Conditioning (Average)

This is one of the oldest nodes in ComfyUI, and it does exactly what its search aliases promise: "blend prompts," "interpolate conditioning," "style fusion." You give it two conditionings, it averages their embeddings in latent space, and the sampler steers toward the blend instead of either prompt alone.

Inputs

• conditioning_to - the conditioning you want to keep. It anchors the blend.
• conditioning_from - the conditioning you're mixing in.
• conditioning_to_strength - 0 to 1. How much of to survives. 1.0 means pure to; 0.0 means pure from; 0.5 is a genuine 50/50 mix.

The math is a weighted sum in embedding space: to * strength + from * (1 - strength). That's the whole mechanism, and it's why the node behaves subtly differently from blending prompt text. Averaging embeddings can merge concepts in ways a string concatenation can't - two styles genuinely interpolate rather than listing next to each other.

How it works under the hood

The source reveals the practical details. conditioning_from gets truncated to to's token length, and if it's shorter, it's zero-padded to fit. The pooled output comes from conditioning_to. There's also a warning baked into the code: if conditioning_from contains more than one conditioning entry, only the first is actually applied. So the node assumes one-from, and anything fancier than that silently ignores your extra inputs.

Where it shines

Style mixing is the canonical use - take a portrait prompt and a painterly style prompt, blend at 0.6, get something in between. It's also the honest way to do "genre fusion" when you want the composition of one prompt and the mood of another. And because it operates on conditioning, you can chain it: blend two blends.

The trap

The two conditionings need to be compatible - same encoder family, same token dimension. Blending an SD1.5 CLIP conditioning with an SDXL one is a shape mismatch the node papering over with padding, and the result is nonsense that renders as "vaguely influenced by both but definitely neither." The zero-padding keeps the math from crashing, not the semantics. If you're on a model with one text encoder, blend conditionings from that same model only. Also worth noting: to and from are directional - flipping them with conditioning_to_strength held the same is not the same blend, because to keeps the pooled output.

Ships with ComfyUI core, ancient, stable, no install. For a beginner it's one of the safest "tricks" to play with - two encode nodes, one average, instant prompt mixing.