BlehBlendConditioning

ComfyUI's built-in conditioning combine and average nodes are pretty blunt instruments - they don't handle multiple conditioning items gracefully, and they don't respect the time ranges you might have attached to a prompt (from something like a ConditioningSetTimestepRange node upstream). This node is the more capable version: it blends two conditionings while actually understanding both of those things.

The one thing it explicitly does not understand is anything like regional conditioning masks - it respects time ranges, not spatial ones. And the node's own description makes a point worth internalizing before you build a workflow around prompt blending in general: conditioning is a sequence of encoded tokens, not concepts. Blending "a cute dog" with "sketch of a cat" doesn't blend the ideas of dog and cat - it blends token 1 with token 1, token 2 with token 2, and so on, so "a" gets mixed with "sketch," "cute" with "of," and so forth down the sequence. It's positional interpolation, not semantic interpolation. Manage expectations accordingly, especially with prompts of very different lengths or sentence structure.

Inputs that matter: conditioning_1 and conditioning_2 are the two things being blended. blend_mode defaults to lerp (plain linear interpolation, the safe default), though the pack exposes roughly 486 blend modes total if you want to dig into the same blending library bleh uses for latents - most of those are experimental and not something you need on day one. strength (0.5 default) sets how much of conditioning_2 gets mixed in. blend_cond_tensor (default true) controls whether the base conditioning tensor itself gets blended at all - turn it off if you only want to blend a secondary tensor like pooled_output while keeping conditioning_1's actual tokens untouched, which is what blend_tensors (default "pooled_output") names as a comma-separated list.

Then there's the part that actually matters once your two prompts aren't the same length, which is most of the time: size_mismatch_strategy decides what happens. zero pads the shorter one with zeros - ComfyUI's normal behavior. match_cond_1/match_cond_2 forces the result to one side's shape. replicate tiles the smaller tensor to fit the larger one, which the tooltip notes is well-suited to CLIP conditioning specifically, since size mismatches there land in increments of the 77-token CLIP limit. error just refuses rather than guessing, if you'd rather know something's wrong than get a silently weird blend. size_mismatch_alignment_mode (left/right/center) only matters for the padding strategies, controlling where the padding goes relative to your actual content.

metadata_base_mode (default prefer_cond_1) decides which conditioning's non-tensor metadata - timestep ranges, area info, that sort of thing - survives into the merged result, preferring conditioning_1's if it exists and falling back to conditioning_2's otherwise.

Output is a single CONDITIONING, wired into your sampler exactly like a normal prompt encode. Install is the standard route for the whole pack - ComfyUI Manager or a manual git clone into custom_nodes, then restart. No extra dependencies. If a blend comes out as word soup rather than a coherent mix of the two prompts, that's the token-level blending behavior described above doing exactly what it's supposed to, not a bug - try prompts closer in length and structure, or drop strength before assuming something's broken.