Seedream 5 工作原理

从想法到成品图像三步完成。由链式推理视觉智能驱动。

Start creatingarrow_forward
Three floating luminous figures showing text, neural mesh, and finished artwork connected by a golden thread.
  1. 1

    Describe What You Want

    Type a natural-language prompt. You can be specific ("a corgi wearing aviator sunglasses, studio lighting, shallow depth of field") or loose ("a cozy autumn coffee shop"). Seedream 5 understands both. Upload reference images if you want to lock a character's face, a brand style, or a product. Structured prompts help, but aren't required — the model fills in sensible defaults.

  2. 2

    Seedream 5 Reasons Through Composition

    This is what's new in Seedream 5: chain-of-thought reasoning. Instead of generating pixels greedily, the model first plans composition, lighting, color palette, and subject placement — then generates. It can pull in real-time web search for current references (events, products, locations) rather than relying on stale training data. You see progress stream via SSE as the image builds.

  3. 3

    Edit, Iterate, Download

    Not perfect on the first try? Use plain-English edits ("make the sky warmer," "remove the lamp," "give her a red jacket"). Or use precision editing to change one element while keeping the rest. When ready, download your finished image. Need maximum resolution? Seedream 4.5 outputs native 4K — 3840×3840, 3840×2160, or 2160×3840. All outputs include commercial-use rights.

5 Pro Tips

lightbulb#1

Start broad, then refine. One vague prompt + three targeted edits beats one overly-engineered prompt.

lightbulb#2

Reference images lock identity. Upload up to 10 reference images for a character — Seedream 5 keeps the face consistent across completely different scenes.

lightbulb#3

Text prompts that read as sentences beat keyword lists. "A woman laughing in a sunlit kitchen" outperforms "woman, laughing, kitchen, sunny, portrait."

lightbulb#4

Use 1:1 for social, 16:9 for hero images, 9:16 for vertical video thumbnails. Choose early — switching aspect ratio mid-iteration burns credits.

lightbulb#5

When in doubt, ask Seedream to "think about the composition" in the prompt itself. Chain-of-thought reasoning triggers on cues like "plan the shot" or "arrange the scene thoughtfully."

常见问题

不需要。Seedream 5 的链式推理会为简短提示填入合理的默认值。详细提示给您更多控制,但简短提示适用于大多数创意探索。

这取决于模型。Seedream 5 Pro 最高渲染 2720×1530,Seedream 5 Lite 最高 3072px,Seedream 4.5 输出原生 4K,支持三种比例:3840×3840(1:1 正方形)、3840×2160(16:9 横向)、2160×3840(9:16 纵向)。无需放大 — 开箱即可打印。

可以。用自然语言描述编辑("将夹克改为红色"),或使用精确编辑来更改一个元素同时保留其余部分。编辑消耗的积分与新生成相同。

Ready to Create?

75 free credits waiting. No card, no trial expiry.

Start creatingarrow_forward
How Seedream 5 Works — Describe, Reason, Create | Seedream Studio