Doubao视觉丨API

Doubao is ByteDance's answer to the vision-model game, and its strong suit, per the pack's own tooltip, is UI analysis, document understanding, and scene recognition - the kinds of jobs where you're reading a screenshot, a form, or a diagram rather than asking for art criticism. DoubaoVisionAPI wires that into a ComfyUI graph: image in, question in, text answer out, with a short conversation history. It's one platform only - 火山引擎 (Volcano Engine), ByteDance's cloud - so there's no platform dropdown to fight over; the single choice is which Doubao model to use.

It's part of ComfyUI-QING's API category and uses the same shared framework as the pack's other vision nodes, so key handling and history management are familiar if you've set up any of them.

How it works. The image and text_input (default "请描述这张图片的内容") go to the Volcano Engine endpoint; the model field offers nine Doubao variants, including the current-generation Doubao-Seed-1.6 family, Doubao-1.5-vision-pro, and the UI-specific Doubao-1.5-UI-TARS. The tooltip on the model field is unusually specific about what each is for: Seed-1.6 is the newest, -thinking does chain-of-thought visual reasoning, -flash is the fast real-time option, and UI-TARS is the interface-element recognizer. The image_quality knob (auto/low/high) is there if you need to trade accuracy for speed.

The inputs that matter.

• image - the screenshot, document scan, or scene you're analyzing.
• text_input - your question.
• model - pick the variant that matches the job (9 choices).
• max_tokens (default 4096) and history (default 10 rounds).
• Optional: temperature, top_p, image_quality, clear_history.

Outputs: generated_text, conversation_info, total_tokens. Same trio as the rest of the pack - text result, conversation metadata, and token usage.

How to install. ComfyUI-QING installs as one pack - search "ComfyUI-QING" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/GAO-SHIQING/ComfyUI-QING
cd ComfyUI-QING
python install_dependencies.py

Configure the Volcano Engine key in Settings → QING → API Configuration (or the matching env var) and restart. The README's clone URL is typo'd (GAOSHI-QING, missing the H) - use the URL above or Manager.

Troubleshooting. Because the platform is fixed, the key is the main thing to get right - a Volcano Engine (火山方舟/Volcengine Ark) API key, not some other ByteDance product key. If you get auth errors, that's almost always it. Second, model availability shifts as ByteDance retires old versions - if a listed model 404s, switch to the current Seed generation rather than assuming the pack is broken. And the standard API-node reminders: your image and prompt go to ByteDance's servers (relevant if the screenshot is sensitive), and Doubao enforces its own content policy. For purely local captioning of a training set, the open-source captioner path in the KB is still the better default; this node shines when you specifically want Doubao's UI/document strengths inside a live workflow.