Seedream5.0 Lite - Text to Image

What you get

Prompt in, image out, hosted Seedream 5.0 Lite. No reference image, no local weights, no VRAM. It's the cheapest sensible way into the Seedream 5.0 family from inside ComfyUI, and the family has a specific reputation worth stating plainly: strong adherence to complicated instructions and clean graphic output, weaker than the photoreal specialists at skin and light.

If you're comparing this against local generation, the calculus is the one that applies to every API node - free-versus-metered, yours-versus-theirs. Where it wins is when you want this particular model's way of composing a busy prompt without negotiating with a 20 GB download, and when the machine running ComfyUI isn't the machine you'd want doing diffusion.

Inputs

• prompt - required, documented at 3 to 3000 characters. The default is a paragraph of cafe-design marketing copy; delete it. It tells you the endpoint was documented for commercial use, not that it rewards verbosity.
• aspect_ratio - required, defaulting to 1:1. Pick deliberately; this is the field that most often makes a good generation unusable in a layout.
• quality - required. basic is 2K, high is 3K, ultra is 4K. The tiered ladder is Lite's distinguishing feature over the 4.5 nodes, which only offered basic/high.
• output_format - png or jpeg.
• nsfw_checker - default False on this text-to-image node (it defaults True on the Lite image-to-image sibling, so check which node you're on before assuming).
• timeout_seconds - 1200 default, range 30–7200. callback_url - leave empty for normal ComfyUI use.

Outputs are the pack's image standard: image, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
pip install -r ComfyUI-KIE-Nodes-Next/requirements.txt

Restart ComfyUI and set the key once under Settings → KIE.ai Nodes Next → Connection. It validates before saving, shows a masked fingerprint instead of your key, and never writes the secret into workflow JSON. On a server, KIE_API_KEY in the environment takes priority over the saved value.

How to not waste money

Two rules. Draft at basic, and only climb to ultra on the take you're keeping - the resolution tier is exactly where a casual session turns into an expensive one, and Lite-vs-Pro pricing chatter in the community exists because people learned this the hard way. Second, don't reroll blindly: each reroll is another metered call, so change one thing per attempt and read the on-node credit receipt so you can see the cost accumulating in real time.

One more thing about the ecosystem rather than the node. Seedream 5.0 discussion online is noticeably astroturfed - the same hosting platforms get name-dropped in post after post, and commenters call it out. That doesn't make the model bad. It does mean you should test it on your own image, with your own prompt, before you believe a comparison post claiming it beat everything.

And the standing API-node trade: your prompt leaves your machine, and the provider's moderation sits on top of whatever you configure locally. There's no patch for that - no weights to edit, no community bypass. For drafting stock-style imagery, fine. For material you'd rather keep private, use something local.