☁️BizyAir InstructPixToPixConditioning

What it is

InstructPix2Pix was one of the earlier "just tell it what to change" image editors - you give it a source image and a plain-English instruction ("make it winter," "give her a red hat") instead of a prompt describing the whole scene from scratch, and it edits accordingly. No mask, no denoise dial to fight with. This node is the conditioning step that makes that pattern work: it takes your source image and folds it into the conditioning so the model has full-frame context to edit from, rather than generating something new that happens to look similar.

The mechanism is simpler than it sounds. Where a normal txt2img conditioning just carries your prompt, this one also concatenates the source image's latent alongside it, so every denoising step has the original picture sitting right there as a reference. That's structurally different from inpainting - there's no mask telling the model what's off-limits, so the whole frame is fair game for the edit, which is exactly the trade-off you want when the instruction genuinely changes the whole image's mood or lighting rather than one object in it.

The inputs that matter

• positive / negative - your conditioning, type BIZYAIR_CONDITIONING, coming from a BizyAir text-encode node upstream (this is where your edit instruction lives).
• vae - BIZYAIR_VAE. As with the rest of this pack, it's a BizyAir-specific type that won't connect to a stock local VAE node.
• pixels - the image you're editing.

There's no mask input here at all, which is the whole point - that's what separates this node from InpaintModelConditioning sitting right next to it in the same category.

Outputs are positive, negative (both BIZYAIR_CONDITIONING), and a latent - wire all three into your BizyAir sampler the way you would for any other conditioning node in this pack.

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 need a BizyAir API key configured - this conditioning step itself is lightweight, but the sampler it feeds is a cloud call, and nothing runs without an authenticated account behind it.

Common issues

The edit changes more than you asked for. This is the known trade-off of whole-frame instruction editing rather than a bug: since there's no mask constraining the change, unrelated regions can drift a little with every generation. If a specific region absolutely has to stay pixel-identical, this isn't the right node for the job - reach for BizyAir_InpaintModelConditioning with a mask instead.

Instruction gets ignored, or the output barely changes. Vague instructions tend to underperform specific ones - "make it look better" gives the model much less to work with than "add warm sunset lighting from the left." Keep the instruction concrete and about the change, not a full re-description of the scene, since a full re-description tends to fight the source-image conditioning instead of working with it.

Type mismatch errors on the conditioning or VAE sockets. The BIZYAIR_* types are deliberately walled off from ComfyUI's stock CONDITIONING/VAE types, so a regular CLIP Text Encode or VAELoader won't plug in here - everything feeding this node needs to be the BizyAir equivalent.

Nothing happens at all. Check the API key and account credits before assuming the node itself is broken - every node in this pack that touches a sampler is a network call under the hood, and auth or billing failures tend to surface as silent or generic errors rather than clear ones.