Seedream 4.5 API and Pricing, Explained Through a Real Pipeline
Seedream 4.5 is ByteDance's flagship image model: 4K output, targeted edits on an image you already have, and composition from several references at once. Here is how it meters, what an image really costs, what the free and uncensored questions actually mean in practice, and how to reach it through a published API endpoint.
Seedream 4.5 is ByteDance's flagship image model, and most of the questions people arrive with are not about how it looks. They are about what it costs, whether it can edit a photo they already have, whether there is a free way to use it, and what it will refuse. Fair questions, none of them answered by a gallery of sample images.
This post covers the real spec sheet, how image generation is metered, the two settings that move your bill, and how to call Seedream 4.5 as a published API endpoint without writing a ByteDance integration first.
What you get before you talk about price
Seedream 4.5 is the high end of the image catalog rather than the fast lane. The numbers on its spec sheet are the reason:
| Capability | Seedream 4.5 |
|---|---|
| Resolution | 1K, 2K, or 4K |
| Aspect ratios | 18, from 1:1 through 21:9 and 9:21, plus auto |
| Input | Text and image |
| Reference images | Up to 10 per node on the canvas |
| Batch per node | 1 to 4 images, each charged separately |
The resolution ceiling is the headline: generations come back with enough pixels for print, large-format placement, and aggressive crops, which removes the upscaling step that usually ends an image workflow. The aspect ratio list is the quieter win. Eighteen ratios including 21:9 and 9:21 means a banner, a vertical Story frame, and a square thumbnail come out of the same node with nobody cropping by hand. Full details are on the Seedream 4.5 model page.
The edit path is the part people miss
Seedream 4.5 takes image input, not just text. That opens two modes that a text-to-image model cannot do at all.
The first is targeted editing. Give it an image and an instruction and it changes the thing you asked about while leaving the rest of the frame alone. Replace the sky with storm clouds, and the foreground stays where it was. That is why a lot of people search for the model at all: they have the asset and want one part of it different, not a fresh generation that lands nearby.
The second is multi-reference composition. The image node accepts up to ten references, and Seedream resolves them into one coherent frame: this product, that style, this scene. That is a different operation from "roughly match this style." All of the references are meant to survive into the output.
Both modes run through the same node: wire an image into its reference input, or drag images onto it, and write the instruction as the prompt. One useful edge case is that leaving the prompt empty passes a wired or dropped image straight through as a reference source for whatever comes next, at no charge. That is how you feed an existing asset into a graph without paying to regenerate it.
How image generation bills, and which settings move it
There is no flat per-image sticker price to quote, and any post that gives you one is guessing. Image models meter on the provider side by what you asked for, and the two settings you were already choosing are the levers on your bill:
- Resolution. 1K, 2K, and 4K are not the same charge. Four thousand pixels wide is a materially larger generation than one thousand, and the price follows the pixels. If the asset is going to be viewed on a phone, 4K is money spent on detail nobody will see.
- Batch size. The image node generates 1 to 4 images in parallel, and each one counts as a separate charge. A batch of four is four generations, not a discount bundle.
So draft at 1K, decide, then re-render the winner at 4K. It is the same discipline as drafting video on a cheap model before a premium render, and it applies harder to images because the loop is so much faster. Gemini 2.5 Flash Image is the node default for exactly the drafting half of that loop. We compared the two in Seedream 4.5 vs Gemini 2.5 Flash Image: Seedream wins every row with a number attached, and Gemini stays the better model to think in.
What you actually pay on Treza
Seedream 4.5 through a Treza pipeline is metered from a prepaid credit balance. One credit equals $0.01, packs start at $5, credits never expire, and there is no subscription of any kind. A failed generation charges nothing.
For scale rather than a worst case: across billed image generations in a production week ending 2026-07-19, measured over the whole image catalog rather than any single model, the mean charge was $0.04 and the highest single charge was $0.23. Images are the cheap part of a media pipeline. The expensive node is almost always the video one, worth remembering when you decide how much to spend on the still that feeds it.
The billing detail that matters most for iteration is the failure rule. If a generation errors out or the provider refuses the prompt, you are not charged. That is what makes exploratory work on an expensive model reasonable rather than nerve-racking.
Is Seedream 4.5 free?
Not on Treza, and not honestly anywhere running the real model for you. Building on the canvas costs nothing: lay out a graph, wire nodes together, save it, spend nothing. Generation spends real provider money, so it draws from your balance. What Treza does not do is charge a monthly fee for the privilege. No plan, no seat minimum, and no month-end forfeiture of credits you paid for and did not use. A month where you generate nothing costs you nothing, which for spiky creative work is the whole difference.
The moderation question, answered straight
People ask whether Seedream 4.5 is censored or whether there is an unfiltered version. The factual answer: Seedream 4.5 is ByteDance's model and it applies ByteDance's moderation on their side of the API. Treza does not run an uncensored mode, cannot disable another company's input moderation, and does not offer a model that promises to. If the provider refuses a prompt, the generation fails and you are not charged.
What we do handle is the recovery. A refusal partway through a multi-node run does not force a re-run from the top. Edit the prompt on the refused node and resume, so the expensive nodes that already succeeded are not paid for twice. This comes up more with video than images: Seedance 2.0 rejects reference images containing recognizable real people, and the fix is to swap that one node to a model that accepts them rather than abandon the run.
Calling Seedream 4.5 as an API without building the ByteDance integration
Direct API access means provider credentials, their request shape, their error semantics, and your own retry and fallback logic before you have written a line of your product.
Through a Treza pipeline, Seedream 4.5 is a node on a canvas instead. You place the image node, pick the model, set resolution and aspect ratio, wire it into the chain, and publish that chain as a single versioned endpoint. Your app calls the typed /invoke route with JSON in and JSON out, or points an existing OpenAI SDK at the OpenAI-compatible route. Swap the model on the node later, drop in FLUX.2 Pro for a different look, and the endpoint your product calls does not change.
The same pipeline is reachable from an agent. Treza's hosted MCP server exposes pipelines as tools that Claude and other MCP clients call directly, so "generate the key visual at 4K and hand me the URL" is something an agent can carry out rather than describe.
Putting it in a pipeline instead of a single call
A single image is rarely the deliverable. The pattern Seedream is best at is being the first node rather than the only one.
Art direct a key frame in Seedream 4.5, then pass it to a video node as a first-frame keyframe. Seedream and Seedance are siblings from the same lab and share an aesthetic, so the still and the motion stay in one visual family instead of drifting apart. That two-node graph is the backbone of most production work here. A fuller video pipeline puts a script step ahead of it and narration, captions, or publishing behind it, then reruns the whole thing on a schedule or from an API call without anyone opening the canvas again.
Starting from blank, the AI image generator page has the image side set up, and the AI video generator page has the template that puts a video model behind a single generate action. For the longer walkthrough, see how to build an AI image generation pipeline.
Frequently Asked Questions
How much does a Seedream 4.5 image cost?
There is no single flat rate, because image models meter by what you asked for. Resolution and batch size are the two settings that move the charge. For scale: across billed image generations in a production week ending 2026-07-19, measured across the whole image catalog, the mean charge was $0.04 and the highest single charge was $0.23.
Can Seedream 4.5 edit an image I already have?
Yes. It accepts image input as well as text, so an existing image plus an instruction changes the part you asked about and leaves the rest of the frame intact. On the canvas, wire the image into the image node's reference input and write the edit as the prompt.
How many reference images can Seedream 4.5 use at once?
The image node accepts up to ten, and Seedream composes them into one frame rather than treating them as loose inspiration. That is what makes product-plus-style-plus-scene briefs work in a single generation.
Is there a free version of Seedream 4.5?
Building pipelines on the canvas is free, but generation spends real provider money and is charged from a prepaid credit balance. There is no subscription, packs start at $5, credits never expire, and a failed generation charges nothing.
Is Seedream 4.5 censored?
The model applies ByteDance's own moderation on the provider side, and no platform reselling it can turn that off. Treza does not offer an uncensored mode. A refused prompt fails without charging you, and in a multi-node run you can reword that node and resume rather than re-run the graph.
Does Seedream 4.5 support 4K output?
Yes. It generates at 1K, 2K, or 4K, and 4K is the reason to reach for it over a flash-tier model. If your workflow currently ends in an upscaling step, Seedream usually removes it.
Can I call Seedream 4.5 through an API without a ByteDance account?
Yes. On Treza the model is a node inside a pipeline you publish as a versioned HTTP endpoint, with a typed /invoke route and an OpenAI-compatible route, so your product calls one API regardless of which model sits behind the node.
