Seedream3Txt2Img(NYJY)

ComfyUI is usually a local-generation town, but not everything worth generating fits in your GPU. Seedream 3.0 - ByteDance's image model, part of the famously API-only Seedream line - is one of those. Seedream3Txt2Img wraps it: type a prompt, get back an image, zero VRAM used and zero models downloaded.

It's one of the Volcengine cloud nodes in ComfyUI_NYJY (aidenli's pack). The whole point of these nodes is that the heavy lifting happens on ByteDance's Volcano Engine servers and your machine just sends text and receives pixels. Think of it as an image API with a ComfyUI face.

How it works. The node calls the Ark image endpoint with your prompt and returns a base64 image it decodes straight into a normal ComfyUI IMAGE tensor. The model defaults to doubao-seedream-3-0-t2i-250415 - Seedream 3.0, the April 2025 text-to-image build. The date in the ID is the model version, and since the field is free text you can point it at other Ark model IDs if you know them.

The inputs that matter:

• ratio - nine presets, and notice the pixel sizes: 1:1 1328x1328, 3:4 1104x1472, 9:16 936x1664 and friends. Seedream 3.0 lives on the same 1328-based resolution grid the Qwen-Image models use, not the 1024 grid you're used to. Pick the aspect ratio you want and it's already correct.
• override_with / override_height - set both above zero to force a custom pixel size and ignore the ratio preset; leave both at 0 to trust the preset.
• guidance_scale - default 2.5, range 1–10. This is Seedream 3.0's strength-of-instruction knob. This one you'll actually tune.
• seed - fixed value for reproducible output, 0 for random.
• watermark - off by default; on if you're okay with a visible platform mark.
• api_key - empty means "use the pack config", filled means "use this instead".

The outputs. Image is the generated picture as a standard IMAGE - preview it, save it, feed it onward. Width and Height report the returned dimensions, handy if you're wiring size into downstream nodes.

The gotchas. Three, and they're shared across the whole Volcengine section of this pack. First, failure handling: when the API errors, the node returns a blank black 512×512 image instead of raising - the real error goes to the ComfyUI console, so a black output means "read the log", not "the model drew a black rectangle". Second, the API key: you need a Volcano Engine Ark key, and signing up as a non-China user has historically meant real-name verification walls - check before you commit. Third, this is the older generation: Seedream 4.0 (in the pack's Seedream4Txt2Img node) is the current one and handles 2K output; reach for 3.0 when you want the lighter, older model.

Install. It comes with the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/aidenli/ComfyUI_NYJY
pip install -r requirements.txt

or ComfyUI Manager → "Install via Git URL" → https://github.com/aidenli/ComfyUI_NYJY. The dependency that matters for this node is volcengine-python-sdk[ark].

When to reach for it. When you want Seedream 3.0 quality without local compute, or you're deliberately staying on the older, cheaper model. If you want the latest, the Seedream 4 nodes are right there in the same pack.