BYOKey ByteDance Seed LLM

This is the node for ByteDance's chat models - Seed, Doubao, and anything else you can reach through the Volcengine Ark inference API - inside a ComfyUI workflow. BYOKey_ByteDance_LLM is the BYOKey pack's answer to the official ByteDance LLM node: same OpenAI-shaped chat request, same STRING output, but it calls Ark directly with your key instead of going through Comfy's proxy. If you're already paying Volcengine for a Doubao endpoint - and the README explicitly targets users who route through third-party relays, which matters a lot for people in China - this is how that key gets used inside ComfyUI.

What you'd use it for is the usual LLM-in-the-graph work: rewriting a rough idea into a structured prompt, captioning images, generating text that other nodes consume. The one thing to know up front: Ark is OpenAI-compatible, so this node's request shape is identical to the generic BYOKey_LLM - the difference is which defaults, auth, and docs it ships with, and that it's tuned for the Ark host.

How it works

It POSTs to {base_url}/chat/completions - standard OpenAI chat format, which Ark speaks natively - with your prompt, an optional system_prompt, and any images you wire in as inline base64 data URIs for the vision-capable models. The reply comes back as a single output STRING.

The inputs that matter

• api_key - your Volcengine Ark key, sent as a Bearer token. This is the one thing you must get right; it's the entire point of the pack.
• base_url - defaults to https://ark.cn-beijing.volces.com/api/v3. The tooltip notes this is the international default; override it if your Ark project lives in a different region or you're going through a relay.
• model - a free-text field defaulting to doubao-1-5-pro-32k-250115. Here's the important part: on Ark you can also pass endpoint ids - the tooltip explicitly calls out deepseek-v3.1 as an example, so you can run DeepSeek's model through the same node if your Ark account has it enabled. Whatever model id or endpoint your Volcengine console exposes, type it here.
• system_prompt - foundational instructions for the model's behavior.
• max_tokens (default 4096, up to 128000), temperature (default 0.7), seed - the standard dials, with a generous token ceiling if you're generating long text.

Output is output (STRING). It feeds anything that takes a string: a prompt field, a text display node, another node's input.

Installing it

It's one of 37 nodes in the comfyui-byokey pack, and the pack installs clean with no extra dependencies - only what ComfyUI bundles (torch, aiohttp, PIL, numpy, av). Use ComfyUI Manager (search "byokey") or:

cd ComfyUI/custom_nodes
git clone https://github.com/MeteorAndy/comfyui-byokey.git

Restart ComfyUI and it appears under api/byokey/text. No models download; the GPU work happens on Volcengine's side.

Gotchas

Because Ark is OpenAI-compatible, the temptation is to use the generic BYOKey_LLM node with a base_url override and skip this one - and honestly, that works. This node earns its keep by shipping the right defaults and the Ark-specific documentation so you don't have to remember the endpoint. The real-world failure mode with Ark is the key/endpoint setup on Volcengine's side (making sure your model is actually deployed and its inference endpoint id is what you paste into model), not the node itself. And the pack-wide caution: your API key lives in the widget and gets saved into workflow JSON - never share a workflow with the key still in it. For a Chinese-market pack this is doubly true, since the whole point of the README is that relay keys are commonly shared around. Keep yours out of anything you post.