BailianVLOption(NYJY)

The vision-side sibling of BailianChatOption, and it does the same one job: turn a ticked checkbox into a model ID string. Wire its output into BailianVL's model input and you never have to remember a vision-model ID again. It's a picker, not a processor - zero inference happens here.

How it works

Six booleans, one per vision model the author surfaced from Alibaba's Bailian marketplace:

• qwen2.5-vl-72b-instruct - the default and the heavyweight; dense natural-language descriptions. The one most people actually use.
• qwen2.5-vl-7b-instruct - the small open-ish Qwen2.5-VL build, faster and cheaper than 72B.
• qwen-vl-plus - Alibaba's low-cost general VL tier.
• qwen-vl-max - the high-capability commercial VL tier on the platform.
• qvq-72b-preview - Qwen's vision-reasoning preview model; it "thinks" about the image before answering, which reads very differently in the output.
• qwen-omni-turbo - the omni model, which is multimodal beyond just vision (text, audio, video on the platform side).

Tick one and the node emits that model's ID as a STRING. Tick several and it follows list order, so in normal use you keep exactly one checked. None of these runs locally - they're all served from Bailian's DashScope endpoint, and the account requirements (real-name verification, API key in config.json, proxy off) are the same ones you already hit with BailianVL.

The one output

model - a STRING like qwen2.5-vl-7b-instruct, meant to plug straight into BailianVL's model field. No optional inputs, no other outputs.

Installing

Same pack install as everything else here: ComfyUI Manager → Install via Git URL → https://github.com/aidenli/ComfyUI_NYJY, restart. No model files, no extra config for the node itself.

Should you bother?

If you're the only person using your workflow, honestly no - typing qwen2.5-vl-72b-instruct into BailianVL takes two seconds. The case for this node is the shared or reused workflow again: you hand someone a graph, they want to compare the 72B result against the cheap qwen-vl-plus, and six checkboxes are more discoverable than an undocumented model ID in a text field. It's also genuinely useful for A/B testing during development - swap the tick, re-run, compare answers, no retyping. Thin node, but it does the one thing it does without surprises.