ByteDance Seedream 4.5 & 5.0

Let's get the elephant in the room out first: despite the name and the "4K" in the description, this node does not run Seedream on your machine. There's no model file, no VRAM cost, no checkpoint to download. Seedream is ByteDance's closed flagship image line - the same models you'd call from their API or hit in the Doubao app - and this node is just a well-wrapped doorway to it that lives inside ComfyUI itself.

That's the whole deal with ComfyUI's "partner nodes": a family of API-backed nodes that ship in core, bill you through a Comfy account instead of a vendor key, and drop the result back into your graph as a normal IMAGE tensor. So you can generate with Seedream, then feed the output into your local upscaler, a ControlNet pass, or a Wan video workflow without ever leaving ComfyUI. For a beginner this is the easiest closed-model on-ramp there is: Settings → User to log in, Settings → Credits to top up, queue, done.

What it actually does

Seedream is a unified model - one endpoint does text-to-image, single-sentence editing, and reference-based generation. The node leans into all three. Feed it just a prompt and you get a generation. Feed it an image plus something like "make her wear a red coat" and it edits. Feed it one or more reference images and it can do character/style-consistent multi-image work.

The model dropdown is where you pick your tier: seedream 5.0 pro (the current flagship), 5.0 lite, 4.5, or 4.0. Pro bills by resolution - roughly 9.5 credits for 1K, ~19 for 2K - so the fancier tier genuinely costs more per run.

The inputs that matter

• prompt - your text, for creating or editing. Keep the editing phrasing sentence-shaped; Seedream's "precise single-sentence editing" is the party trick.
• size_preset - recommended sizes out of the box. Pick Custom and the width/height fields wake up (1024 to ~6240 wide, 1024 to ~4992 tall, step 2). Defaults are 2048×2048, which is already well past local-SDXL territory.
• image - optional. One image for img2img, several for multi-reference work.
• sequential_image_generation - set to auto and the model decides whether to produce a small series of related images (story scenes, character variations) instead of one. max_images caps it; total input+output can't exceed 15.
• seed, watermark (an "AI generated" tag), and fail_on_partial (abort if any requested image comes back missing - leave it on unless you like silent failures).

Output is a single IMAGE, which wires into Preview Image, Save Image, or anything else that eats an image tensor.

Getting it and gotchas

It ships with ComfyUI core - no Manager, no install. But you need a Comfy account with credits, and you should run a recent ComfyUI; the partner-node roster changes fast and the docs flat-out recommend the nightly build. Two real-world pains to know about: login is only allowed from localhost unless you set up the Comfy account API key for LAN access, and - the big one - the content filtering on the API side has changed under people before. There was a stretch in 2025 where Seedream-via-Comfy went from permissive to heavily censored overnight and users with pre-bought credits got burned. It settled down, but the lesson stands: this is a hosted service, the filter is theirs, not yours, and credits are prepaid and non-refundable. If you need guarantees, that's what the open-weight ecosystem is for.