Seedream 4 Text-to-Image

The plainest node in the pack, and that's its virtue. Seedream 4 Text-to-Image takes a prompt and a size and returns a Seedream 4 (seedream-4-0-250828) image as a normal IMAGE tensor. No mode switches, no dynamic inputs, no batch logic - if the unified Seedream 4 node feels like a lot of machinery for a simple "make me a picture" task, this is the stripped-down version. It's also the honest baseline: whatever you're going to do with Seedream 4 - edit, blend, animate - it starts here.

Seedream 4 is ByteDance's current flagship image model and it is strictly API-only, as the KB's ByteDance panel lays out: the entire Seedream line is a paid product, always has been, and no weights are coming. So this node is a thin client for the BytePlus images endpoint - prompt in, image URL back, tensor out. The trade is dead simple: you get flagship quality without a GPU, and you pay per image.

Inputs and outputs that matter

• prompt - everything. Seedream rewards detailed, structured prompts: subject, style, lighting, composition. This is the one input you'll actually fiddle with.
• width / height - output dimensions, default 2048×2048, up to 4096. Keep the aspect ratio sensible for your subject; the API bills roughly on resolution class, so 2048 is the sweet spot for most work.
• seed - set a value for reproducible results, leave -1 for random.

Output: one IMAGE tensor. Save it, upscale it, feed it into the image-to-video nodes - it's a first-class tensor.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/FloyoAI/ComfyUI-Seed-API
cd ComfyUI-Seed-API
pip install -r requirements.txt

Restart, and set your BytePlus API key in config.ini (under [API]) or as the BYTEPLUS_API_KEY environment variable. ComfyUI Manager users: search "ComfyUI Seed API Integration".

Where people get burned

The blank-black-image failure is this pack's catch-all error path - a black 512×512 output means the request failed and the real reason (API key, quota, model not enabled on your BytePlus account) is in the console. On the craft side, the mistake people make with Seedream is treating it like SDXL: short tag prompts. It reads natural language well, so whole sentences about the scene beat comma soup - but it still needs direction on style and lighting. And be aware of the sizing model: you pick width/height explicitly here, and an unusual resolution can surprise you, so stick to standard aspects unless you know why you're deviating. Paid per image, so iterate on the prompt rather than firing off long shots.