doubao-seed-1-6

This pack isn't only images and video - doubao-seed-1-6 is the text-generation node, and it's the quiet engine underneath the comic book workflow. It calls ByteDance's Doubao seed-1.6 LLM and drops the result back into your graph as a plain STRING. Story writing, captions, prompt enhancement, scene planning - if you need a language model inside a workflow, this is the node. And because it's the same model the comic node uses, it's also the way to script your own comic pipeline instead of taking the all-in-one path.

How it works

Standard chat-completions under the hood: your prompt plus an optional system_prompt are sent to the chosen mirror's text endpoint, and the reply comes back as generated_text. You get three model tiers: doubao-seed-1-6-250615 (the standard, default), plus two -flash variants (250615 and 250828) that trade some quality for speed and cost. Because it's a full LLM API, the tuning knobs you'd expect are all there - max_tokens, temperature (0–2), top_p, presence_penalty and frequency_penalty to discourage repetition - plus a stream toggle if you want streaming output. There's a 300-second timeout, which the code notes is there because story generation genuinely can take a while.

The mirror selection matters more than it usually does here: mirror_site (comfly / t8_mirror / volcengine) chooses the reseller, and each has its own key in SeedReam4_config.json. The node will also read the api_key field or the DOUBAO_API_KEY env var.

Outputs are generated_text (the reply), response_info (the raw response metadata), and usage_info (token counts - handy for keeping an eye on what a long story is costing you).

Install & setup

Same pack as everything else:

cd ComfyUI/custom_nodes
git clone https://github.com/xuchenxu168/ComfyUI_doubao_seed
cd ComfyUI_doubao_seed
pip install -r requirements.txt

Restart ComfyUI, configure your key, and the node appears under Ken-Chen/Doubao. Wire generated_text into a Show Text node to see output.

Gotchas

It's a metered LLM call, so the two things to watch are cost and moderation. max_tokens has a 4000 cap and defaults to 1000 - a creative-writing job with a long system_prompt will hit that ceiling mid-story, so bump it when you're generating anything longer than a paragraph. And the content policy is ByteDance's, not yours: a prompt that gets refused returns empty generated_text with the reason in response_info, which is also where you'll look first when a call silently fails. One more nuance - if you're using this to write comic scenes, the comic node's own text path already exists; this standalone node is for when you want the text without the full comic pipeline, or want to script one yourself.