Seedream 5.0 Flash · Image Editing

Two jobs, one node. Either you hand it a single photo and describe a change, or you hand it a face, a garment and a location and ask for a frame that contains all three. Up to ten references, fifteen aspect ratios, output anywhere from 1K to 2K for $0.029 - and all three resolution tiers cost the same, which is the detail that decides how you use it.

The Flash tier is the cheap one in the Seedream 5.0 line. If you're picking between this and Seedream 4.0's edit node, the split is simple: 4.0 gives you exact pixel dimensions from a preset list, 5.0 Flash gives you aspect ratios, transparency, and JPEG/PNG output. Ratio-shaped work goes here.

What it does under the hood

Nothing renders locally. The node encodes each connected image (PNG first, falling back to high-quality JPEG only if the PNG would exceed the 10 MiB upload cap), uploads them, takes a price quote from SpicyAPI, submits the task and polls for the result. What comes back is a normal IMAGE tensor plus, when the model returned transparency, a matching MASK. Your GPU is only involved in loading and saving.

References are sockets that grow: connect an image into image_urls and a new socket appears. Order matters - the prompt should say what each one contributes, in the order you wired them.

The inputs worth setting

• image_urls - one to ten sources. More references raise the quote, so add them deliberately.
• prompt - describe the picture you want back and what each reference is for. "Use picture 1 for the face, picture 2 for the jacket" beats hoping it infers that.
• aspect_ratio - the default is auto, which keeps the shape of your first reference. That default is right most of the time; override it when you need a 9:16 deliverable out of a landscape source.
• resolution - 1k, 1.5k, 2k, all at one price. Take the 2k unless you have a reason not to.
• output_format - jpeg by default; pick png if you care about staying lossless or you want alpha.
• background - opaque normally, transparent when you want the subject cut out for real. Transparent needs png output and a single PNG input that already has alpha, which is the part people miss.

Outputs are images, an alpha MASK (1 = opaque) for compositing, and task id, a STRING that identifies the request in SpicyAPI's console logs or for support.

Install

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

Restart, then Settings > SpicyAPI → paste a key from spicyapi.ai/console/keys, or set SPICY_API_KEY on a server that has no settings panel. One more restart and the node list comes from your account. ComfyUI Manager: search SpicyAPI. No dependencies install with this pack, and there are no weights - it's a client.

Costs, and where it bites

$0.029 per image, shown on the node, quoted before submission. Set Settings > SpicyAPI > Max cost per run so a fat reference set can't quietly make an expensive run; the quote is compared against that ceiling and the job never leaves if it's over. Re-queueing the same graph is safe - control after generate starts on fixed, so you get the cached image instead of a second charge. Interrupting a run in ComfyUI doesn't cancel it on the server.

Troubleshooting

The errors you'll plausibly hit, in the pack's own words: No SpicyAPI key is set (key missing or mistyped), 40310 (your SpicyAPI email isn't verified yet), 40004 (no deployment serves the option combination you picked - change the named option, nothing charged), and any message ending in request_id ... (hand that id to support). A model missing from the menus needs a ComfyUI restart, and if you're behind a proxy, HTTPS_PROXY has to be set correctly or the request fails with a message about a firewall or proxy intercepting it.