A model detail page for one product in this catalogue. It records what the catalogue declares, what the publisher documents, and where the two do not agree.
Updated:
Doubao Seedream 5.0 Pro
Choose a mode, write the prompt, then read the quoted cost before you generate.
Nine briefs the publisher published, in the language each was written in
Each of these is an image from the publisher’s own pages, shown with the prompt the publisher printed beside it. They are grouped here because they are the cases where a prompt was published in full: one Spanish brief, one Japanese brief, and the rest Chinese.
All nine images are the publisher’s own artwork, retrieved on 19 September 2026. Rights remain with the publisher; they are shown with source attribution and no claim of ownership. Every file arrived as published, so any marking it carries is the publisher’s and this site added none. Prompts are reproduced in the language the publisher wrote them in and are not translated, because a prompt is the model input and translating one would change the output. Two briefs among the publisher’s cases are published only in abbreviated form and are therefore not shown in this gallery; they appear among the cases below with captions that name no prompt.
Copied from the publisher’s page, where this prompt is printed beside this image. The wording is unchanged, including its original language, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page can accept it: the catalogue returns this product as available and priced.
Copied from the publisher’s page, where this prompt is printed beside this image. The wording is unchanged, including its original language, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page can accept it: the catalogue returns this product as available and priced.
Copied from the publisher’s page, where this prompt is printed beside this image. The wording is unchanged, including its original language, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page can accept it: the catalogue returns this product as available and priced.
Copied from the publisher’s page, where this prompt is printed beside this image. The wording is unchanged, including its original language, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page can accept it: the catalogue returns this product as available and priced.
Copied from the publisher’s page, where this prompt is printed beside this image. The wording is unchanged, including its original language, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page can accept it: the catalogue returns this product as available and priced.
Copied from the publisher’s page, where this prompt is printed beside this image. The wording is unchanged, including its original language, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page can accept it: the catalogue returns this product as available and priced.
Copied from the publisher’s page, where this prompt is printed beside this image. The wording is unchanged, including its original language, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page can accept it: the catalogue returns this product as available and priced.
Crea una infografía vertical 4:5 sobre la cultura del Día de los Muertos en México. Usa ilustraciones claras de altar, flores de cempasúchil, calaveras de azúcar, velas, pan de muerto y fotografías familiares.(制作一张关于墨西哥亡灵节文化的竖版 4:5 信息图。使用清晰的祭坛、万寿菊、糖骷髅、蜡烛、亡灵面包和家庭照片插画。)
Copied from the publisher’s page, where this prompt is printed beside this image. The wording is unchanged, including its original language, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page can accept it: the catalogue returns this product as available and priced.
Copied from the publisher’s page, where this prompt is printed beside this image. The wording is unchanged, including its original language, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page can accept it: the catalogue returns this product as available and priced.
A July 2026 model whose leading capability is not exposed here
ByteDance Seed published Seedream 5.0 Pro on 8 July 2026. The first capability in the publisher’s own list is layer decomposition, which was new in this version and lets one image be taken apart into a base layer plus up to sixteen separate layers. That control is not part of this catalogue’s form for the product.
The gap is not a matter of opinion about features. The publisher’s API reference for image generation lists thirteen body parameters. This site’s catalogue returns five form fields for the same product. Comparing the two lists is the single most useful thing a reader can do on this page, because four of the eight parameters that are missing are the ones that make this version different from the one before it.
This page therefore does two jobs. It describes what the product is, from the publisher’s own pages. It also records precisely which parts of that description the workspace on this site can reach today, and which parts it cannot.
The publisher count of thirteen is taken from the request-parameter list on the image-generation API page and counts the model field itself. The catalogue count of five is read from the live product record for site-19, whose form fields are prompt, image, size, response_format and watermark.
Surface
Parameters or fields
Includes layer decomposition
Publisher API reference, image generation, updated 9 September 2026
13 body parameters
Yes, boolean, default false
This catalogue, product doubao-seedream-5.0-pro, site-19
5 form fields
No
What the publisher says it does, under its own four headings
The release announcement organises the version into four claims rather than a feature list, and they are worth quoting as the publisher wrote them, because three of the four are about a single behaviour: holding a shape while something inside it changes.
Four capability names are quoted from the release blog. The publisher also states where the model is available at launch: the Volcano Ark experience centre immediately, with the Doubao and Jimeng apps to follow. Rolling out is the publisher’s own wording, not a statement that they are live.
Complex information visualisation: data, concepts and dense text turned into professional layouts, aimed at content with a high information density.
Interactive precise editing: point selection, lasso selection, sketch rendering, colour and material replacement, layer separation and multi-image fusion combined freely, described by the publisher as editing at pixel level.
Realistic photographic and portrait texture: light, material and skin grain, holding both computer-generated and photographic qualities.
Native multilingual input and generation: direct input and high-quality rendering in more than ten commonly used languages, carrying localised characteristics.
Thirteen parameters on one side of the table, five on the other
Listing both sets side by side is more useful than summarising the difference, because a reader who is deciding whether to use this product through this site needs to know which single control is absent, not that some are.
Two of the missing parameters are the ones that would change what a visitor could do rather than how the request is shaped. Layer decomposition is the version’s headline capability. Output format decides whether a layered result arrives as a transparent PNG, which is what makes a layer useful in a design tool at all.
The watermark parameter is present in both, but the defaults point in opposite directions. The publisher’s default is true and this catalogue exposes it as an unticked boolean, so the workspace and the API disagree about what happens when nobody touches the control.
Defaults are quoted from the publisher’s parameter table. The catalogue column records what the site-19 product record actually publishes, which was read on 19 September 2026. A parameter being absent from the form does not mean the underlying service rejects it; it means a visitor on this site cannot set it from the page.
Parameter in the publisher API
Publisher default
Present in this catalogue form
prompt
required
Yes
image
optional, single or list
Yes, capped at 10
layer_decomposition
false
No
size
not stated
Yes, as free text
optimize_prompt_options
not stated
No
output_format
jpeg
No
background
opaque
No
response_format
url
Yes, as a url or base64 choice
sequential_image_generation
disabled
No
sequential_image_generation_options
not stated
No
stream
false
No
tools
not stated
No
watermark
true
Yes, as a boolean
model
required, the model identifier
Set by the workspace, not by the visitor
Pro is narrower than the cheaper tiers on four counts
The same API page documents Seedream 5.0 lite, 4.5 and 4.0 in a second block, and that block grants those models capabilities the Pro tier does not have. On this page’s evidence, paying more buys precision rather than breadth.
The publisher states the Pro limits as a flat list of things the model does not do: set generation, web search and streaming output are all named as unsupported, and sequential generation cannot be configured at all. The lite block directly above grants set generation, which makes the inversion explicit rather than inferred.
That matters for anyone comparing the tiers on price alone. A cheaper image model in this family can return a group of related images from one brief and this one cannot, so the Pro tier is the wrong choice for storyboard grids and the right one for holding a single complex frame together.
Both columns are the publisher’s own text, taken from the two capability blocks on one page. The page does not explain why the flagship is narrower, and this page does not guess.
Behaviour
Seedream 5.0 Pro
Seedream 5.0 lite, 4.5 and 4.0
Images per reference set
2 to 10
2 to 14
Set generation
Not supported
Supported with sequential generation set to auto
Largest set from one brief
One image, since sets are unsupported
Up to 15 images, and reference count plus output count stays within 15
Web search during generation
Not supported
Available on lite
Streaming output
Not supported
Not listed as unavailable
What a layer-decomposition call returns
The publisher publishes a worked example for this feature, and the response body is more informative than the feature description, because it shows the shape of a layered result rather than promising one.
The request sends one image, a set of coordinates marking the text blocks and the subject, the boolean turned on, and a size of 2K. The response comes back as an array rather than a single image: the base layer at index zero, a full two-thousand-and-forty-eight-pixel square, followed by the separated layers at increasing indices.
Each separated layer carries its own bounding box in both absolute pixels and normalised coordinates, a name and a short description, which is what makes a layer addressable later. The usage block reports one input image and eight generated images for the call, so a single brief produced nine files in total counting the base.
All six rows are read from the response example on the publisher’s image-generation API page. The example uses a size of 2K in the request and returns a base layer of 2048 by 2048, which is the clearest statement on the page of what 2K means for this model. The layer count in this one example is eight; the documented maximum is sixteen.
Field in the published response
Value in the publisher’s example
Base image size
2048 by 2048, jpeg, index 0
First separated layer size
1273 by 265, png, index 1
Bounding box, absolute pixels
383, 120, 1655, 384
Bounding box, normalised
187, 59, 808, 188
Layer name and description
The masthead text, described as large yellow serif lettering
Usage for the call
One input image, eight generated images, 23107 output tokens
The size control a visitor actually sees
This is the part of the page that could only be established by reading the code and the live catalogue together, and it does not match what the model’s own configuration file says.
The workspace adds an aspect-ratio selector and a resolution selector from the local model table, but only when the catalogue has not already published a field of that kind. For this product the catalogue publishes size as a free-text field, so the aspect-ratio selector is never added; and it publishes no quality field, so the resolution selector is not added either. What a visitor sees is one empty text box labelled output size.
The local table is not silent on the subject. It declares a set of aspect ratios, three resolution options and a reference-image ceiling of fourteen. None of those three reaches the form, because each is superseded by what the catalogue publishes. The reference-image cap a visitor gets is ten, which is the catalogue figure and also the publisher’s stated maximum for multi-reference input.
Read from the live product record for site-19 and from the workspace code that merges catalogue fields with the local table. The rule the code applies is that a catalogue field wins and the local table only fills a gap the catalogue leaves. The consequence is worth stating plainly: on this page none of the three local declarations is in force.
Control
Local model table says
Catalogue publishes
What the form renders
Aspect ratio
A named ratio list, default auto
A size field, free text
No ratio selector; a text box
Resolution
Three options, default 2K
Nothing
No resolution selector
Reference images
Up to 14
Up to 10, at least 1
Up to 10
What the publisher says the model still gets wrong
The release announcement ends with a limitation the publisher wrote itself, and it is more specific than the usual forward-looking closing paragraph.
The limitation is the publisher’s own sentence. The timing and pricing figures are third-party and are reported as claims rather than facts: this page has not measured latency and cannot quote a resale price, because the publisher does not set one. Where two sources disagree the disagreement is the finding, and neither figure is presented here as the true one.
The publisher states that the model has made progress on complex infographics and interactive editing but that finer-grained text rendering and pixel-level edit retention still have room to improve.
A third-party review measured sixteen panels and found the model’s long-form body copy less reliable than its headline text, while noting it held factual detail in a geographic test that a rival model did not.
Two reviewers publish generation times that differ by more than an order of magnitude. One states about two seconds per image; another measured between twenty-six and eighty seconds with a median of fifty-two, describing this model as the slowest in its batch.
The same reviewers disagree about the cost of access, quoting two different per-image prices from the same third-party host.
The artwork on this page, and what it cannot show
Every image here comes from the publisher, and every one was opened and looked at before it was used. Nothing on this page is a user submission, a competitor render or a synthetic placeholder.
Screening was visual and per file, not by thumbnail. No kept image carries a visible AI label, a publisher watermark, a third-party wordmark or another site’s logo. The magazine masthead and the can marking in the ten-language poster are the publisher’s own brand.
Thirty images are used, all from the publisher’s release blog and project page. Six of them are before-and-after pairs, used as the comparison cases above, because the publisher publishes those edits as two separate files rather than one combined panel.
The prompts are reproduced in the language the publisher wrote them in, which for this model is mostly Chinese, with one Spanish brief and one Japanese brief. They are not translated. A prompt is the model’s input, and rewriting one into another language produces a different image rather than a translation of the same image.
The publisher publishes five demonstration clips alongside these stills, including a two-hundred-and-fifty-megabyte layer-separation clip. None is embedded here. The page is a stills gallery and embedding it would not add a case that the stills do not already carry.
Where the publisher publishes a long prompt only in abbreviated form, this page does not quote one at all. That is why two case captions above describe their image without a prompt: the publisher cuts those briefs off with a note saying how many characters were omitted, and quoting half a prompt as though it were one is worse than quoting none.
Examples
Before
After
The clearest single demonstration of what the publisher means by interactive editing. The input carries six purple boxes and handwritten English notes; the output keeps the layout, the furniture positions and the window light direction while restyling everything the notes touch. The interest is that the marks are spatial rather than verbal, which is the distinction the publisher draws between saying what to generate and saying where to change it.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
Before
After
An edit that changes what the subject is rather than how it looks. The publisher brief was to evolve the lizard into a crystal dragon while keeping the creature plausible, and the output holds the original silhouette, footprint and shadow while replacing the surface with faceted crystal. Keeping the animal anatomy is the part that is easy to get wrong, and the caption on the publisher page says so.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
Before
After
The brief asked for the whole frame to be reconstructed as a night chase scene rather than for a single object to be swapped, so the colour, the weather, the time of day and the depth of field all move together. It is the case on this page where the gap between the input and the output is widest, and it is the reason the publisher lists the capability as reconstruction rather than retouching.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
Before
After
A sketch-to-interface conversion with an explicit instruction to preserve the layout. The value for a reader is the constraint rather than the picture, because holding a rough layout while filling it with finished components is the part that usually drifts; the brief names the headline text exactly and asks for the structure to stay put.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
Before
After
A restyling case where the elements being changed are small in area but carry most of the meaning. Changing hair and makeup without moving the faces, the bodies or the court is a tighter task than it looks, and it is the same class of edit as the portrait-touch-up case the publisher also publishes.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
Before
After
The right-to-left case. Nothing is redesigned: the colour, the icons, the photograph and the module structure are held, and the text is translated and the reading direction is mirrored. The publisher treats this as a separate claim from generating text in many languages, because reversed reading order is a layout problem rather than a vocabulary problem.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
A cover composition that has to hold a masthead, three cover lines, a subject and a barcode without any of them colliding. It is included because a magazine cover is the layout most likely to expose a text renderer: the masthead is large, the cover lines are small, and both sit over a busy photographic subject.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
A menu is a test of two things at once: does the layout survive an odd page proportion, and does every price land beside the dish it belongs to. The publisher publishes this one in English while the Thai menu case on the same page is published in Chinese, so the pair shows the same house style in two scripts.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
Four languages in one poster, which is the densest multilingual case on the page. The interest is that the same four instructions have to fit the same column width in Chinese, English, Japanese and Korean at once, so the layout cannot be tuned for one language and quietly broken in the other three.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
An infographic about a subject with numbers that can be checked, which makes it the most falsifiable image on the page. Every figure, unit and percentage in a diagram like this is a place where a generator can be plausibly wrong, and the publisher does not claim otherwise.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
A product page built almost entirely from a single repeated object: bottle, apple, leaf, repeated at several scales down the page. It is the case where a generator most often loses count or changes the object between repetitions, so it is included as a shape that is easy to read at a glance.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
A wide action frame rather than a portrait. The interest is the water: droplets caught mid-flight, mud thrown up behind the shoes and rain on the ground all have to agree with one light source, which is the kind of physical consistency that studio-lit product shots do not test.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
The publisher publishes ten language versions of one layout as a single image, which is the most direct evidence on this page for the multilingual claim. Ten scripts is fewer than the dozen-plus languages the publisher claims elsewhere, and the image shows the ones chosen for display rather than the full supported set.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
A nine-by-sixteen education poster with more text than artwork. The publisher publishes this brief with an abbreviated prompt, so this caption does not quote one; the image is included for its shape, since a tall poster with several headed blocks is a different test from a single wide frame.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
Sixteen panels in one frame, each numbered and each part of a sequence, which is the strictest layout constraint on this page. The publisher publishes this brief with an abbreviated prompt, so the caption quotes no prompt; the image is included because panel-to-panel continuity in a grid is a different problem from making one strong picture.
ByteDance Seed, Seedream 5.0 Pro project page, retrieved 19 September 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
Questions
Does this page publish a price for Seedream 5.0 Pro?
It publishes the catalogue figure and nothing else. The site-19 catalogue returns seven points per image as a fixed unit price across all three task types, shown as a single rate rather than a range. That is the number the workspace quotes from. The page does not convert it into a currency and does not publish any third-party rate, because resale prices change and are not set by the publisher.
Can I actually generate an image with this model from this page?
Yes. The catalogue returns this product as available, with all three task types callable and priced, so the workspace can send a task. It pins the model when the page opens so nothing else has to be selected. This differs from some other model pages on this site, where the catalogue marks the product temporarily unavailable and the form renders but every mode is inert.
Why is there no aspect-ratio or resolution selector in the workspace?
Because the catalogue publishes a size field for this product and no quality field. The workspace only adds its own ratio and resolution selectors when the catalogue has left that kind of field out, and here it has not, so both additions are skipped and a single free-text size box is shown instead. The local model table does declare a ratio list and three resolution options, but neither reaches the form.
What is layer decomposition, and can I use it here?
It is the publisher’s feature for taking one image apart into a base layer plus up to sixteen separate layers, each with its own bounding box, name and description, so the pieces can be moved independently. The publisher documents a boolean parameter that switches it on. That parameter is not among the five fields this catalogue publishes, so it cannot be set from this page, and this is the clearest example on the page of a documented capability that the workspace does not reach.
How many reference images can I send?
Up to ten, and at least one. The catalogue publishes a reference-image limit of ten and the publisher states a range of two to ten for multi-reference generation, with a separate single-reference path that takes one image. The workspace uses the catalogue figure, so it accepts up to ten even in the modes where the publisher describes a two-image floor.
Does it return a set of related images from one brief?
No. The publisher names set generation as unsupported on this model and says the set-generation parameter cannot be configured at all. The same page grants set generation to Seedream 5.0 lite, 4.5 and 4.0, which can return up to fifteen related images with a reference-image count and an output count that together stay within fifteen. For storyboard grids and variants, the cheaper tiers in this family are the ones that can do it.
What is the largest image it can produce?
Two thousand and forty-eight pixels on a side, which the publisher calls 2K. The worked response example uses a 2K request and returns a base layer of 2048 by 2048. Nothing on the page documents 4K support, and although the local model table lists a 4K option, that list does not reach the workspace form, so the option is not offered to a visitor here.
Why are the prompts on this page not translated into other languages?
Because they are model inputs. The publisher wrote most of these briefs in Chinese, one in Spanish and one in Japanese, and each was written for the script it uses; the multilingual cases depend on the language of the brief. A prompt is the exact text that produced the image, so translating it would replace evidence with paraphrase. Each language version of this page therefore shows the same prompts in their original wording.