Seedream 5.0 Pro · Text to Image

Pro is the high-detail end of ByteDance's Seedream 5.0 line: a prompt of up to 5,000 characters, fifteen aspect ratios, 1K or 2K, from about $0.036 an image. One enable_thinking toggle separates it from the Flash tier, and it starts switched on.

Here's the thing about that toggle: you're billed per image, not per token or per second. Running the thinking pass costs you wall-clock time and nothing else. So unless you're iterating on prompt wording and want the plain, unaffected read of what you wrote, there's little reason to turn it off - a longer wait for a better frame at the same price is a good trade in almost every direction.

How it works

The node sends your prompt and format options to SpicyAPI, waits for the hosted model to render, and downloads the result as a standard ComfyUI IMAGE. There's no local model involved, so there's no VRAM floor, no checkpoint, and no sampler to fight. That's the entire pitch of this pack in one sentence: the model is somebody else's, the node is just the pipe.

The cost of that convenience is the standard one for API wrapper nodes - your prompt leaves your machine, the render is metered, and you can't do a thing about the deployment's behaviour beyond changing the options it exposes. If you need pixel-level control and reproducibility, a local checkpoint is still the better answer. If you need a big, well-composed frame from a long description in one shot, this is cheaper than an hour of prompt-fiddling.

Inputs

• prompt - subject, composition, style, lighting. Put any literal text you want rendered in quotes and keep it short.
• aspect_ratio - fifteen options, 1:1 by default. Match your destination: the model renders more composition into a tall frame than a square.
• resolution - 1k or 2k, default 2k.
• output_format - jpeg or png.
• enable_thinking - default on. A refinement pass over your prompt before generating; off is quicker and more literal.
• seed - fixed on purpose. Queueing the same graph twice reuses the result instead of billing a second render; set control after generate to randomize when you want a variation.

Output: images (your IMAGE), and task id, a STRING naming the SpicyAPI request - useful only when you're chasing a failure with support.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart ComfyUI. Then Settings > SpicyAPI → paste a key from spicyapi.ai/console/keys, or set SPICY_API_KEY in the environment on a server. Restart one more time so the node list is rebuilt from your account, which is also how you pick up models released after the pack's bundled snapshot. ComfyUI Manager: search SpicyAPI. Nothing else installs - no pip packages, no model files.

Price and the guardrails

About $0.036 per image, rate shown on the node and confirmed by a quote before submission. Set Settings > SpicyAPI > Max cost per run as a ceiling and a run quoted above it won't be submitted at all. The other guardrail is subtler: the pack forces seeds to fixed at creation time, so a habitual Ctrl+Enter doesn't buy you a second image by accident. And as with every node in this pack, cancelling a run stops your wait, not the job - it completes on the server and is charged if it succeeds, with the output sitting in SpicyAPI's console for two weeks.

Troubleshooting

No SpicyAPI key is set means the key isn't configured where the pack looks. 40310 is an unverified account email. 40004 means no deployment serves the combination you requested; the message names the option to change and nothing is billed. Errors ending in request_id ... are meant to be handed to support verbatim. Models missing from the menu need a ComfyUI restart, and if you're behind a corporate proxy, HTTPS_PROXY must actually work or requests fail with the pack's firewall-and-proxy message.