☁️BizyAir JanusImageGeneration

What it is

This is the text-to-image half of DeepSeek's Janus-Pro - a model that's unusual in the ComfyUI ecosystem because the same weights can also look at images and answer questions about them (that's the other node in this pack's Janus-Pro section). As a pure image generator it's not chasing Flux or SD3.5-level photorealism; Janus-Pro's actual claim to fame is the unified architecture, not raw output quality. It's worth trying specifically when you want generation and understanding in one model rather than juggling two separate ones, or you're just curious what a genuinely unified multimodal model produces.

The inputs that matter

• model / processor - the pair from BizyAir JanusModelLoader.
• prompt - plain text, multiline, defaults to "A beautiful photo of" as a starting stub.
• seed - reproducibility, as usual.
• batch_size - has an unusual floor: minimum 4, maximum 9. You can't request just one image from this node; it always generates at least four candidates in a batch. Plan your workflow around picking the best of a batch rather than expecting a single clean output.
• cfg_weight - classifier-free guidance strength (default 5, range 1-10). Higher pushes the output to hew closer to the prompt at some cost to variety.
• temperature - sampling randomness (default 1, range 0.1-2). Lower values give more conservative, repeatable outputs; higher values give more variation between batch members.
• top_p - nucleus sampling cutoff (default 0.95).

Output is images (IMAGE) - the generated batch, ready to feed into whatever comes next in your graph.

How to install it

ComfyUI Manager: search "BizyAir," install, restart. Or manually:

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

You need a BizyAir account with an API key set up - this node is entirely a cloud call, and Janus-Pro itself genuinely won't run on non-CUDA hardware even if you wanted to self-host it, so the cloud route isn't just convenience here, it's the only option for a chunk of people.

Common issues

Expecting a single image and getting four to nine. That's not a bug, it's the node's floor - batch_size bottoms out at 4. If your downstream workflow assumes exactly one image, add a selection step (or just accept the first) rather than assuming something's broken.

Output quality below what you'd get from a dedicated image model. Janus-Pro's design goal was unifying generation and understanding in one model, not topping image-quality leaderboards. If pure photorealism or prompt fidelity is the priority and you don't need the understanding side of Janus-Pro at all, a dedicated generator elsewhere in the pack (or your normal txt2img pipeline) will usually outperform it.

Prompt seems to get half-followed. The default stub prompt ("A beautiful photo of") is a hint about how this model wants to be prompted - as a photo-style description more than an instruction. Try writing prompts in that descriptive register rather than as a command, and lean on cfg_weight if adherence still feels loose.

Job fails or times out. Check your API key and account credits first - this is a network round-trip to SiliconFlow's infrastructure, and auth or billing issues are the most common cause of a failed run that has nothing to do with your prompt or settings.