Seedream 5.0

Overview

Seedream 5.0 is ByteDance Seed's current image generation model line. ByteDance's own one-line description of the Lite tier is precise and worth quoting in full:

"A unified multimodal image generation model endowed with deep thinking and online search capabilities, featuring an all-round upgrade in its understanding, reasoning and generation capabilities."

Seedream 5.0 Lite was announced on February 13, 2026. ByteDance's launch post says only that the model "is live on platforms such as Dreamina AI and Volcano Ark Experience Center," without dating that rollout; the Ark snapshot is doubao-seedream-5-0-lite-260128, whose suffix denotes January 28, 2026. It is the first Seedream that reasons before it renders, and the first that can ground a generation in a live web search.

ByteDance frames the shift as a change in the model's role rather than its resolution. Seedream 5.0 Lite moves from a system that "passively responds to instructions" to "a professional creative assistant capable of observation, comprehension, and logical reasoning." ByteDance organises the model's deep thinking around four verbs: reading, seeing, drawing, and writing.

A Seedream 5.0 Pro tier was announced on June 23, 2026 at the Volcano Engine FORCE conference. See the Seedream 5.0 Pro section below for an important caveat about its sourcing.

Capabilities

From ByteDance Seed's launch post and product page:

• Deep thinking — multi-step visual reasoning before generation, aligned with physical laws. The model reasons about a scene rather than pattern-matching a prompt to pixels.
• Online search — live internet retrieval, which ByteDance says lets the model "more effectively respond to time-sensitive generation needs." The feature is toggleable, on or off, per request.
• Information visualisation — turning complex concepts into diagrams and explanatory graphics.
• Style transfer via reference images — applying the look of a supplied reference to new content.
• Advanced image editing from vague instructions — the model interprets underspecified edit requests rather than requiring precise masks and coordinates.
• Complex multi-subject generation — scenes with several interacting subjects.
• Precise style and layout control — attention to compositional detail, per the product page.
• Single-image editing and text-to-image in one unified architecture.
• Intelligent interpretation of user intent — inferring what the user meant, not only what they typed.

The online search capability is the genuinely novel one. Prior Seedream releases could only draw on what was in their training data; Seedream 5.0 Lite can look something up first.

Technical Specifications

ByteDance Seed's blog publishes little hard specification, but Volcano Engine's API reference and pricing page publish the operational limits. We list only what the vendor stated.

• Model line: Seedream 5.0
• Tiers: Lite (documented), Pro (announced, not yet documented — see below)
• Ark model ID: doubao-seedream-5-0-lite-260128
• Type: unified multimodal image generation model — text-to-image, image editing, and generation share one architecture
• Distinguishing features: deep thinking; toggleable online search
• Availability: Dreamina AI (Jimeng AI), Volcano Ark Experience Center, Volcano Engine API
• Weights: closed
• Maximum resolution: 4K. The size parameter accepts 可选值：2K、3K、4K ("optional values: 2K, 3K, 4K").
• Default size: 2048x2048 pixels
• Explicit dimensions: total pixels must fall within [2560x1440 = 3686400, 4096x4096 = 16777216], and the aspect ratio within [1/16, 16]. Both constraints must hold simultaneously.
• Largest documented outputs: 4096x4096 (1:1) and 6240x2656 (21:9)
• Pricing: ¥0.22 per generated image
• Rate limit: 500 IPM (images per minute)
• Absolute Elo score: Not published by ByteDance. Only the relative claim against Seedream 4.5 is official.

8K is genuinely unsupported. The total-pixel ceiling of 16,777,216 is exactly 4096x4096, so no 8K output is reachable through the API by either the named-resolution or explicit-dimension route.

ByteDance's own characterisation of the Lite tier's scale: "Seedream 5.0 Lite is a relatively small model." Its launch post also frames the release as "not higher resolution or faster speed, but the deep thinking behind 'reading', 'seeing', 'drawing', and 'writing'."

Seedream 5.0 Pro

Sourcing caveat, stated plainly: Seedream 5.0 Pro was announced on June 23, 2026 at the Volcano Engine FORCE conference in Beijing, alongside Seed 2.1 Pro, Seedance 2.5, and an audio generation model.

The conference itself is documented on a ByteDance property, not merely in the trade press: Volcano Engine's own FORCE 2606 event page confirms the dates (2026.06.23–06.24), the venue (Beijing, 国家会议中心二期 — China National Convention Center Phase II), and a headline billing of 豆包大模型 全新重磅发布 ("Doubao large model, major new release"). That page does not, however, name Seedream 5.0 Pro anywhere.

As of July 8, 2026, ByteDance has published no product page, technical blog post, model card, model-directory entry, API reference, or price for Seedream 5.0 Pro. ByteDance Seed's model directory still lists only Seedream 5.0 Lite, Seedream 4.5, and Seedream 4.0. So while the conference is primary-sourced, the Pro tier's existence as a named product is not, and its capabilities certainly are not. We record the announcement and its date, and nothing further. Capability claims, resolution figures, and pricing for the Pro tier circulating on third-party sites are speculative, and we do not reproduce them.

One adjacent primary fact, easy to misread as evidence of a Pro tier: Volcano Engine's model list does carry a non-Lite model ID, doubao-seedream-5-0-260128, noted as sharing a row with doubao-seedream-5-0-lite-260128. Its 260128 suffix dates it to January 2026, five months before FORCE, and Volcano Engine publishes no price for it. It is not a "Pro" tier, and should not be cited as one.

The FORCE announcements together sketch a full multimodal production stack under Volcano Engine — Seed 2.1 for language and agents, Seedream 5.0 Pro for images, Seedance 2.5 for video, and an audio model — but the details of the image tier remain undocumented at the primary source.

Use Cases

• Time-sensitive visual content — the online search grounding is designed for exactly this: generating imagery that reflects current events, trends, or real-world specifics the model was never trained on.
• Information graphics and explanatory diagrams — ByteDance highlights information visualisation from complex concepts as a first-class capability.
• Instruction-driven photo editing — editing from vague natural-language instructions rather than precise masks.
• Style-consistent asset production — applying a reference image's style across a set of generations.
• Multi-subject scene composition — layouts with several interacting subjects, where naive models lose track of who is doing what.
• Marketing and social creative — the Dreamina/Jimeng consumer surface is aimed squarely at this.
• Physically plausible scenes — the deep-thinking pass is explicitly aligned with physical laws, targeting the classic failure modes where objects float or clocks have impossible hands.

Performance / Benchmarks

ByteDance publishes relative claims, not absolute numbers. We reproduce them as stated and add nothing.

MagicArena (ByteDance-reported)

• Seedream 5.0 Lite "has made significant progress in key fundamental metrics, including overall Elo score, instruction following, and editing consistency."
• Its Elo score exceeded Seedream 4.5, with "significant improvements in knowledge reasoning, editing response, and consistency."

MagicBench (ByteDance's internal multi-dimensional evaluation)

• "Significant improvement across core dimensions including prompt following and alignment."

That is the complete set of performance claims ByteDance has made public. No absolute Elo score, no percentage, no head-to-head figure against a competing model has been published.

Precise-looking numbers attributed to Seedream 5.0 elsewhere — prompt-adherence percentages, typography accuracy rates, visual-logic scores out of ten, generation latencies — do not appear in any ByteDance source. We omit them.

Limitations

ByteDance names the first three itself, in its own launch post — a level of candour worth noting:

• Structural stability — room for improvement, per ByteDance.
• Realism — room for improvement, per ByteDance.
• Aesthetics — room for improvement, per ByteDance.
• It is a small model. ByteDance describes Seedream 5.0 Lite as "a relatively small model." The forthcoming Pro tier presumably addresses this, but it is undocumented.
• Resolution caps at 4K. The API's total-pixel ceiling is 4096x4096, so there is no 8K path. There is also a minimum: explicit dimensions below 2560x1440 total pixels are rejected, which rules out small thumbnails in a single call.
• No open weights. No self-hosting, no local fine-tuning, no offline generation. For an open-weight alternative see Stable Diffusion 3.5.
• Search grounding introduces a dependency. Web-grounded generation inherits whatever is wrong, stale, or missing in the retrieved results.
• Pro tier is announced but undocumented. Do not build a plan around Seedream 5.0 Pro's capabilities until ByteDance publishes them.
• Content policy constraints. As a Chinese-market model, generation is shaped by content rules that differ from those governing US-hosted models.

Pricing & Access

ByteDance publishes per-image pricing for the Lite tier on Volcano Engine's 模型价格 (model pricing) page, under 图片生成模型 ("image generation models"). The column header is 单价 元/张 — "unit price, yuan per image."

Billing is 按成功输出图片数量计费 ("billed by the number of successfully output images"). Volcano Engine adds two clarifications: 组图场景按实际生成的图片数量计费 ("in image-set scenarios, billing follows the number of images actually generated") and 因审核等原因未成功输出的图片不计费 ("images not successfully output, for reasons such as moderation, are not billed").

No price is published for a Seedream 5.0 Pro tier, because no such product is documented.

Access routes, as stated by ByteDance:

• Dreamina AI (known as Jimeng AI in China) — the consumer creative app where Seedream 5.0 Lite is live.
• Volcano Ark Experience Center — ByteDance's model playground, where the model can be tried directly.
• Volcano Engine API — the programmatic route, via the Ark platform.

There are no open weights and no self-hosted deployment option. Rates change; confirm against the Volcano Engine pricing page before budgeting.

Ecosystem & Tools

• Dreamina AI / Jimeng AI — ByteDance's consumer creative application, and Seedream's main consumer surface.
• Volcano Ark — model platform, Experience Center, and API.
• Seedream 4.5 and 4.0 — the preceding generations, both still listed in ByteDance Seed's model directory.
• Seedance — the sibling video generation line. Seedance 2.5 was previewed at the same FORCE conference that announced Seedream 5.0 Pro.
• Seed 2.1 (Doubao) — the language and agent flagship in the same Volcano Engine stack.
• Seed1.5-VL — image understanding, as distinct from Seedream's image generation.

Community & Resources

• Deeper Thinking, More Accurate Generation: Introducing Seedream 5.0 Lite — the official launch post, and the source of every capability and MagicArena claim on this page
• Seedream 5.0 Lite product page — official description and API access
• ByteDance Seed Models directory — canonical model list; lists Lite, 4.5 and 4.0, but no Pro tier
• Volcano Engine image generation API reference — the size parameter, 2K/3K/4K, and the pixel and aspect-ratio ranges
• Volcano Engine model pricing — the ¥0.22-per-image rate
• Volcano Engine FORCE 2606 event page — ByteDance's own page for the conference at which the Pro tier was announced
• ByteDance Seed — research team homepage