Introduction: ByteDance Seedream's Dual Core for Still Images
ByteDance Seedream is known for high-quality output and flexible controls. Within Seedream 5 Lite, two complementary modes anchor most workflows: text-to-image for creation from prompts alone, and image-to-image for reference-driven edits, style transfer, and multi-image fusion.
The practical question is always the same: when should you stay in pure text generation, and when do you need reference images?
This guide compares both modes using complete parameter shapes and turns differences into a simple selection framework.
Seedream hub: /home/seedream
I. Snapshot: One Table for Both Modes
| Dimension | Seedream 5 Lite Text-to-Image | Seedream 5 Lite Image-to-Image |
|---|---|---|
| Core function | Generate from text only | Edit / generate from refs + text |
| Input | Prompt only (up to ~3000 chars) | 1-14 reference images + prompt |
| Output (basic / high) | 2K / 3K | 2K / 4K |
| Aspect ratios | 8 options | 8 options |
| Reference images | None | Up to 14 |
| Typical use | Concept art, ads from scratch | Style transfer, local edits, fusion |
| Learning curve | Easy | Medium |
| Positioning | Creative generation | Precise control |
II. Two Models Deep Dive
Model 1: Seedream 5 Lite Text-to-Image
The baseline Seedream 5 Lite route for generating brand-new images purely from language.
{
"model": "seedream/5-lite-text-to-image",
"prompt": "Image description (max ~3000 chars)",
"aspect_ratio": "1:1 / 4:3 / 3:4 / 16:9 / 9:16 / 2:3 / 3:2 / 21:9",
"quality": "basic / high"
}Highlights: no reference assets required; long prompts; basic targets 2K-class output, high targets 3K-class output.
Use cases: zero-to-one concepts, hero visuals without source photos, exploration before refinement.
Route: /home/seedream/5-lite-text-to-image
Model 2: Seedream 5 Lite Image-to-Image
Reference-first workflow for edits, material changes, and blending multiple visual cues.
{
"model": "seedream/5-lite-image-to-image",
"prompt": "Edit or fusion instructions (max ~3000 chars)",
"image_urls": ["Reference image URLs (1-14)"],
"aspect_ratio": "1:1 / 4:3 / 3:4 / 16:9 / 9:16 / 2:3 / 3:2 / 21:9",
"quality": "basic / high"
}Highlights: requires at least one image; up to 14 references; high reaches 4K-class output (higher ceiling than text-only high); typical limits include jpeg/png/webp up to ~10MB per file.
Use cases: style or material swaps with structure preserved, localized edits, multi-reference fusion.
Route: /home/seedream/5-lite-image-to-image
III. Head-to-Head Comparison
3.1 Resolution by quality tier
| Quality | Text-to-Image | Image-to-Image |
|---|---|---|
| basic | 2K-oriented | 2K-oriented |
| high | 3K-oriented | 4K-oriented |
Takeaway: if maximum pixel budget matters and you already have references, image-to-image high often wins.
3.2 Input flexibility
| Capability | Text-to-Image | Image-to-Image |
|---|---|---|
| Prompt-only input | ✅ | ❌ |
| Single reference | ❌ | ✅ |
| Multi-image fusion | ❌ | ✅ (max 14) |
| Preserve layout / structure | ❌ | ✅ |
3.3 Task routing cheat sheet
| Task | Recommended mode | Reason |
|---|---|---|
| “Draw a cat on grass” with no photo | Text-to-Image | Pure creation |
| Change fur color on an existing cat photo | Image-to-Image | Anchored edit |
| Tech cover with no assets | Text-to-Image | Fast ideation |
| Summer photo → winter mood, same composition | Image-to-Image | Structure lock |
| Merge composition A with palette B | Image-to-Image | Multi-reference fusion |
IV. Selection Decision Tree
What is your task?
|
|-- No reference images at all
| `-- Create from scratch -> Text-to-Image
|
|-- You have 1-14 reference images
| |-- Change material / style -> Image-to-Image
| |-- Local element swap -> Image-to-Image
| |-- Fuse multiple references -> Image-to-Image
| `-- Keep structure, change look -> Image-to-Image
|
`-- Pixel budget priority
|-- Need ~3K-class peak from prompt-only -> Text-to-Image (high)
`-- Need ~4K-class peak with refs -> Image-to-Image (high)V. Why Seedream 5 Lite Stands Out
5.1 Resolution ladder
Text-to-image high emphasizes crisp 3K-class delivery; image-to-image high pushes toward 4K-class output when references justify the extra detail.
5.2 Long prompts
Both modes accept very long natural-language briefs (on the order of thousands of characters) so art direction, lighting, and brand cues can live inside one prompt.
5.3 Eight aspect ratios
Shared ratio set spans square, portrait, landscape, and cinematic 21:9—enough for social, landing pages, and widescreen hero art.
5.4 Deep reference stacks
Image-to-image supports up to fourteen references—strong for mashups and iterative pipelines where multiple cues must coexist.
VI. Final Recommendations
| Scenario | Pick | Why |
|---|---|---|
| Zero-reference creation | Text-to-Image | Prompt-only freedom |
| Style or material transfer | Image-to-Image | Preserves structure while restyling |
| Surgical edits | Image-to-Image | Grounded in pixels you already trust |
| Heavy multi-image fusion | Image-to-Image | Up to 14 references |
| Maximum resolution with references | Image-to-Image (high) | 4K-class ceiling |
One-line playbook:
- No reference images → Text-to-Image.
- Need edits, fusion, or structure preservation → Image-to-Image.
- The modes complement each other: draft with text-to-image, polish with image-to-image.
Run both parameter sets directly on FuseAI Tools—start from /home/seedream.