☁️BizyAir JanusModelLoader

What it is

Janus-Pro is DeepSeek's unified multimodal model - the same weights can both generate images from a prompt and look at an image and describe or answer questions about it. That's an unusual pairing; most models in ComfyUI do one job or the other. This loader is the entry point for BizyAir's Janus-Pro integration: load it once here, and it feeds both the generation node and the understanding (image Q&A) node elsewhere in this pack's Janus-Pro section.

When Janus-Pro landed in early 2025, the community's first question was basically "does this even run outside DeepSeek's own demo app?" - someone shipped a ComfyUI wrapper within days, and it does run locally if you've got a CUDA GPU with enough VRAM. It genuinely does not run at all on non-CUDA hardware (an Intel Arc user in that early thread hit a hard wall trying). BizyAir's version sidesteps that entirely by running it on their cloud, so the GPU and CUDA requirement becomes their problem instead of yours.

The inputs that matter

Just one field:

• model_name - currently a single option, deepseek-ai/Janus-Pro-7B, the 7-billion-parameter release.

You get two outputs from one load: model (BIZYAIR_JANUS_MODEL) and processor (BIZYAIR_JANUS_PROCESSOR). Both are required - wire the pair into BizyAir JanusImageGeneration if you want it to generate, or into BizyAir JanusImageUnderstanding if you want it to describe/answer questions about an image. Same loaded model, two different jobs downstream.

How to install it

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

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

Restart ComfyUI afterward. As with every node in this pack, you need a BizyAir account and API key configured before this loader - or anything it feeds - will actually run.

Common issues

Confusing this with a dedicated image generator or a dedicated captioner. Janus-Pro is neither exclusively - it's one model doing both, through architecturally separate encoding paths internally (the "decoupled visual encoding" is the whole point of the Janus line: understanding and generation don't fight over the same visual representation the way earlier unified models did). If you only need captioning, you're still loading the same generation-capable model; that's expected, not a sign you picked the wrong loader.

Unfamiliar BizyAir folder showing up unprompted. Worth a general note for this pack: ComfyUI Manager had a period where its missing-node inference misattributed nodes from unrelated custom node packs to BizyAir, so some people found this pack installed on their system without meaning to and understandably got spooked by an unfamiliar folder with Chinese branding. It's a legitimate commercial node pack from SiliconFlow, not malware, and the mapping issue was patched.

Downstream nodes error immediately after loading. Check the API key first. This loader itself doesn't pull any weights to your machine - there's nothing local to go wrong - so a failure here or immediately downstream almost always traces back to authentication or account credits rather than the model choice.