EvoLink Seedream 4.5 (Official)

Seedream is ByteDance's image model - the thing behind Doubao, and one of the few current-gen image models you will never, ever be able to run locally. There are no weights to download; Seedream is an API product and stays that way by corporate posture, not by accident. So if you want it inside a ComfyUI workflow, you use a wrapper node like this one. The EvoLink Seedream 4.5 node is that wrapper: type a prompt (Chinese included), and ~30 seconds later a real image drops into your graph like a local sampler made it.

What it actually is

EvoLink is a pay-per-use cloud API that resells closed models - Seedream, Sora, Veo, Wan 3.0, Z-Image and friends. This node is the official EvoLink one for Seedream 4.5, and it covers text-to-image, image-to-image and image editing in a single node. That "one variant, one node" split is deliberate: because each model gets its own node, the parameters you see are the parameters the model actually accepts. No mystery fields that silently do nothing.

Mechanically it's a thin HTTP client wearing a node costume. It uploads any reference images to EvoLink's file channel, fires off a task to /images/generations, polls until it's done, then downloads the result and hands it back as a normal IMAGE tensor. The result links EvoLink returns expire after 24 hours, which is exactly why the node bothers to download for you - wire the image output into a Save Image node and it's yours permanently.

The inputs that matter

prompt is required, supports Chinese, and up to 32,000 characters. quality is the resolution tier - 2K or 4K - and it's where your credit spend lives, so start at 2K for drafts. size takes auto, a ratio like 16:9, or explicit pixels like 2048x2048.

For image editing, plug a Load Image into image1 and describe the change. You get four reference slots (image1–image4) plus an image_urls_text box for existing public URLs, and they mix freely. Up to 15 images per run if you set n - but the model only produces multiple images when the prompt asks for them, so write "四张不同角度的..." if you actually want four.

Outputs

Five of them, same as every EvoLink node. image is the result, downloaded and ready to save. task_id is your receipt for the job log at evolink.ai, status_info summarizes model/time/credits, result_urls has the raw 24-hour links, and response_json is the full platform response if you want to build automation on it.

Installing and troubleshooting

This node ships inside the comfyui-evolink pack, so install that once and you get every EvoLink node at the same time. ComfyUI Manager → Custom Nodes Manager → search EvoLink → Install, then fully restart ComfyUI - close the console window, don't just refresh the browser. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/deeplearning-goethe/comfyui-evolink

Only dependency is requests, which ComfyUI already has. No model downloads, no GPU needed.

You'll need an API key from evolink.ai/dashboard/keys (sk- prefix). Fill api_key once and it's stored locally; leave it blank afterward so your key doesn't ride along in shared workflow JSON. EVOLINK_API_KEY env var works too.

The usual beginner trips: the node not appearing in your search after install (wrong folder nesting, or you didn't fully restart), and "it finished but I can't see the image" (the image output wasn't connected to a Preview/Save node - rerun after connecting; it's cached, so it won't charge you again). Error codes map cleanly: 401 is a bad/deleted key, 402 is no credit, 429 is rate limiting, and content-policy rejections don't charge you at all. Rerunning an unchanged workflow hits ComfyUI's cache - no new request, no new charge.