EvoLink Seedream 5.0 Lite (Official)

If you've spent any time in image-gen circles lately, you've seen Seedream 5.0's name in the "what's actually good right now" arguments. ByteDance runs the whole Seedream line as API-only - no open weights, no local install, ever - so the only way to get it into a ComfyUI graph is a wrapper node. EvoLink Seedream 5.0 Lite is the volume tier of that family: the one you reach for when you need a pile of images, not one perfect shot.

Why Lite exists

The Seedream family splits into Pro and Lite. Pro gives you the best single image the model can produce; Lite trades a little of that polish for throughput, and it's the only one that will happily hand you a batch. Set n anywhere from 1 to 15, mention the count in your prompt ("十二张不同配色方案..."), and you get a full grid back in one task, at one resolution tier. For storyboards, concept sheets, style exploration, or anything where you're going to pick one and discard the rest, that's the workflow this node was built for.

It does the same trick as every EvoLink node under the hood: reference images get uploaded to EvoLink's own file channel (no third-party image host, no imgbb key - the old community packs made you do that dance), a task goes to /images/generations, and the node polls until done, downloads, and drops a IMAGE tensor into your graph. That download step matters because the raw result links expire in 24 hours; the node handles it so you don't have to.

The inputs that matter

• prompt - required, Chinese-friendly. For multi-image runs, the count has to be in the prompt itself.
• quality - 2K / 3K / 4K. This is the resolution tier and your main cost lever; start at 2K.
• n - 1 to 15, the hard cap on how many images one task can return.
• size - auto, a ratio (1:1, 9:16, 9:21...), or explicit pixels (2048x2048).

For image-to-image or editing, plug a Load Image into image1 (there are four slots, plus an image_urls_text box for existing public URLs - they mix). Up to 15 reference images... well, n reference slots up to the model's limit.

Outputs

image is the batch - a [B,H,W,C] tensor where B is however many images you asked for, so it plugs straight into Preview/Save Image. task_id is your charge receipt for the job log, status_info summarizes model/time/credits, and result_urls / response_json give you the raw links and full platform payload for chaining or automation.

Install and gotchas

This node ships in the comfyui-evolink pack - install once, get all EvoLink nodes. ComfyUI Manager → search EvoLink → Install, then a full restart (close the console, not just the browser tab). Or:

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

Only requests as a dependency, no model downloads, no GPU. Grab an sk- key from evolink.ai/dashboard/keys, fill api_key the first time, then leave it blank so it doesn't get embedded in workflow JSON you share.

The usual trips: "didn't show up after install" (restart fully, check the folder isn't double-nested), and "finished but no image" (the image output needs to actually connect to a Save/Preview node). 401 = bad key, 402 = no credit, 429 = too fast. Content-policy blocks don't charge. And if you rerun an unchanged batch, ComfyUI caches it - no second charge for the same output.