☁️BizyAir Seedream 4.5

Seedream is ByteDance's image-generation line, and it's never showing up as a checkpoint you download - ByteDance runs a strict split where research components go open (Depth Anything, PuLID, SDXL-Lightning) and anything it considers a finished product stays API-only. Seedream is squarely in the second bucket. This node is BizyAir's wrapper around it: a plain ComfyUI node that quietly makes a cloud call every time you queue it, billed against your BizyAir credits rather than your GPU's watts.

4.5 is the newest revision in this pack - the model id is doubao-seedream-4-5-251128, dated later than the plain BizyAir Seedream4 node's doubao-seedream-4-0-250828 sitting elsewhere in the same pack. If you're choosing between the two, this is the one to reach for.

What it does

Type a prompt, hand it a size (2K, 4K, or Custom with your own custom_width/custom_height in the 480–4096 range), and it generates. Where it gets interesting is the reference-image side: the optional images input takes up to 14 reference photos, and BizyAir gives you a genuinely nice bit of UI for it - the inputcount slider (1–14) dynamically adds more image slots when you click the node's "Update inputs" button, instead of forcing you to guess ahead of time how many you'll need. That's useful for multi-image composition - blend a subject from one photo with a background or style cue from another.

max_images (1–15) caps how many outputs you get back, with a catch worth knowing: your reference image count plus your generated image count can't exceed 15 total. Feed it 10 references and you'll only get 5 outputs regardless of what max_images says.

optimize_prompt is a two-way switch - disabled (default) sends your prompt as-is, standard runs it through ByteDance's own prompt-rewriting pass first for higher-quality output at the cost of taking noticeably longer. Worth toggling on when you're not getting what you expected from a terse prompt.

Custom sizing isn't free-for-all: the tooltip specifies a total-pixel budget (roughly 2.56M to 16.8M pixels) and an aspect ratio bound of 1:16 to 16:1. Stay inside those and Custom behaves; go outside and expect the request to get clamped or rejected.

Outputs

images (the generated batch) and a passthrough bizyair_model_name label - mostly useful if you're chaining this into a node that logs or displays which model actually ran.

Installing it

ComfyUI Manager: search BizyAir, install, restart. Or by hand:

cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git

Then restart ComfyUI. No model weights to download - the whole point of this node is that it doesn't run locally. What you do need is a BizyAir account and an API key set on first use (the pack prompts you with a "click to login" link); without it, every one of these API nodes just fails at the call.

Where people get burned

It's a billed cloud call, not a free local node. BizyAir added paid billing in April 2025, and image generation on a closed frontier model costs real credits per request - budget for it if you're iterating fast on prompts.

BizyAir occasionally gets installed by accident. ComfyUI Manager's missing-node auto-resolution has misattributed unrelated node names to BizyAir before (a well-documented case involved the completely separate Nunchaku pack), which is why some people find a BizyAir folder they never meant to install and understandably get suspicious. It's a legitimate, actively maintained pack from SiliconFlow - not malware - but if you didn't mean to install it, remove it cleanly through Manager rather than hand-deleting folders.

Reference-image math trips people up. The 14-reference-image cap and the "references + outputs ≤ 15" rule aren't obvious from the node alone - if max_images seems to be doing nothing, check how many reference images you're feeding in first.