Seedream 4.0
Last reviewed
May 31, 2026
Sources
11 citations
Review status
Source-backed
Revision
v2 · 1,958 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 31, 2026
Sources
11 citations
Review status
Source-backed
Revision
v2 · 1,958 words
Add missing citations, update stale details, or suggest a clearer explanation.
Seedream 4.0 is a unified image generation and editing model built by the Seed team at ByteDance. Released on September 9, 2025, it folds into one network the two jobs that earlier ByteDance models handled separately: making pictures from a text prompt, and editing existing pictures from written instructions. ByteDance describes it as a single, unified architecture that joins text-to-image generation with image editing. The model produces output at resolutions up to 4K, returns a 2K image in roughly 1.8 seconds, and reaches users through ByteDance consumer apps such as Doubao and Dreamina along with the Volcano Engine developer API.[1][3][4]
Seedream 4.0 arrived in a fast moving stretch for image generation, and within days of release it sat at the top of public preference leaderboards. It ranked first on both the text-to-image and image editing boards run by LMArena, and on the matching boards from Artificial Analysis, ahead of Google's Nano Banana, the model branded Gemini 2.5 Flash Image, and competing with OpenAI's GPT Image.[2][3] It sits in ByteDance's line of text-to-image systems as the successor to Seedream 3.0 and the SeedEdit editing models.[4]
ByteDance has shipped its image work under two related names. The Seedream models cover text-to-image generation, while SeedEdit covered instruction based editing of an image that already exists. Seedream 3.0, released earlier in 2025, brought native 2K output and faster sampling than the version before it. The Seed group also works on video, including the Seedance model, so the image releases sit inside a broader media research effort rather than standing alone.[1][4]
Seedream 4.0 changes the structure by dropping the split between generation and editing. Instead of one model that paints from scratch and a separate one that revises, a single network does both, pulling together the generation work of the Seedream line and the editing work of SeedEdit.[1][4] Coverage of the launch treated that merge as the headline change in this version, since it lets one model do work that used to need two.
The timing put it straight into a race with Google. Nano Banana, Google's Gemini 2.5 Flash Image, had been released on August 26, 2025, and that model went viral for its natural edits and steady characters.[10] Seedream 4.0 followed about two weeks later, so the two were compared at once, and the public arenas became the place where that comparison played out.[2][3]
The practical payoff of the unified design is that one prompt can do work that used to need two tools. A user can ask for a fresh image, then in the same model ask to recolor it, swap a background, add or remove an object, change a pose, or restyle the whole frame, all through plain language.[1][4] Because generation and editing share the same network, the model carries context between those steps instead of treating an edit as an unrelated new request.
This matters for jobs that mix the two modes. Product mockups, marketing variations, and storyboards often start with a generated base and then need many small revisions. Keeping both in one model cuts the round trips and helps the look stay consistent from the first draft to the last edit.[4]
ByteDance also points to capabilities that go past plain rendering. The model is pitched for knowledge based generation, meaning prompts that need some world knowledge or reasoning, such as building a diagram or an educational visual where the content has to be correct rather than only good looking.[4] Its text rendering improved over earlier versions, which helps with posters, ads, and layouts that carry typography. It also supports in context generation, producing a linked series of images that hold a narrative or a style across the set, which fits comic panels, tutorials, and step by step sequences.[4] Because all of this runs in one model rather than a chain of separate tools, a single conversation can move from an idea to a finished set without exporting an image and loading it into another program, which is part of why the launch framed the unified design as the main advance.[1][4]
Seedream 4.0 generates images at up to 4K, meaning dimensions as large as 4096 by 4096 pixels, with a working range that starts around 1024 by 1024.[3][7][8] Native high resolution is a real difference from many rivals, which often top out lower and lean on a separate upscaler to reach print sizes. The 4K ceiling means a single generation can be used at poster or large screen size, while the 2K and lower settings trade some detail for speed when a draft is enough.[3][7]
Speed is the other selling point. ByteDance reported that the model returns a 2K image in about 1.8 seconds, and it claims the new architecture runs roughly ten times faster than Seedream 3.0.[3][4] Larger 4K renders take longer, and real latency depends on the platform and the load at the time, but the headline figure points to a model tuned for interactive use rather than slow batch runs.
Beyond single prompts, Seedream 4.0 accepts reference images so the output can match a given subject, character, or style. ByteDance documents multi-reference generation, and write-ups of the release note support for up to six reference images, while third party API hosts such as Replicate allow up to ten input images in one request.[4][7] In a single call the model can return up to 15 images, which suits grouped or sequential sets.[4][7]
That multi image support feeds the consistency features. Given a reference subject, the model tries to hold that subject steady across a batch, so a character keeps the same face and outfit across several scenes, or a product keeps the same shape across several angles. Generating a related group in one call, rather than one image at a time, helps that consistency hold.[4][7]
Seedream 4.0 reaches users through several ByteDance channels. Consumers can use it inside Doubao, ByteDance's assistant app, and inside Dreamina, the creative tool known in China as Jimeng.[1][4] Developers and businesses call it through the Volcano Engine API, ByteDance's cloud platform, where it appears as a Doubao Seedream model.[9] ByteDance also offers the model internationally through BytePlus and its ModelArk service, the overseas counterpart to Volcano Engine, so teams outside China can call the same model.[3][6]
The model is available through third party inference hosts as well. Services such as fal.ai and Replicate expose both the text-to-image and the editing endpoints, and both list a price of about 0.03 US dollars per generated image, which works out to roughly 30 dollars per 1,000 images.[3][7][8]
| Attribute | Detail |
|---|---|
| Developer | ByteDance, Seed team |
| Released | September 9, 2025 |
| Type | Unified text-to-image generation and image editing |
| Output resolution | Up to 4K (4096 by 4096 pixels), from about 1024 by 1024 upward |
| Reported speed | About 1.8 seconds for a 2K image |
| Reference images | Multiple (up to 6 documented; up to 10 on some API hosts) |
| Outputs per request | Up to 15 images |
| Indicative price | About 0.03 USD per image (about 30 USD per 1,000) |
| Availability | Doubao, Dreamina (Jimeng), Volcano Engine, BytePlus ModelArk |
| Predecessors | Seedream 3.0 (generation), SeedEdit (editing) |
The clearest external signal for Seedream 4.0 came from public preference arenas. These run blind pairwise tests, where people pick the better of two images without seeing which model made each, and the votes feed an Elo style score that ranks the systems. Soon after release in September 2025, Seedream 4.0 took the number one spot on both of LMArena's image boards, the text-to-image arena and the image edit arena, surpassing the previous leaders including Nano Banana.[2][5] Artificial Analysis reported the same outcome on its two image leaderboards, text to image and image editing, again placing Seedream 4.0 first and ahead of Google's model.[3][11]
Topping both the generation and the editing boards at once is the notable part. Many models do well at one or the other but not both, so leading the two together backed up the claim that the unified design did not trade one skill for the other.[3] These rankings reflect human preference on everyday prompts rather than a fixed accuracy score, so they shift as new models arrive, but at launch Seedream 4.0 sat at the front.[2][3]
| Model | Developer | Released | Unifies generation and editing | Main access |
|---|---|---|---|---|
| Seedream 4.0 | ByteDance | September 2025 | Yes | Doubao, Dreamina, Volcano Engine API |
| Nano Banana (Gemini 2.5 Flash Image) | August 2025 | Yes | Gemini app and API | |
| GPT Image (gpt-image-1) | OpenAI | 2025 | Yes | ChatGPT and OpenAI API |
Against ByteDance's own past work, Seedream 4.0 is both a merge and a step up. Seedream 3.0 handled generation and reached 2K, while SeedEdit handled instruction based edits. Version 4.0 puts those into one model, pushes resolution to 4K, and speeds up generation, so it replaces two tools with one that aims to beat each on its own ground.[3][4]
Against outside models, the comparison usually runs through Nano Banana and GPT Image, since all three pair generation with conversational editing. Nano Banana earned wide praise for natural edits and character consistency, and Seedream 4.0 was measured directly against it on the arenas, where it took the lead.[2][3] GPT Image, OpenAI's model behind picture creation in ChatGPT, is the other common reference point.[4] Seedream 4.0's distinct pitches in that group are its native 4K output, its fast generation, and a price near three cents per image, alongside arena results that put it ahead of the best known systems at launch.[3][4] Several writeups read those results as a sign that a Chinese lab had matched or passed the leading Western image models on everyday prompts, and done it at a lower price per image.[3][5]
Seedream 4.0 is a closed model. ByteDance has not released its weights, so it is reachable only through company products and the API, which rules out local or fully offline use and leaves access subject to regional rollout and platform terms.[1][6] Like other image models, it can still struggle with precise text inside an image, with fine structural detail, and with edits that are meant to leave the rest of a picture untouched, and its outputs pass through safety filters that block some requests. The headline speed applies to 2K output, since larger 4K renders take more time, and the per image price adds up across high volume use or many variations.[3][7] Arena standing, finally, is a snapshot of human preference at one moment, and the quick pace of image research means leaderboard positions can change as competing models update.[2][3]