Seedream Image Generate

Seedream is ByteDance's flagship image model, and it is never coming to your disk. The company's posture is blunt: components and accelerators ship open, anything that competes with the flagship stays on the API. So if you want Seedream 5.0 Pro quality inside a ComfyUI graph, this is the node you do it through - it's a thin wrapper around the Volcengine ARK image API, cloud generation with an API key instead of a checkpoint. Your 3090 gets a vacation.

That framing matters, because it changes what "using" the node means. There's no models folder to fill, no VAE to match. You install one Python SDK, paste a key, and every image costs whatever ByteDance charges per call. Worth it if you need Seedream's specific strengths - the multi-reference identity work the community raves about, top-tier text rendering - and pointless if your workflow is a fully local, private pipeline.

The inputs that matter

Most of the node is pass-through to the API, but three settings decide everything:

• model - four choices: Seedream 4.0, 4.5, 5.0, and 5.0-pro. Defaults to the pro. Don't assume newer is strictly better for your use case; 5.0 Pro actually has a smaller resolution ceiling than the others (more on that below).
• prompt - multiline, works in Chinese and English. This is the whole game.
• aspect_ratio - 14 presets from 1:1 to 21:9, plus 2K/3K/3.5K/4K. Each maps to a fixed size; there's no arbitrary resolution here. That's what the V2 variant is for.

Then the multi-image feature: image1 through image5 optional image inputs. None connected = text-to-image. One or more = image-to-image, and this is the identity-preservation sweet spot - feed a few reference shots and Seedream holds the subject across style changes. That's the use case people actually run this for.

Multi-image output (read this twice)

The default config returns exactly one image, and the path to more is non-obvious. To generate several variants you need all three of: sequential_image_generation on enabled (or auto), max_images above 1, and stream set to True. Miss the stream toggle and you silently get one image. The node then works the streaming response for image_generation.partial_succeeded events and collects only the ones that carry a real URL or base64 payload - so the count you get back can be less than max_images. The console logs the collection progress; the text output lists every image's URL.

Outputs and the reproducibility lie

You get images (a list of real IMAGE tensors - wire it into SaveImage and it works) plus a text log with the generation info. But here's a detail the node is honest about in its own source: the seed input is normalized and never sent to the API. The Seedream API doesn't support seeding yet, so the seed is workflow-tracking sugar. Don't expect identical reruns - if you need to reproduce a result, you reproduce the call, not the seed.

Install

ComfyUI Manager → search "Seedream Image Generate", or:

cd ComfyUI/custom_nodes
git clone https://github.com/dzy1128/Seedream-Image-Generate-ComfyUI
pip install volcengine-python-sdk[ark]

You must also activate the model in the Volcengine console (model square → find a Seedream model → API接入 → pick your key → 开通模型) and set ARK_API_KEY. No model files, no heavy deps beyond the SDK - the "heavy" part is that the API needs network access and has quotas, so budget for a few failed calls while you tune prompts. Watch the resolution ceiling: 5.0 Pro is capped around 2048×2048 while the older models stretch to 4096×4096. If you hit a "resolution" error in text, that's why - switch models or use the V2 node with explicit dimensions.