Seedream Image Generation & Editing

Complete guide to ByteDance BytePlus ModelArk Seedream image-generation models — three active versions (5.0 / 4.5 / 4.0), 4K output, multi-image fusion, batch sequence generation, reference-image editing, all on a single endpoint.

Overview

Seedream is the flagship image-generation model series from ByteDance BytePlus ModelArk, with a unified generation-editing architecture: text-to-image, single-image editing, multi-image fusion, and batch sequence generation all run through one /v1/images/generations endpoint — only the parameters differ. APIYI has a strategic partnership with BytePlus and integrates every active version on day one.

🎨 Highlights: three active versions (5.0 / 4.5 / 4.0) on unified billing + 4K output + up to 10 reference images for fusion + batch (input + output ≤ 15) + leading text rendering. Ideal for e-commerce hero images, ad posters, product photography, and content creation — anywhere high quality plus readable text matters.

All image APIs are synchronous — there is no task ID to poll, and if your client disconnects the result is lost while the request is still billed. Set a generous timeout for this model; see Image API Essentials & Best Practices.

Which Seedream is this? APIYI serves Seedream on official overseas BytePlus (international) resources — not the mainland Doubao / Volcengine edition. The international edition applies a comparatively lighter content-moderation policy than the domestic one, giving you more creative freedom — a real advantage of this platform, but it does not mean moderation-free: BytePlus still runs built-in content safety checks, and violating prompts or reference images are rejected with 400/403 (rejections are not billed). Use within compliance.

Text-to-Image API

POST /v1/images/generations. Generate images from prompts at 1K / 2K / 3K / 4K or exact pixel sizes.

Image Editing API

Same endpoint with image parameter. Single-image editing, multi-image fusion, and batch sequence (up to 15 images).

Historical Versions

5.0 / 4.5 / 4.0 spec comparison, pricing differences, and migration guide.

Why APIYI’s Seedream

Drop-in replacement for the BytePlus ModelArk official channel, optimized for production use along three axes — stability, cost, integration:

Strategic partnership · stable resources

Authorized direct connection to BytePlus ModelArk. Request and response behavior is identical to the upstream — no protocol bypass, safe for production.

Unlimited concurrency · enterprise-ready

Linear scaling for batch generation, multi-image fusion, and sequence generation — no Tier-style account limits. Default 500 RPM, contact sales for higher quotas.

Same price + up to 20% off via top-ups

Default unit price matches BytePlus official pricing. Combined with top-up bonuses, the effective price drops to as low as 80% of list.

Global zero-friction access

No overseas server or proxy required. Connects directly to api.apiyi.com from mainland China data centers, residential networks, and overseas nodes. No need to set up routing for BytePlus ap-southeast-1 / eu-west-1 regions.

OpenAI-compatible · zero code change

The path /v1/images/generations is identical to OpenAI. Point the OpenAI SDK’s base_url at APIYI and call the API as-is. Pass extension parameters (image / sequential_image_generation etc.) via extra_body. Note that OpenAI’s n parameter is not supported upstream (silently ignored — you still get 1 image); use sequential_image_generation for multi-image output.

Professional support · enterprise concierge

Deep expertise in image-generation use cases — multi-image fusion, text rendering, batch asset production. End-to-end support from PoC to production rollout.

Key Features

4K high-fidelity output

4.0 / 4.5 support native 4K (4096×4096) with rich detail layers — ideal for posters and print. 5.0-lite tops out at 3K but offers a more refined overall experience.

Unified generation-editing

Text-to-image, single-image editing, multi-image fusion, and batch sequence all share one endpoint and one parameter set. Switch modes via image and sequential_image_generation.

Multi-image fusion · up to 10 references

image accepts a URL array. Refer to “image 1 / image 2” in the prompt for explicit ordering. Pair with sequential_image_generation: "disabled" for subject-consistent fusion.

Text rendering breakthrough

The 4.5 release dramatically improved small-text legibility. Posters, ad copy, and product text are clear and accurate — best in class.

Batch sequence (up to 15)

sequential_image_generation: "auto" plus max_images outputs a coherent series — perfect for storyboards, brand visuals, and product series.

≈ 15s per image · balanced speed

Typical single-image latency is ~15s; 4K + hd may extend up to a minute. 500 RPM default, scalable on request.

Flexible sizes · arbitrary aspect

Resolution presets (1K/2K/3K/4K) or exact pixels. Total pixels ∈ [1280×720, 4096×4096], aspect ratio ∈ [1/16, 16].

Drop-in OpenAI SDK

Set base_url=https://api.apiyi.com/v1 and call with the official OpenAI SDK. Extension params go through extra_body. Zero code change for migration.

Pricing

Per-image billing, same price as BytePlus official. Top-up bonuses lower the effective unit price further.

Model	APIYI Price	List Price (RMB est.)	Status
seedream-5-0-pro-260628	$0.12 / request (1 image)	≈ ¥0.84 / request	🆕 Pro tier (~2 min per image; use 5.0-lite for everyday work)
seedream-5-0-260128	$0.035 / image	≈ ¥0.245 / image	✅ Recommended (latest)
seedream-4-5-251128	$0.04 / image	≈ ¥0.28 / image	✅ Recommended
seedream-4-0-250828	$0.03 / image	≈ ¥0.21 / image	🟡 Maintained (still callable)

Billing notes:

• Billed per generated image, regardless of prompt length or fusion mode
• seedream-5-0-pro is billed at a fixed $0.12 per request (1 image per request; batch sequence not supported). Officially the model uses two output-pixel price tiers (one price for ≤2.36M px, one for >2.36M px) plus a per-image fee for reference images beyond the first; APIYI simplifies this to a flat per-request price — no tiers, input-image fees included. This model carries no official discount of any kind; APIYI prices it on a supply-guarantee basis — effectively at no margin once top-up bonuses and tax costs are factored in — and any price change will be announced in advance
• In sequential_image_generation: "auto" mode, billing is by actual output count (e.g. max_images: 4 outputs 4 → 4 billed)
• Failed requests (4xx / blocked by moderation) are not billed
• Free trial: 200 free images on first onboarding (provided by BytePlus)
• Top-up bonus details: see Top-up Promotions

Technical Specs

Dimension	seedream-5-0-pro	seedream-5-0	seedream-4-5	seedream-4-0
Model ID	seedream-5-0-pro-260628	seedream-5-0-260128	seedream-4-5-251128	seedream-4-0-250828
Model ID alias	—	seedream-5-0-lite-260128	—	—
Release date	2026-06-28 (UTC+8)	2026-01-28 (UTC+8)	2025-11-28 (UTC+8)	2025-08-28 (UTC+8)
Resolution tiers	1K / 2K + exact WxH (≤ 4.19M total px; at 16:9 the longest edge reaches 2720, ~2.7K)	2K / 3K	2K / 4K	1K / 2K / 4K
Output format	png / jpeg	png / jpeg	jpeg	jpeg
Prompt optimization	standard / fast	standard	standard	standard / fast
Text-to-image	✅	✅	✅	✅
Single-image editing	✅	✅	✅	✅
Multi-image fusion	✅ (up to 10)	✅	✅ (up to 10)	✅
Batch sequence	❌ (400 if passed)	✅	✅	✅
Streaming output	❌ (400 if passed)	✅	✅	✅
Max images per minute (RPM)	500	500	500	500
Single-request input + output	input ≤ 10, output 1	≤ 15	≤ 15	≤ 15
Typical latency	~2 minutes	~30s	10-20s	10-15s
Response field	same	data[].url or data[].b64_json	same	same

Generation Time Comparison

Measured single-request latency per version (measured 2026-07, UTC+8; wall-clock time from request to full response — expect normal per-request variance):

Scenario	seedream-4-0	seedream-4-5	seedream-5-0	seedream-5-0-pro
Text-to-image (1K/2K)	7-11s	8-13s	29-34s	110-130s
Text-to-image (top tier)	~15s (4K)	~18s (4K)	~37s (3K)	~134s (WxH 2720×1530, max size)
Editing / multi-image fusion	~11s	17-21s	38-40s	115-132s
Batch sequence (2 images, edit + auto)	—	~36s	—	— (not supported)
Recommended client timeout	≥ 60s	≥ 60s	≥ 120s	≥ 240s

seedream-5-0-pro consistently takes ~2 minutes per image (110-132s measured across every run, no exceptions). This is expected behavior for a deep-reasoning image model, not a fault. Before adopting pro, confirm your product can absorb the latency: interactive flows (user waiting on screen) are a poor fit — use 5.0-lite (~30s) instead; pro suits offline batch production where image quality and instruction adherence matter most.

API Endpoints

Endpoint	Purpose	Content-Type
POST /v1/images/generations	Text-to-image / single-image editing / multi-image fusion / batch sequence — all modes share one endpoint, switched via request body params	application/json

Domain choice: api.apiyi.com is the primary endpoint. vip.apiyi.com and other gateway domains are also available, with identical behavior. You do not need to use BytePlus native domains like ark.ap-southeast.bytepluses.com / ark.eu-west.bytepluses.com — APIYI normalizes everything to OpenAI-compatible paths.

Key Parameters in Detail

size (output size)

Two value families — pick one:

Preset tiers (model decides aspect ratio):

Tier	Approx pixels	Supported by
1K	~1024×1024	4.0 / 5.0-pro
2K	~2048×2048 (default)	all versions
3K	~3072×3072	5.0 only
4K	~4096×4096	4.5 / 4.0

Exact pixels (custom):

• Total pixels ∈ [1280×720, 4096×4096]
• Aspect ratio ∈ [1/16, 16]
• Default: 2048x2048

Valid examples: 1920x1080 (FullHD), 3840x2160 (landscape 4K), 1080x1920 (phone portrait), 2560x1440 (landscape 2K) Invalid examples: 5000x5000 (over the cap), 100x1600 (aspect under 1/16)

Sizes with total pixels above 4096×4096 return 400. Extreme aspect ratios (close to 1/16 or 16) may stretch unnaturally — prefer presets or common 16:9 / 9:16 / 1:1.

The 5.0-series models use a different exact-pixel range than 4.x (higher floor, lower cap). Out-of-range sizes return 400 with the valid range in the error message. Measured reference: 5.0-lite floor is around 2560×1440; 5.0-pro caps at 4.19M total pixels (max 2048×2048; at 16:9 the longest edge reaches 2720×1530 ≈ 2.7K, verified working) — there are no 3K/4K presets.

image and sequential_image_generation (mode switches)

The /v1/images/generations endpoint covers both text-to-image and editing/fusion. Two parameters together select the mode:

Mode	image	sequential_image_generation	Notes
Pure text-to-image	omitted	omitted or "disabled"	1 output
Single-image editing	["url1"]	"disabled"	Edit based on 1 reference
Multi-image fusion	["url1", "url2", ...]	"disabled"	Up to 10 references; refer as “image 1 / image 2”
Batch sequence	optional	"auto" + sequential_image_generation_options.max_images	N coherent outputs, N ≤ max_images and input + output ≤ 15

seedream-5-0-pro does not accept the sequential_image_generation parameter — passing any value (including "disabled") returns 400. For editing / fusion with the pro model, pass image only and omit the parameter entirely; the same applies to stream.

For full code examples, see Text-to-Image and Image Editing.

Best Practices

1. Pick the right version

   • Best overall experience → seedream-5-0-260128 (most features, but 3K cap)
   • 4K + strong text rendering → seedream-4-5-251128 (4K + text breakthrough)
   • 4K + best price → seedream-4-0-250828 (cheapest 4K)
   • Top image quality / complex instructions (professional work) → seedream-5-0-pro-260628 ($0.12/request, ~2 min per image, 1K/2K only — not recommended for everyday use)
2. Prefer preset sizes

   1K/2K/3K/4K are tuned by BytePlus for stable speed and quality. Use exact pixels only when you have a real aspect requirement. Note version differences in supported tiers.

3. Refer to images explicitly

   With multiple image URLs, write the prompt with explicit references — “Place the person from image 1 into the scene of image 2 using the colour palette of image 3” — instead of letting the model guess.

4. Control batch sequence cost

   sequential_image_generation: "auto" + max_images: 4 outputs 4 — billed × 4. Validate with max_images: 1 first, then scale up.

5. Pick output format by use case

   5.0 / 5.0-pro support png and jpeg; 4.5 / 4.0 only jpeg. Use the 5.0 series + png when you need transparent backgrounds or lossless detail, jpeg for size-sensitive scenarios.

6. Set client timeout ≥ 60 seconds

   Single image is ~15s, but batch sequence (4 images) or 4K + hd may take 30-60s. Start with a 60s client timeout and show progress feedback in the UI. seedream-5-0-pro takes ~2 minutes per image — use a timeout of 240s or more.

7. Disable watermark when needed

   Set watermark: false to remove the BytePlus watermark (defaults vary by version, so set it explicitly). Required for commercial assets.

Error Codes & Retries

Status	Meaning	Recommended action
400	Invalid parameters (size out of bounds, image array > 10, unsupported tier)	Validate against the supported tier of the version in use
401	Invalid token	Check Bearer Token
403	Blocked by content moderation	Adjust prompt or replace reference images
429	Rate limit (default 500 RPM) or insufficient balance	Exponential backoff; contact sales for higher RPM
5xx	Gateway / upstream error	Retry 1-2 times
Timeout	Long-tail request	Client timeout ≥ 60s (batch sequence or 4K + hd may reach 60s)

Client recommendations:

• Start with a 60-second request timeout (batch sequence or 4K + hd may take a minute)
• Apply exponential backoff for 5xx and timeouts (suggested 2 retries)
• Log the x-request-id response header for support tickets

FAQ

5.0 Pro / 5.0 / 4.5 / 4.0 — which to pick?

Need	Recommended
Latest features + best overall	seedream-5-0-260128
4K + strong text rendering (posters, ads)	seedream-4-5-251128
4K + best price	seedream-4-0-250828
Long-running stable batch production	seedream-4-0-250828 (proven, cheapest, fast prompt mode)
Top image quality / complex instructions (professional work)	seedream-5-0-pro-260628 ($0.12/request, ~2 min per image — skip unless you need it)

See Historical Versions for full comparison.

Why does image editing also use the generations endpoint?

Seedream uses a unified generation-editing architecture — there is no separate /v1/images/edits endpoint. Unlike OpenAI’s gpt-image-2 (multipart upload to /v1/images/edits), Seedream uses application/json and passes image URLs as an array in the image field.

Benefits: protocol consistency, parameter reuse, easy mode switching. See Image Editing for details.

Does the image field accept base64?

Yes (verified by testing). Use a data URI: data:image/<format>;base64,<base64 string> with the <format> in lowercase (e.g. data:image/jpeg;base64,...). URLs and base64 entries can be mixed in the same array. For large local images, uploading to an image host and passing a URL is still preferable to keep the request body small.

Multi-image fusion limit? Batch sequence limit?

• Multi-image fusion (image array): 4.5 / 5.0-pro explicitly support up to 10. 5.0 / 4.0 also support multi-image, though no explicit ceiling is documented.
• Batch sequence (max_images): bounded by the global rule input references + output ≤ 15. When combining fusion and sequence, count the total. Note 5.0-pro does not support batch sequence (passing sequential_image_generation returns 400).

Does b64_json need a data:image prefix?

Depends on response_format:

• response_format: "url" (default) → data[0].url is a temp signed URL, render directly with <img src=...>
• response_format: "b64_json" → data[0].b64_json is a plain base64 string (no data:image/...;base64, prefix). Decode and write to disk, or prepend the prefix manually for browser rendering.

Is streaming output supported?

Supported on 5.0 / 4.5 / 4.0 via stream: true. Streaming is especially useful for long prompts and high-resolution images — the frontend can render partial results progressively. seedream-5-0-pro does not support streaming — passing stream returns 400.

Rate limit?

Default 500 images per minute (Max Images per Minute), unified across versions. Contact sales for higher quotas.

Are failed requests billed?

No. BytePlus has built-in moderation. Moderation rejects and parameter errors return 400 / 403 and are not billed. Other zero-billed errors: 401 (invalid token), 429 (rate limited). Only successful generation (200 with valid response) is billed.

Can I use the official OpenAI SDK?

Yes, with no code changes. Point base_url at https://api.apiyi.com/v1 and pass extension params (image / sequential_image_generation / watermark etc.) via extra_body:

from openai import OpenAI

client = OpenAI(api_key="sk-your-key", base_url="https://api.apiyi.com/v1")
resp = client.images.generate(
    model="seedream-5-0-260128",
    prompt="...",
    size="2K",
    extra_body={
        "image": ["https://.../ref.png"],
        "sequential_image_generation": "disabled",
        "watermark": False,
    }
)

Who owns the generated images?

Generated images can be used commercially and non-commercially. See the BytePlus terms of service for details.

Are transparent backgrounds supported?

seedream-5-0 / seedream-5-0-pro support png output and can produce transparent backgrounds when prompted (“transparent background, alpha channel”). seedream-4-5 / 4-0 output only jpeg and do not support transparency — post-process to remove the background yourself.

Can I cancel a generation in flight?

No. /v1/images/generations is synchronous. Once submitted, the request runs to completion. The server still completes and bills even if the client disconnects. Set client timeouts and avoid the assumption that disconnect saves cost.

Related Docs

• Text-to-Image Playground — POST /v1/images/generations with five language code samples
• Image Editing Playground — image + sequential_image_generation patterns
• Historical Versions — 5.0 / 4.5 / 4.0 comparison and migration
• API Manual — general invocation guide
• Image Generation Sandbox — try it online
• BytePlus official docs: docs.byteplus.com/en/docs/ModelArk/1824121 — Seedream 4.0-5.0 tutorial

Seedream is provided through APIYI’s strategic partnership with BytePlus ModelArk. Three versions share a single integration, billing, and authentication path — pick by need. For questions or feedback, file a ticket from the console.