workflow
  • Excellent character consistency
  • Multi-reference and batch capabilities
  • Fast inference compared to quality level
  • Natural language editing

Limitations

  • No manual seed control
  • No negative prompt support
  • Fixed aspect ratios only (no completely custom ratios)
  • Slightly higher cost than SeeDream 3 ($0.03 vs $0.025)

Cost

$0.03 per generation (regardless of resolution)

Migration from SeeDream 3

If you're upgrading from SeeDream 3:

  • Resolution limits increased: 2048x2048 → 4096x4096
  • New parameter structure (size presets instead of raw dimensions)
  • Removed: seed support, negative prompts
  • Added: image_input array, multi-image generation, higher resolutions
  • Slightly higher cost but significantly more features