☁️BizyAir JanusImageUnderstanding

What it is

This is the "look at this image and tell me about it" side of DeepSeek's Janus-Pro - vision Q&A and captioning, running on the same unified model that also generates images elsewhere in this pack's Janus-Pro section. Point it at an image, ask a question in plain English, get a text answer back. It's genuinely handy for auto-captioning a batch of images for dataset prep, sanity-checking what a generation actually produced, or just asking specific questions about a picture ("what's the lighting direction here?", "describe the outfit") that a fixed caption model wouldn't answer.

The inputs that matter

• model / processor - the pair from BizyAir JanusModelLoader. Same loaded model as the generation node; this one just routes into the understanding path instead.
• image - what you're asking about.
• question - plain text, multiline, defaults to "Describe this image in detail." Swap it for anything specific - a targeted question generally gets a more useful answer than the generic default.
• seed - reproducibility for the response.
• temperature - default 0.1, notably low (range 0-1). That's a deliberate choice for a Q&A/description task: you generally want consistent, literal answers here, not creative variation, so the default is tuned conservative.
• top_p - nucleus sampling cutoff (default 0.95).
• max_new_tokens - caps response length (default 512, up to 2048). Bump this if you're asking for a genuinely long, detailed description and getting cut off.

Output is a single text (STRING) - wire that into a text display node, a save-text node, or feed it forward as a prompt for something else entirely (a common pattern: caption an image here, then use that caption as the prompt for a different generation).

How to install it

ComfyUI Manager: search "BizyAir," install, restart. Manual clone:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAir.git

You'll need a BizyAir API key configured. Janus-Pro genuinely requires CUDA and real VRAM to self-host, so this cloud route is doing real work for anyone without that hardware on hand, not just saving you a download.

Common issues

Answer is vague despite a specific question. Try lowering temperature further if it's not already near the default, and keep the question narrow and concrete - "what color is the car" gets a better answer than "tell me about this image" even with a specific-sounding question, because the model still has to parse what you actually want out of a broad phrasing.

Response gets cut off mid-sentence. That's max_new_tokens hitting its ceiling - raise it (up to 2048) if you're asking for long, detailed descriptions.

Using this when you actually just want a caption for training data. It'll do the job, but if you're captioning a large batch, factor in that every call here is a network round-trip billed against your BizyAir account - for genuinely large datasets, a locally-run dedicated captioning model might be the more economical choice, with this node reserved for smaller batches or cases where you specifically want Janus-Pro's answers.

Job fails outright. Check the API key and account credits before anything else - this node has no local fallback path, so authentication and billing issues are the most likely cause of a hard failure.