🤖 Leon Seedream 3 API

ByteDance runs the sharpest two-track posture in this business: every flagship model stays API-only while the components underneath ship free. Seedream - its image generation line - is firmly on the closed side, which is why the only sane way to use it from ComfyUI is a node like this. Leon_Seedream3_API_Node gives you the whole Seedream 3 family through one dropdown and one HyprLab key: seedream-3 for straight generation, dreamina-3.1 for generation with LLM-enhanced prompts, and seededit-3 for editing (it requires a reference image and throws a ValueError without one). ByteDance's image models are consistently strong on complex scenes and text rendering, so if a local model keeps mangling your subject, this is worth a look.

How it works

Standard pack plumbing - the shared HyprLabImageGenerationNodeBase POSTs an OpenAI-style payload to https://api.hyprlab.io/v1/images/generations with a Bearer token, retries with exponential backoff (five attempts), and converts the result to an RGBA tensor. The interesting part is how the node fences off parameters per model, because the three model strings in that dropdown take different settings:

• seedream-3 takes legacy_size (small/regular/big), aspect_ratio, and guidance_scale (the tooltip notes it needs at least 1.0).
• dreamina-3.1 takes resolution (1K/2K), aspect_ratio, and enhance_prompt - the "enhance the prompt with an LLM" toggle that's the whole point of this variant.
• seededit-3 takes guidance_scale and an image; aspect_ratio is explicitly not supported, and the node warns you if you set it.

The node is strict about the mismatches - set enhance_prompt on seedream-3 and it logs a warning that the parameter only applies to dreamina. It's the rare API node that documents its own footguns, which you'll appreciate the first time you flip models.

Inputs that matter

• model - the big decision, since it changes which other inputs matter.
• input_images_array - required for seededit-3. Feed it from the pack's Image Array Builder (IMAGE_ARRAY type).
• guidance_scale - 0–10, default 2.5. The Seedream family responds to guidance the way SDXL users think of CFG; don't crank it to 10 expecting sharper output, it'll just distort.

Outputs: image, image_url, seed - the pack standard.

Installing

ComfyUI Manager (search Leon) or clone:

cd ComfyUI/custom_nodes/
git clone https://github.com/l3ony2k/comfyui-leon-nodes comfyui-leon-nodes
pip install -r requirements.txt

Restart, paste a HyprLab key. No downloads, no VRAM - remote inference.

Common issues

The highest-frequency beginner fail is wiring a plain IMAGE into input_images_array when trying seededit-3 - it's an IMAGE_ARRAY socket, so use the Image Array Builder. After that, the "this parameter doesn't apply to this model" warnings in the console are informational but worth reading; they tell you which knob you're twiddling for nothing. And since every run is metered, do your prompt tuning on seedream-3 before you spend a dreamina enhance_prompt pass on an idea you're not committed to.