SeeDream 4.5 (BudgetPixel)

Seedream is ByteDance's flagship image line, and - like everything ByteDance actually sells - it has never been open-weighted. The whole Seedream family is API-only, full stop. So when you want Seedream 4.5 inside a ComfyUI graph, this node is the way in: the official BudgetPixel wrapper that sends your prompt (and optionally up to nine reference images) to the hosted model and drops the result back as a normal IMAGE. No weights to download, no 24GB VRAM, no ComfyUI install fight. Just a key and some credits.

Why you'd reach for it

Seedream 4.5 is the strong-aesthetics workhorse of the line - 2K and 4K output, good at keeping a subject consistent across a composition, and genuinely good at following a detailed prompt. Its party trick is multi-reference composition: wire in up to nine reference images and it'll blend subjects, styles and scenes from all of them, which is the closest thing a closed model gives you to a freeform collage tool. For product shots, character sheets, or "put this in that" editing, it's a strong default. The 4.5 line also reads as the budget-conscious Seedream tier at a flat 45 credits per generation - same as Qwen Image Pro's 1K - which makes it the one you try first before paying up for 5.0 Pro.

How it works

Same thin-client pattern as every BudgetPixel node. The IMAGE tensors on reference_images are serialized to PNG, sent inline or via the upload endpoint, a job is created against the API, and the node polls until it's done before downloading the result back into the graph. You'll see the progress bar creep; don't assume the graph hung, video-sized jobs just take a while. And remember the async gotcha: interrupting the run stops the wait, not the job - it keeps rendering and billing server-side.

The inputs that matter

• prompt - required. Text-to-image if no references are connected, editing/composition if they are.
• reference_images - up to 9. This is the whole multi-reference story, so it's the socket you'll actually build workflows around.
• size - 2K (default) or 4K. Flat pricing, so 4K costs nothing extra - unusual and worth knowing if output quality is the point.
• sequential_image_generation - set to auto and the model returns a group of related images, up to max_images. The model decides the exact count and you're billed per image returned, so treat max_images as a spending ceiling, not a count.
• num_images - 1–4, billed per image.

Output is images (IMAGE) - straight into Save Image or any downstream node.

Install and key setup

Standard for the pack: ComfyUI Manager → search BudgetPixel → install, or:

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

The only real dependency is requests. You'll need a BudgetPixel API key (included in every paid plan) via the BUDGETPIXEL_API_KEY env var or a budgetpixel.json next to the pack - the key is deliberately never stored in workflow JSON.

Where people get burned

The sequential_image_generation "auto" mode surprises people who expect exactly max_images images and instead get a bill for whatever the model returned. If you want one image, leave it disabled. The other classic is ComfyUI's output caching: an unchanged prompt reruns nothing, so use force_rerun when you actually want a fresh generation (and are OK paying for it). And on a free account you'll hit api_access_not_enabled - the API needs a paid plan.