Seedream 5.0 Pro · Image Editing

Pro edit is the Seedream 5.0 node you reach for when the instruction is complicated: combine a face, a garment and a location into one frame, or revise one photo while keeping everything you didn't mention intact. One to ten references, fifteen aspect ratios, 1K or 2K, from about $0.036 an image - and a enable_thinking toggle that's on by default.

That toggle is the only thing here you won't find on the Flash and Lite siblings. Leave it on and the model spends a pass refining your prompt before it draws. You're charged per image, not per thought, so the cost of leaving it on is seconds, not money - a nice rarity. Turn it off when you're iterating fast on wording and want a straight read of what you typed.

What the node is doing

It's a client. On Run, every connected image gets encoded (PNG, or JPEG if the PNG would cross the 10 MiB upload ceiling), uploaded, and referenced in a job submitted to SpicyAPI after a price quote. The result comes back as a ComfyUI IMAGE batch plus an alpha MASK. Your machine never loads a model - the appeal if you're on a laptop, and the reason this pack needs no GPU at all.

Inputs worth knowing

• image_urls - one to ten sources, as growing sockets: connect one and the next appears. Ordering is meaningful, so say in the prompt which picture does what.
• prompt - the instruction. Name the sources ("picture 1's face, picture 2's jacket") rather than hoping the model infers roles.
• aspect_ratio - auto by default, which picks the supported ratio closest to your first reference. Override it deliberately for a delivery format.
• resolution - 1k or 2k, default 2k. There's no third tier here, unlike Flash.
• background - opaque, or transparent to get the subject cut out for real. Transparent requires output_format png and a single PNG source that already has an alpha channel; that combination is the bit people get wrong and then blame the model for.
• output_format - jpeg or png.
• enable_thinking - on by default. Off is faster, and slightly more literal about your wording.
• seed - fixed by default so re-queueing reuses the cached result rather than charging you again. Switch control after generate to randomize when you want a variation.

Outputs: images, alpha (1 = opaque), and task id, the STRING that identifies this request in SpicyAPI's logs.

Install

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

Restart, add a key from spicyapi.ai/console/keys in Settings > SpicyAPI (or export SPICY_API_KEY), restart again to pull your account's model list. Manager users: search SpicyAPI. No dependencies, no weights, no downloads - this pack installs as pure Python over what ComfyUI already ships.

Cost control and the usual errors

About $0.036 per image, quoted before submission and printed when the run finishes. Settings > SpicyAPI > Max cost per run is your ceiling; anything quoted above it stops before it's charged. Note that adding references pushes the quote up, which is normally fine for a compositing job and occasionally not.

Errors, mapped: No SpicyAPI key is set when the key is missing or wrong; 40310 when your account email is unverified; 40004 when no deployment serves the combination you picked - change the named option, nothing charged; anything ending request_id ... is a support-bound identifier. A missing model needs a ComfyUI restart. A cancel in the UI leaves the server job running, and it's billed if it succeeds.