SeeDream 5.0 Pro (BudgetPixel)

SeeDream 5.0 Pro is the precise one. Where the rest of the Seedream line is about generating or composing, this node is built around instruction-based editing with full layer control - you hand it up to nine reference images, describe the edit, and it keeps the structure while changing what you asked to change. That's the job ByteDance prices up: 55 credits per image at 1K, 110 at 2K, and - the bit that surprises everyone - input images bill 5 credits each after the first. Your first reference image is free; the second through ninth each cost. It's the only Seedream node here that charges you for what you feed it, and it's the reason a careful composition can cost more than the generation itself.

Why you'd reach for it

Anything where identity and structure have to survive the edit. Change the background of a product shot, swap an outfit, move a subject into a new scene - the "full layer control" framing is ByteDance's claim that the model separates content from style well enough to edit one without wrecking the other. It's also the Seedream with the least batch appeal: no grouped/sequential generation mode, so it's one prompt → one considered image, which is exactly what a final-asset workflow wants.

The inputs that matter

• prompt - required. The instruction. With references connected, this is an edit command, not a description.
• reference_images - up to 9. First is free; each additional bills 5 credits per generated image. Don't wire in a 9-image batch out of habit - every extra input is money.
• size - 1K (55 credits) or 2K (110 credits) per image.
• num_images - 1–4; billed per image.

Output is images (IMAGE), straight into Save Image. There's no seed widget here (no seed in the schema), so reproducibility is whatever the server decides per run - if you need the same result twice, you'll pay for it twice.

How it works

The usual BudgetPixel mechanics: IMAGE inputs get serialized to PNG and sent to the hosted API, a job is created, the node polls until the server finishes, then the result downloads back into the graph. Same async caveat as every node in the pack - interrupting the graph cancels the wait, not the job, and a server-side job that keeps running keeps billing.

Install and key setup

Standard pack install: ComfyUI Manager → BudgetPixel, or:

cd ComfyUI/custom_nodes
git clone https://github.com/budgetpixel-ai/comfyui-budgetpixel.git
pip install -r comfyui-budgetpixel/requirements.txt

Only requests as a dependency; no model files. API key from a paid plan via BUDGETPIXEL_API_KEY or a budgetpixel.json in the pack folder - never in the workflow JSON.

Where people get burned

The per-input-image billing is the trap. People wire nine references "for best quality" and double their generation cost. Start with one image; add references only when you're actually using them. And if you want to prototype cheaply, run drafts through Lite at 40 credits flat before spending 55–110 per image here. Standard pack gotchas apply too: use force_rerun to defeat ComfyUI's caching when you truly want a new generation, and a free account will hit api_access_not_enabled because the API rides on paid plans.