Skip to content
HiAPI Docs
HiAPI
Playground Get API Key
English
English 中文

Seedream 5.0 Lite Image to Image API

POST /v1/tasks

All models are called through the Unified Async API POST /v1/tasks endpoint; only the input fields differ (see input parameters below).

Model summary

Model name seedream-5.0-lite/image-to-image
Type Image generation (image-to-image)
Endpoint POST /v1/tasks
Pricing See HiAPI Pricing

ByteDance Seedream 5.0 Lite image editing: instruction-based edits with up to 14 reference images — rewrite in-image text, swap materials, blend references — while untouched areas stay intact. 2K and 4K cost the same.

Production guidance

Production guidance
  • For production, pass callback.url at the top level of the request body so HiAPI can notify your service when the task reaches a terminal state.
  • GET /v1/tasks/:id is better for local debugging, low-volume jobs, or fallback reconciliation if a callback is missed.
  • Use callback.when=final. Both success and fail are terminal states, so your service should deduplicate by taskId.

Best suited for

Instruction-based editing

Describe the change in plain language — recolor, add elements, relight. Areas you do not mention stay untouched, ideal for poster restyling and product touch-ups.

promptimage_urls
In-image text rewriting

Rewrite signs and titles inside the image while keeping the original font style and glow — bilingual swaps work in a single edit.

prompt
Multi-reference composites

Blend up to 14 reference images using "the X from the first image, the Y from the second" phrasing — still billed as one output image.

image_urls
Material and mood swaps

Glass to ice, daylight to rainy night — whole-scene material or atmosphere changes with composition and subject positions preserved.

prompt

Request parameters

model string required

Fixed value seedream-5.0-lite/image-to-image.

example seedream-5.0-lite/image-to-image
input object required

Business parameters. Put Seedream 5.0 Lite Image to Image-specific configuration here.

prompt string required

Edit instruction, 3–3000 characters: say what to change and what to keep. Unmentioned areas stay as close to the original as possible.

image_urls string[] required

Array of 1–14 reference image URLs. JPEG/PNG/WebP, up to 10MB each. Billing is per output image, independent of reference count.

aspect_ratio enum required

Width-to-height ratio of the output image.

default 1:1 enum: 1:14:33:416:99:162:33:221:9
resolution enum required

Output resolution tier: 2K standard (≈2300px long edge) or 4K ultra (≈4096px long edge, verified). Both tiers cost the same.

default 2K enum: 2K4K
callback object optional

Optional callback configuration. When set, HiAPI notifies your service when the task reaches a terminal state.

url string required

Required when callback is set; HTTPS URL that receives terminal task notifications.

example https://your-domain.com/hiapi/callback
when enum optional

Callback trigger timing. Use final.

default final enum: final

Example requests

Poster recolor + new element (instruction edit)

Keep layout and subject; swap the palette and add one element.

Request body
{
  "model": "seedream-5.0-lite/image-to-image",
  "input": {
    "prompt": "Keep the poster layout, calligraphy title and bronze vessel unchanged. Change the color grade from aged gold and teal to a deep crimson and ivory palette, and add a subtle red wax seal stamp in the bottom right corner.",
    "image_urls": [
      "https://static.hiapi.ai/previews/seedream-5.0-lite/2026/07/03/1783057555249-f76441ce83f6c0d6-seedream-lite-t2i-g1-bronze.webp"
    ],
    "aspect_ratio": "3:4",
    "resolution": "2K"
  }
}
Glass chess to ice sculpture (material swap)

Whole-scene material and mood change with piece positions preserved.

Request body
{
  "model": "seedream-5.0-lite/image-to-image",
  "input": {
    "prompt": "Transform the glass chess pieces into luminous ice sculptures with intense inner glow, change the dark background to a dramatic vivid aurora borealis storm in magenta and emerald filling the sky, keep the wooden table and every piece position unchanged, frost spreading across the board",
    "image_urls": [
      "https://static.hiapi.ai/previews/seedream-5.0-lite/2026/07/03/1783057569706-dbb85f6269c1f5c0-seedream-lite-i2i-input-chess.webp"
    ],
    "aspect_ratio": "16:9",
    "resolution": "4K"
  }
}
Two-reference hologram composite

Blend two references into a single new scene.

Request body
{
  "model": "seedream-5.0-lite/image-to-image",
  "input": {
    "prompt": "Combine the two reference images: take the ancient bronze ding vessel from the first image and place it as a giant glowing translucent cyan holographic exhibit floating above the rainy neon street from the second image, crowd silhouettes with umbrellas looking up at it, keep the street's neon signs clearly readable",
    "image_urls": [
      "https://static.hiapi.ai/previews/seedream-5.0-lite/2026/07/03/1783057555249-f76441ce83f6c0d6-seedream-lite-t2i-g1-bronze.webp",
      "https://static.hiapi.ai/previews/seedream-5.0-lite/2026/07/03/1783057274330-5b12897efad48ccd-seedream-lite-t2i-cover.webp"
    ],
    "aspect_ratio": "16:9",
    "resolution": "4K"
  }
}

Getting the result

  1. The response returns a taskId immediately without waiting for generation to finish.
  2. In production, prefer waiting for callback.url to receive the terminal notification. For local debugging, poll GET /v1/tasks/:id.
  3. When status=success, download the generated image from output[].url.
  4. When status=fail, fix the request based on the returned error instead of retrying the same invalid payload.

FAQ

How many reference images are supported? Do more images cost more?

Up to 14 reference images (JPEG/PNG/WebP, ≤10MB each). Billing is per output image regardless of how many references you pass. See the pricing page for live rates. View pricing

How do I keep the edit from touching other parts of the image?

Explicitly state what must stay unchanged (for example, "keep the layout and subject positions unchanged"). The model preserves unmentioned areas, and the more specific the requested change, the more stable the result.

Can it rewrite text inside the image?

Yes. Name the original and the new text (for example, "change the sign NIGHT NOODLES to MIDNIGHT RAMEN"). The model keeps the original font style, glow and position, and bilingual swaps work in a single edit.

How do I get the edited image?

The request returns a taskId immediately. When the task reaches a terminal state, download the image from output[].url. In production, pass callback.url at the top level of the request body to receive the terminal notification instead of polling.

Next steps