KIE Seedream 4.5 Text-To-Image [SL]

Seedream is ByteDance's closed text-to-image line - no open weights, ever, per the company's consistent "flagships stay on the API" policy. This node is the front door to the current one: a prompt in, a Seedream 4.5 image out, no local VRAM involved. It's the simplest image generator in this pack - nothing but a prompt, an aspect ratio, and a quality setting - which makes it the obvious first node to grab when you just want to test whether the KIE pipeline works at all.

Why would you pick Seedream 4.5 over Nano Banana Pro in the same pack? It's a different flavor. Nano Banana Pro is Google's strengths - instruction following, in-image text, reference-image editing. Seedream is ByteDance's - punchy, high-detail, sometimes more "designed" output, and it's the cheaper call at the basic quality tier. If you're doing a big grid of concept images, basic Seedream is a very cost-effective way to fill cells; high is for the one you promote to final.

The inputs that matter

• prompt - everything, up to 3000 characters.
• aspect_ratio - 1:1, 4:3, 3:4, 16:9, 9:16, 2:3, 3:2, 21:9 (default 1:1).
• quality - basic or high (default basic).

Output is a single image tensor - standard BHWC float32 - that plugs straight into SaveImage.

How it works

The simplest possible KIE job: no uploads, create a task for seedream/4.5-text-to-image, poll until it renders, download the image, decode to a tensor. Transient 429/5xx errors auto-retry. It's one of the faster jobs in the pack because there's nothing to upload.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/v3devv/ComfyUI-Kie-API-Serverless

or ComfyUI Manager search "Kie-API-Serverless", then restart. No model files, no dependencies beyond what ComfyUI already bundles. Paste your KIE key into the node's api_key field.

Where people get burned

The biggest surprise is moderation - Seedream is a ByteDance model and there's no local bypass for a model you never owned, so anything ByteDance refuses is a hard wall. Budget-wise, high is meaningfully pricier than basic for a quality bump that's often marginal at 1K; use basic for drafts. And while this node is text-only, the pack's edit node is the same model family with an image input - if you find yourself re-prompting the same scene repeatedly, switch to an edit workflow instead of paying for fresh generations.