Blend Latents

WAS's README documents Image Blend ("Blend two images by opacity") but doesn't spell out Blend Latents by name - it sits in the same family conceptually, just one level deeper in the pipeline: instead of mixing two finished images by opacity, it mixes two LATENT tensors before they've been decoded into pixels at all.

Worth being honest about what that actually buys you, because it's easy to assume "blending latents" is just "blending images but earlier." It isn't, quite. An image blend is a straightforward, perceptually meaningful operation - 50% of pixel A plus 50% of pixel B looks like a crossfade. A latent isn't a picture; it's a compressed, non-linear encoding the VAE learned, and linearly averaging two of them doesn't correspond to any equally clean operation in pixel space. Blend two latents near the extremes - 90/10, 95/5 - and you'll mostly get something close to the dominant one. Blend them near 50/50 and you're in riskier territory: the result can come out mushy, ghosted, or structurally confused, because you're asking a space that isn't linear in any perceptual sense to behave as if it were.

That's not a reason to skip the node - it's a reason to know what you're reaching for it for. The legitimate use case is combining two latents that are meant to be close to each other before a final sampling pass: a "structure" latent and a "style" latent generated from related seeds or prompts, or crossfading between two seeds entirely in latent space so the whole pipeline stays on the GPU without an unnecessary decode-encode round trip through pixel space. Push the blend factor toward one side rather than parking it at 50/50 and you'll get more predictable results.

The shape of the node follows directly from that: two LATENT inputs, a blend factor (a float, almost certainly running 0 to 1 the way Image Blend's opacity does), and one LATENT output - feed that into your next KSampler stage or straight to VAE Decode if you're happy with where it landed.

Installing it: through ComfyUI Manager, search "WAS Node Suite" and install - the fast path. By hand: cd ComfyUI/custom_nodes && git clone https://github.com/WASasquatch/was-node-suite-comfyui, then pip install -r requirements.txt (portable installs: python_embeded\python.exe -s -m pip install -r requirements.txt; manual/venv installs: activate the venv first), then restart ComfyUI. No model download and no special dependency attached to this node specifically - it's tensor math, ready as soon as the suite loads.

Where it bites: the failure mode people actually report isn't the node crashing, it's the output looking wrong and assuming something's broken when really the blend ratio is just in the mushy middle ground described above - try pushing it toward one input before concluding the node is at fault. Beyond that, the usual caveat for this whole pack applies: WAS Node Suite has been unmaintained since December 2023 (the README says "Retired" outright), and the community's recurring complaint is the whole suite failing to import after a ComfyUI update rather than any single node breaking in isolation. If Blend Latents won't show up in your node search, check your console for an import error on the suite as a whole first.