☁️BizyAir ReferenceLatent

This node's own description says exactly what it does, and it's worth quoting directly rather than paraphrasing: it "sets the guiding latent for an edit model," and "if the model supports it you can chain multiple to set multiple reference images." That's the whole concept in two sentences. Edit models - the Flux Kontext family and its successors - don't just generate from a text prompt, they generate conditioned on an existing image they're supposed to preserve, transform, or reference while producing the output. Something in the graph has to explicitly hand that reference image's latent representation to the model as an anchor, rather than the model just seeing a prompt in isolation. That's this node.

Why "chain multiple" matters

Most edit-model workflows are single-reference: one source image, one edit instruction, one output. But some edit models support multi-reference conditioning - think "combine the subject from this photo with the pose from that one" - and the node's own description is explicit that chaining several ReferenceLatent nodes together is exactly how you'd set that up, provided the underlying model you're running actually supports multiple references. If it doesn't, chaining extras won't do anything useful; check what your specific edit model supports before building a workflow around multiple reference latents.

The inputs and output

• conditioning (required) - the conditioning you're attaching the reference latent to, typically coming out of your prompt encoding.
• latent (optional) - the reference image's latent. This being optional rather than required is worth noting: it means you can technically pass conditioning through this node without a reference attached, which is presumably a no-op pass-through rather than something you'd deliberately reach for, but it does mean a missing latent connection won't throw an error - it'll just silently do nothing.

Output: BIZYAIR_CONDITIONING - the conditioning with the reference latent attached, ready to continue toward your sampler.

Installing it

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

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

Set your BizyAir API key on first use.

Where people get burned

Because latent is optional, the single most common mistake with this node is building a workflow that looks complete - conditioning goes in, conditioning comes out - while the actual reference image never got wired in. If your edit model's output shows no sign of respecting the reference you thought you'd given it, check this node first: an unconnected latent input is the quiet failure mode here, not an error message.

The second thing worth knowing is that this node's output type, BIZYAIR_CONDITIONING, keeps you inside BizyAir's cloud pipeline - it won't connect to a locally-run edit model's stock conditioning input, and a plain ComfyUI CONDITIONING won't plug into this node's input either. If you're trying to mix a BizyAir edit-model workflow with local nodes, the handoff has to happen at a plain type (an IMAGE or LATENT boundary), not through this node.